docX13AS PDF

X-13ARIMA-SEATS Reference Manual
Version 1.0
Time Series Research Staff

Statistical Research Division
U.S. Census Bureau
Washington, DC 20233
phone: 301-763-1649
email: x12@census.gov
WWW: http://www.census.gov/srd/www/x13as/
January 15, 2013

This page intentionally left blank.
Contents
1 Introduction 1
1.1 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Running X-13ARIMA-SEATS 5
2.1 Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Input errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Specifying an alternate output filename . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Running X-13ARIMA-SEATS on more than one series . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5.1 Running X-13ARIMA-SEATS in multi-spec mode . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5.2 Running X-13ARIMA-SEATS in single spec mode . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5.3 Special Case: File Names Containing Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Log Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.7 Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.8 Program limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 The Specification File and Its Syntax 17
3.1 Examples of Input Specification Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Print and save . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Dates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 General rules of input syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
ii
CONTENTS iii
4 RegARIMA modeling Capabilities 26

4.1 General model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Data input and transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Regression variable specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4 Identification and specification of the ARIMA part of the model . . . . . . . . . . . . . . . . . . 35
4.5 Model estimation and inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.6 Diagnostic checking including outlier detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.7 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5 Points Related to regARIMA Model Estimation 40

5.1 Initial values for parameters and dealing with convergence problems . . . . . . . . . . . . . . . . 40
5.2 Invertibility (of MA operators) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.3 Stationarity (of AR operators) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.4 Cancellation (of AR and MA factors) and overdifferencing . . . . . . . . . . . . . . . . . . . . . . 42
5.5 Use of model selection criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.5.1 Avoid using the criteria to compare models with different sets of outlier regressors when
possible . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.5.2 Model comparisons for transformed data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.5.3 Do not use the criteria to compare models with different differencing operators . . . . . . 48
6 Points Related to Seasonal Adjustment and Modeling Diagnostics 49

6.1 Spectral Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.2 Sliding Spans Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.3 Revisions History Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7 Documentation for Individual Specs 56

7.1 ARIMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.2 AUTOMDL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3 CHECK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7.4 COMPOSITE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7.5 ESTIMATE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.6 FORCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
iv CONTENTS
7.7 FORECAST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

7.8 HISTORY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.9 METADATA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.10 IDENTIFY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.11 OUTLIER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.12 PICKMDL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.13 REGRESSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
7.14 SEATS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.15 SERIES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.16 SLIDINGSPANS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.17 SPECTRUM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.18 TRANSFORM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.19 X11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.20 X11REGRESSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
A Graphics Codes 235
B Print and Save Tables 239

B.1 Print and Save Tables Available for X-13ARIMA-SEATS specs . . . . . . . . . . . . . . . . . . . . . 239
B.2 Special Output Related to the seats Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
B.3 Special tables Used to Save X-13ARIMA-SEATS Output as Percentages . . . . . . . . . . . . . . . 249
C Irregular-component Regression Models Used 251

C.1 Irregular regression models for multiplicative decompositions. . . . . . . . . . . . . . . . . . . . . 251
C.1.1 Obtaining separate trading day and holiday factors . . . . . . . . . . . . . . . . . . . . . . 253
C.1.2 Estimating only holiday effects or stock trading day effects. . . . . . . . . . . . . . . . . . 254
C.1.3 Estimating user-defined flow trading day and/or holiday effects . . . . . . . . . . . . . . . 254
C.2 Irregular regression models for other decomposition modes . . . . . . . . . . . . . . . . . . . . . . 255
C.2.1 Additive Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
C.2.2 Pseudo-Additive Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
C.2.3 Log-Additive Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
C.3 When tdprior is used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Bibliography 259
Index 267
List of Tables
2.1 X-13ARIMA-SEATS Program Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.2 X-13ARIMA-SEATS Program Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 X-13ARIMA-SEATS Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.1 Predefined Regression Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1 Probability that a Chi-Square Variate with ν Degrees of Freedom Exceeds 2ν + ∆AIC
for ∆AIC = 0, 1, 2, 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.1 Revision Measure Calculated for Revision Lag Analysis . . . . . . . . . . . . . . . . . . . 55
7.1 Available Output Tables for Automdl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

7.2 Available Log File Diagnostics for Automdl . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.3 Available Output Tables for Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.4 Available Log File Diagnostics for Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.5 Default Output Tables for Composite Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.6 Other Output Tables for Composite Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.7 Tables Saved As Percentages in the save Argument . . . . . . . . . . . . . . . . . . . . . 83
7.8 Available Log File Diagnostics for Composite . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.9 Default Output Tables for Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.10 Other Output Tables for Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.11 Available Log File Diagnostics for Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.12 Example of ARMA Roots Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.13 Default Output Tables for Force spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
v
vi LIST OF TABLES

7.15 Choices for the target Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.16 Available Output Tables for Forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.17 Choices Available for the estimates Argument . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.18 Default Output Tables for History Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.19 Other Output Tables for History Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.20 Available Log File Diagnostics for History . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.21 Available Output Tables for Identify . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.22 Default Critical Values for Outlier Identification . . . . . . . . . . . . . . . . . . . . . . . 125
7.23 Available Output Tables for Outlier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.24 Available Output Tables for Pickmdl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.25 Available Output Tables for Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.26 Available Log File Diagnostics for Regression . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.27 Predefined regression variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.28 Change of Regime Regressor Types and Syntax . . . . . . . . . . . . . . . . . . . . . . . . 145
7.29 500 Year (1600-2099) means for Easter regressors of different window length w. . . . . . . . . . . 148
7.30 Output Tables for Seats spec that can be Printed and Saved . . . . . . . . . . . . . . . 158
7.31 Output Tables for Seats spec that can Only be Saved . . . . . . . . . . . . . . . . . . . . 159
7.33 Available Log File Diagnostics for Seats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.34 X-13ARIMA-SEATS File Extensions for Special SEATS Saved Output . . . . . . . . . . . . 162
7.35 Available Output Tables for Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
7.36 Default Formats for Each X-11 Format Code . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.37 Default Output Tables for Slidingspans Spec . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.38 Other Output Tables for Slidingspans Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.39 Sliding span lengths for different seasonal filters chosen . . . . . . . . . . . . . . . . . . . . . . . . 183
7.40 Seasonal MA parameter greater than 0 at which the span length increases to the value indicated. 184
7.41 Available Output Tables for Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.42 Available Log File Diagnostics for Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.43 Choices Available for the series Argument . . . . . . . . . . . . . . . . . . . . . . . . . . 190
LIST OF TABLES vii
7.44 Transformations Available Using the function Argument . . . . . . . . . . . . . . . . . . 195

7.45 Available Output Tables for Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
7.46 Default Output Tables for X11 spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
7.47 Other Output Tables for X11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.48 Plots Specified by the print Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
7.50 Available Log File Diagnostics for X11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
7.51 X-13ARIMA-SEATS Seasonal Filter Options and Descriptions . . . . . . . . . . . . . . . . . 209
7.52 Modes of Seasonal Adjustment and Their Models . . . . . . . . . . . . . . . . . . . . . . 212
7.53 Number of Surrounding SI-ratios in Table D 8.B Assumed Affected by a Level Shift 215
7.54 Default Output Tables for X11regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.55 Other Output Tables for X11regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
7.56 Predefined regression variables for x11regression . . . . . . . . . . . . . . . . . . . . . . . 229
A.1 Codes Associated with the X-13ARIMA-SEATS Graphics Metafile . . . . . . . . . . . . . . . 235
B.2 Output Tables for Seats spec that can Only be Saved . . . . . . . . . . . . . . . . . . . . 248
B.3 X-13ARIMA-SEATS File Extensions for Special SEATS Saved Output . . . . . . . . . . . . 249
B.4 Tables That Can Be Saved as Percentages in the save Argument . . . . . . . . . . . . . 250
1 Introduction
Contents
1.1 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
The X-13ARIMA-SEATS seasonal adjustment program is an enhanced version of the X-11 Variant of the Census
Method II seasonal adjustment program (Shiskin, Young, and Musgrave 1967). The enhancements include a
more self-explanatory and versatile user interface and a variety of new diagnostics to help the user detect and
remedy any inadequacies in the seasonal and calendar effect adjustments obtained under the program options
selected. The program also includes a variety of new tools to overcome adjustment problems and thereby enlarge
the range of economic time series that can be adequately seasonally adjusted. Examples of the use of these
tools can be found in Findley and Hood (1999). Basic information on seasonal adjustment is given in Chapter 2
of Dagum and Cholette (2006) and in Chapter 1 of Ladiray and Quenneville (2001), where the X-11 method is
thoroughly documented. See also Bell and Hillmer (1984, 1985), den Butter and Fase (1991), and Klein (1991).
The chief source of these new tools is the extensive set of time series model building facilities built into the
program for fitting what we call regARIMA models. These are regression models with ARIMA (autoregressive
integrated moving average) errors. More precisely, they are models in which the mean function of the time
series (or its logs) is described by a linear combination of regressors, and the covariance structure of the series
is that of an ARIMA process. If no regressors are used, indicating that the mean is assumed to be zero, the
regARIMA model reduces to an ARIMA model. There are built-in regressors for directly estimating various
flow and stock trading day effects and holiday effects. There are also regressors for modeling certain kinds of
disruptions in the series, or sudden changes in level, whose effects need to be temporarily removed from the
data before the X-11 methodology can adequately estimate seasonal adjustments. To address data problems
not provided for, there is the capability of incorporating user-defined regression variables into the model fitted.
The regARIMA modeling module of X-13ARIMA-SEATS was adapted from the regARIMA program developed
by the Time Series Staff of U. S. Census Bureau’s Statistical Research Division.
Whether or not special problems requiring the use of regressors are present in the series to be adjusted,
a fundamentally important use of regARIMA models is to extend the series with forecasts (and backcasts) in
order to improve the seasonal adjustments of the most recent (and the earliest) data. Doing this mitigates
problems inherent in the trend estimation and asymmetric seasonal averaging processes of the type used by
the X-11 method near the ends of the series. The provision of this extension was the most important technical
improvement offered by Statistics Canada’s widely used X-11 program. Its benefits, both theoretical and
empirical, have been documented in many publications, including Geweke (1978), Dagum (1988) and Bobbitt
and Otto (1990) and the articles referenced in these papers.
X-13ARIMA-SEATS is available as an executable program for PC microcomputers (386 or higher with a math
coprocessor) running DOS (version 3.0 or higher), Sun 4 UNIX workstations, and VAX/VMS computers. Fortran
source code is available for users to create executable programs on other computer systems. When it is released,
the X-13ARIMA-SEATS program will be in the public domain, and may be copied or transferred. Computer files
containing the current test version of the program (executables for various machines and source code), this
documentation, and examples, have been put on the Internet at http://www.census.gov/srd/www/x13as/.
1
2 CHAPTER 1. INTRODUCTION
Limited program support is available via regular mail, telephone and email (the preferred mode of communi-
cation) at the addresses given on the title page. If problems are encountered running a particular input file,
providing the input, data and resulting output files will facilitate our identification of the problem.
There are now two seasonal adjustment modules contained in the program. One uses the X-11 seasonal
adjustment method detailed in Shiskin, Young, and Musgrave (1967). The program has all the seasonal adjust-
ment capabilities of the X-11 and X-11-ARIMA programs. The same seasonal and trend moving averages are
available, and the program still offers the X-11 calendar and holiday adjustment routines.
The X-11 seasonal adjustment module has also been enhanced by the addition of several new options,
including
(a) the sliding spans diagnostic procedures, illustrated in Findley, Monsell, Shulman, and Pugh (1990)
(b) the ability to produce the revisions history of a given seasonal adjustment
(c) a new Henderson trend filter routine which allows the user to choose any odd number for the length
of the Henderson filter
(d) new options for seasonal filters
(e) several new outlier detection options for the irregular component of the seasonal adjustment
(f) a table of the trading day factors by type of day
(g) a pseudo-additive seasonal adjustment mode.
The second seasonal adjustment module uses the ARIMA model based seasonal adjustment procedure from
the SEATS seasonal adjustment program developed by Victor Gómez and Agustin Maravall at the Bank of Spain.
All the capabilities of SEATS are included in this version of X-13ARIMA-SEATS , which can generate stability
and spectral diagnostics for SEATS seasonal adjustments in the same way as X-11 seasonal adjustments.
For more details on the SEATS seasonal adjustment method, see Maravall (1995), Gómez and Maravall
(1996), Gómez and Maravall (2001a), and Gómez and Maravall (2001b).
The modeling module of X-13ARIMA-SEATS is designed for regARIMA model building with seasonal economic
time series. To this end, several categories of predefined regression variables are available in X-13ARIMA-SEATS
including trend constants or overall means, fixed seasonal effects, trading-day effects, holiday effects, pulse effects
(additive outliers), level shifts, temporary change outliers, and ramp effects. User-defined regression variables
can also be easily read in and included in models. The program is designed around specific capabilities needed
for regARIMA modeling, and is not intended as a general purpose statistical package. In particular, X-13-
ARIMA-SEATS should be used in conjunction with other (graphics) software capable of producing high resolution
plots of time series.
Observations (data) from a time series to be modelled and/or seasonally adjusted using X-13ARIMA-SEATS
should be quantitative, as opposed to binary or categorical. Observations must be equally spaced in time, and
missing values are not allowed. X-13ARIMA-SEATS handles only univariate time series models, i.e., it does not
estimate relationships between different time series.
X-13ARIMA-SEATS uses the standard (p d q)(P D Q)s notation for seasonal ARIMA models. The (p d q)
refers to the orders of the nonseasonal autoregressive (AR), differencing, and moving average (MA) operators,
respectively. The (P D Q)s refers to the seasonal autoregressive, differencing, and moving average orders.
The s subscript denotes the seasonal period, e.g., s = 12 for monthly data. Great flexibility is allowed in the
specification of ARIMA structures: any number of AR, MA, and differencing operators may be used; missing
lags are allowed in AR and MA operators; and AR and MA parameters can be fixed at user-specified values.
3
For the user who wishes to fit customized time series models, X-13ARIMA-SEATS provides capabilities for the
three modeling stages of identification, estimation, and diagnostic checking. The specification of a regARIMA
model requires specification both of the regression variables to be included in the model and also the type
of ARIMA model for the regression errors (i.e., the orders (p d q)(P D Q)s ). Specification of the regression
variables depends on user knowledge about the series being modelled. Identification of the ARIMA model for
the regression errors follows well-established procedures based on examination of various sample autocorrelation
and partial autocorrelation functions produced by the X-13ARIMA-SEATS program. Once a regARIMA model
has been specified, X-13ARIMA-SEATS will estimate its parameters by maximum likelihood using an iterative
generalized least squares (IGLS) algorithm. Diagnostic checking involves examination of residuals from the
fitted model for signs of model inadequacy. X-13ARIMA-SEATS produces several standard residual diagnostics
for model checking, as well as providing sophisticated methods for detecting additive outliers and level shifts.
Finally, X-13ARIMA-SEATS can produce point forecasts, forecast standard errors, and prediction intervals from
the fitted regARIMA model.
In addition to these modeling features, X-13ARIMA-SEATS has an automatic model selection procedure based
largely on the automatic model selection procedure of TRAMO (Gómez and Maravall 1996, documented in
Gómez and Maravall 2001a). There are also options that use AICC to determine if user-specified regression
variables (such as trading day or Easter regressors) should be included into a particular series. Also, histories
can be generated for likelihood statistics (such as AICC, a version of Akaike’s AIC that adjusts for the length
of the series being modelled) and forecasts to facilitate comparisons between competing models.
The next five chapters detail capabilities of the X-13ARIMA-SEATS program.
• Chapter 2 provides an overview of running X-13ARIMA-SEATS and explains program limits that users
can change.
• Chapter 3 provides a general description of the required input file (specification file), and also discusses
specification file syntax and related issues.
• Chapter 4 discusses the general regARIMA model fit by the X-13ARIMA-SEATS program, summarizes the
technical steps involved in regARIMA modeling and forecasting, and relates these steps to capabilities of
the program.
• Chapter 5 discusses some key points related to model estimation and inference that all users of the
modeling features should be aware of, including some estimation problems that may arise and ways to
address them.
• Chapter 6 discusses some details of key seasonal adjustment diagnostics: spectrums, sliding spans, and
revisions history.
• Chapter 7 gives detailed documentation for each specification statement that can appear in the specifi-
cation file. These statements function as commands that control the flow of X-13ARIMA-SEATS’ execution
and select among the various program options.
The focus in Chapters 2 through 6 is on giving an overview of the use and capabilities of the X-13ARIMA-SEATS
program. In contrast, Chapter 7 is intended as the primary reference to be used when constructing specification
files for running the X-13ARIMA-SEATS program.
4 CHAPTER 1. INTRODUCTION
1.1 Acknowledgements
We are indebted to Statistics Canada, particularly to Estella Dagum, providing us with the source code from
X-11-ARIMA ( Dagum 1980, Dagum 1988 ) to use as the starting point for the seasonal adjustment routines of
X-13ARIMA-SEATS and giving us helpful advice.
We are grateful to Hirtugu Akaike and Makio Ishiguro of the Institute of Statistical Mathematics for per-
mission to incorporate into X-13ARIMA-SEATS the source code of the autoregressive spectrum diagnostics of
BAYSEA (Akaike and Ishiguro 1980).
This version of X-13ARIMA-SEATS would not be possible without the support of Agustin Maravall of the
Bank of Spain and Gianluca Caporello, who provided us the SEATS source code (and subsequent updates) and
were generous with their advice and expertise in integrating SEATS into the X-13ARIMA-SEATS.
Finally, we are indebted to Victor Gómez for providing us with the Fortran code of TRAMO (Gómez and
Maravall 2001a) to enable us to implement an automatic modeling procedure very similar to TRAMO’s in X-13-
ARIMA-SEATS and to Agustı́n Maravall, Gianluca Caporello, and their unit at the Bank of Spain for updates to
the TRAMO and SEATS source code and advice.
2 Running X-13ARIMA-SEATS
Contents
2.1 Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Input errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Specifying an alternate output filename . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Running X-13ARIMA-SEATS on more than one series . . . . . . . . . . . . . . . . . . 7
2.5.1 Running X-13ARIMA-SEATS in multi-spec mode . . . . . . . . . . . . . . . . . . . . . . 8
2.5.2 Running X-13ARIMA-SEATS in single spec mode . . . . . . . . . . . . . . . . . . . . . . 9
2.5.3 Special Case: File Names Containing Spaces . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Log Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.7 Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.8 Program limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Tables
2.1 X-13ARIMA-SEATS Program Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 X-13ARIMA-SEATS Program Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Procedures for installing X-13ARIMA-SEATS are machine-specific; information about this is provided with the
program, and is also available on the Internet at http://www.census.gov/srd/www/x13as/. Having installed
the program on a microcomputer running a DOS operating system, a generic statement to run X-13ARIMA-SEATS
is
path\x13as path\filename
In this statement path\filename.spc is the main X-13ARIMA-SEATS input (specification) file. The program
created a file named path\filename.out as an output file. The path to X-13ARIMA-SEATS is necessary if the file
containing the X-13ARIMA-SEATS program is not in the current directory; similarly for the path to the input file
filename.spc.
Note that only the filename is specified, not the extension; the program will use the filename provided at
runtime to form the filenames for all files generated by the program. For an X-13ARIMA-SEATS run using the
spec file filename.spc, the output will be stored in the file filename.out, the error messages will be stored in the
file filename.err, etc. Thus, if the spec file xuu1.spc is in a PC’s current directory, typing
x13as xuu1
5
6 CHAPTER 2. RUNNING X-13ARIMA-SEATS
and pressing the <return> (or <enter> key) will cause the program to run and create files xuu1.out and
xuu1.err in the current directory.
Program input and output are both discussed briefly below, and more extensively in the documentation that
follows. To run the program under a UNIX (or Linux) operating system, substitute (forward) slashes for the
backslashes in the generic statements above. To run X-13ARIMA-SEATS under other operating systems, specify
paths, etc., using the syntax appropriate for the system. For the DOS, UNIX/Linux and VAX/VMS operating
systems, a quick reference document is also available, giving more detailed instructions on the syntax for running
X-13ARIMA-SEATS in these operating systems.
2.1 Input
To apply X-13ARIMA-SEATS to any particular time series, a main input file, called a specification file, must be
created. This ASCII (or “text”) file contains a set of specifications or specs that X-13ARIMA-SEATS reads to
obtain the information it needs about the time series data, the time series model to be used, the analysis to
be performed, and the output desired. X-13ARIMA-SEATS assumes that the specification file has the extension
.spc. Thus path\filename is sufficient in the above statements. The only input files other than the spec file
that X-13ARIMA-SEATS may need are optional files containing data for the time series being modelled, data for
any user-defined regression variables, values for any user-defined prior-adjustment factors, and model types to
try with the automatic model selection procedure from the pickmdl spec. The names of these files (including
paths) are provided to X-13ARIMA-SEATS by listing them in appropriate specs in the spec file. The use of
such additional input files is optional because the user can alternatively include the data values required in
appropriate places in these specs, and a default set of models for the automatic modeling procedure is available.
Section 7 explains how to write spec files.
2.2 Output
The usual output is written to the file path\filename.out. Individual specs control their contribution to
this output using optional print arguments (discussed in Section 3.2). The save argument is used to create
certain other output files for further analysis (for example, to save a time series of residuals for plotting using
a graphics program). Cautionary note: When save is used, the program constructs the name of the file to
which the specified output is written using naming conventions discussed in Section 3.2. If a file with this name
already exists, it will be overwritten by X-13ARIMA-SEATS and the contents lost. Users should thus take suitable
precautions when saving output. See Section 3.2 for more information.
2.3 Input errors
Input errors are reported as they are discovered by the program, which then prints appropriate error messages.
These error messages are also stored in a file named path\filename.err. When the program can localize the
error, the line in the spec file containing the error will be printed out with a caret (^) positioned under the
error. If the program cannot localize the error, then only the error message will be printed. If the error is
2.4. SPECIFYING AN ALTERNATE OUTPUT FILENAME 7
fatal, then ERROR: will be displayed before the error message, sometimes with suggestions about what to
change. For nonfatal errors, WARNING: will be printed before the message. WARNING messages are also
used sometimes to call attention to a situation in which no error has been committed, but some caution is
appropriate.
X-13ARIMA-SEATS first reads the whole spec file, reporting all input errors it finds. This way the user can
try to correct more than one input error per run. Frequently, however, the only informative messages are those
for the first one or two errors. These errors may result in other errors, especially if input errors occur in the
series spec. The program will stop if any fatal errors are detected. Warnings will not stop the program, but
should alert users to check both the input and output carefully to verify that the desired results are produced.
2.4 Specifying an alternate output filename
As was noted before, for an X-13ARIMA-SEATS run using the spec file filename.spc, the output will be stored in
the file filename.out, the error messages will be stored in the file filename.err, etc. For the purpose of examining
the effects of different adjustment and modeling options on a given series, it is sometimes desirable to use a
different filename for the output than was used for the input. The general form for specifying an alternate
filename for the output files is
path\x13as path\filename path\outname (2.1)
This X-13ARIMA-SEATS run still uses the spec file filename.spc, but the output will be stored in the file
outname.out, the error messages will be stored in the file outname.err, etc. All output files generated by this
run will be stored using the path and filename given by the user, not the path and filename of the input
specification file.
2.5 Running X-13ARIMA-SEATS on more than one series
In a production situation, it is essential to run more than one series in a given X-13ARIMA-SEATS run. X-13-
ARIMA-SEATS allows for running multiple series in two modes:
(a) multi-spec mode, where there are input specification files for every series specified;
(b) single spec mode, where every series will be run with the options from a single input specification
file.
Before X-13ARIMA-SEATS can be run in either mode, a metafile must be created. This is an ASCII file
which contains the names of the files to be processed. Two types of metafiles are used by the X-13ARIMA-SEATS
software: input metafiles (for multi-spec mode) and data metafiles (for single spec mode).
If an error occurs in one of the spec files in a metafile run, the program will print the appropriate error
messages. Execution will stop for that series and the program will continue processing the remaining spec files.
A listing of all the input files with errors is given in the X-13ARIMA-SEATS log file, described in Section 2.7.
2.5.1 Running X-13ARIMA-SEATS in multi-spec mode

Before X-13ARIMA-SEATS can be run in multi-spec mode, an input metafile must first be created. This is an
ASCII file which contains the names of the files to be processed by X-13ARIMA-SEATS in sequence. An input
metafile can have up to two entries per line: the filename (and path information, if necessary) of the input
specification file for a given series, and an optional output filename for the output of that series. If an output
filename is not given by the user, then the path and filename of the input specification file will be used to
generate the output files. The input specification files are processed in the order in which they appear in the
input metafile.
For example, to run the spec files xuu1.spc, xuu2.spc and xuu3.spc, the input metafile should contain the
following:
xuu1
xuu2
xuu3
This assumes that all these spec files are in the current directory. To run these files if they are stored in the
c:\export\specs DOS directory, the metafile should read:
c:\export\specs\xuu1
To run X-13ARIMA-SEATS with a input metafile, use the following syntax:
x13as -m metafile
where metafile.mta is the metafile and -m is a flag which informs X-13ARIMA-SEATS of the presence of a
metafile.
For example, if the metafile defined above is stored in exports.mta, type
x13as -m exports
and press the return key to run the corresponding spec files.
Note that when the name of the input metafile was given in the example above, only the filename was
specified, not the extension; .mta is the required extension for the input metafile. Path information should be
included with the input metafile name, if necessary.
The filenames used by X-13ARIMA-SEATS to generate output files are taken from the spec files listed in the
metafile, not from the metafile itself. The example given above would generate output files named xuu1.out,
xuu2.out and xuu3.out corresponding to the individual spec files given in the metafile exports.mta, not a
comprehensive output file named exports.out. To specify alternate output filenames for the example above,
simply add the desired output filenames to each line of the input metafile, e.g.,
c:\export\specs\xuu1 c:\export\output\xuu1
2.5. RUNNING X-13ARIMA-SEATS ON MORE THAN ONE SERIES 9
2.5.2 Running X-13ARIMA-SEATS in single spec mode
To run X-13ARIMA-SEATS on many series using the same specification commands for each series, it is necessary
to create a data metafile. A data metafile can have up to two entries per line: the complete filename (and path
information, if necessary) of the data file for a given series, and an optional output filename for the output of
that series. If an output filename is not given by the user, then the path and filename of the data file will be
used to generate the output files. Note: In a data metafile, no extension is assumed for the individual data
files. The extensions must be specified, along with the path and filename, if the data files are not in the current
directory.
The data files are processed in the order in which they appear in the data metafile. The options used to
process each data file are provided by a single input specification file identified at runtime. This means that
all the data files specified in the data metafile must be in the same format. Also, certain formats supported by
X-13ARIMA-SEATS should be avoided; see the description of the series spec in Section 7.15 for more details.
For example, to process the data files xuu1.dat, xuu2.dat and xuu3.dat, the data metafile should contain
the following:
xuu1.dat
xuu2.dat
xuu3.dat
This assumes that all these data files are in the current directory. To run these files if they are stored in the
c:\export\data DOS directory, the metafile should read:
c:\export\data\xuu1.dat
To run X-13ARIMA-SEATS with a data metafile, use the following syntax:
x13as specfile -d metafile
where metafile.dta is the data metafile, -d is a flag which informs X-13ARIMA-SEATS of the presence of a data
metafile, and specfile.spc is the single input specification file used for each of the series listed in the data
metafile.
For example, if the data metafile with three series used for illustration above is named exports.dta, type
x13as default -d exports
and press the return key to process the corresponding data files using the default.spc input specification file.
Note that when the name of the data metafile was given in the example above, only the filename was
specified, not the extension; .dta is the required extension for the input metafile. Path information should be
included with the data metafile name, if necessary.
The filenames used by X-13ARIMA-SEATS to generate output files are taken from the data files listed in the
metafile, not by the metafile itself. The example given above would generate output files named xuu1.out,
xuu2.out and xuu3.out corresponding to the individual data files given in the metafile exports.dta, not a
comprehensive output file named exports.out. To specify alternate output filenames for the example above,
simply add the desired output filenames to each line of the data metafile, e.g.,
c:\export\data\xuu1.dat c:\export\output\xuu1
2.5.3 Special Case: File Names Containing Spaces
In many current operating systems, it is permissable to have blank spaces in file names or paths - for example,
c:\My Spec Files\test.spc. When specifying such a file in an input or data metafile, the user must enclose
the entire filename with quotation marks ("). Otherwise, the program will assume that the first entry in the
metafile is only the text up to the first space.
For example, if the specfiles used in the second example in Section 2.5.1 were stored in the c:\export specs
DOS directory, then the input metafile should read:
"c:\export specs\xuu1"
Running X-12-ARIMA on the input metafile given above would generate output files named xuu1.out,
xuu2.out and xuu3.out in the c:\export specs directory.
This convention applies to data metafiles and alternate output filenames provided in metafiles as well. The
following data metafile would read data files from the directory c:\export data and store the output files into
the directory c:\export output
"c:\export data\xuu1.dat" "c:\export output\xuu1 a"

Running X-12-ARIMA on the data metafile given above would generate output files named xuu1 a.out,
xuu2 a.out and xuu3 a.out in the c:\export output directory.
Be careful that the opening and closing quotation marks fully contain the filenames with no extra spaces,
and that there are matching opening and closing quotation marks for each file.
2.6. LOG FILES 11
2.6 Log Files

Every time X-13ARIMA-SEATS is run, a log file is produced where a summary of modeling and seasonal ad-
justment diagnostics can be stored for every series or spec file processed. When X-13ARIMA-SEATS is run in
multi-spec or single spec model, as described in the previous section, the log file is stored with the same name
and directory as the metafile (for multi-spec mode) or data metafile (single spec mode), with an extension of
.log. For example
x13as -m exports
runs each of the spec files stored in exports.mta and stores user-selected diagnostics into the log file exports.log.
If only one series is processed, the output directory and filename is used along with the .log file extension
to form the name of the log file.
Users can specify which diagnostics are stored in the log file by using the savelog argument found in
the automdl, check, composite, estimate, history, pickmdl, regression, seats, series, slidingspans,
transform, x11, and x11regression specs. The descriptions of the individual specs in Section 7 give more
details on which diagnostics can be stored in the log file.
As mentioned in the previous section, if an error occurs in one of the spec files in a metafile run, a listing of
all the input files with errors is given in the log file.
2.7 Flags
In the previous section, the flags -m and -d were required in the command line to obtain the desired run. There
are several other input and output options that are specified on the command line. The general syntax for the
command line can be given as
path\x13as arg1 arg2 . . . argN
where the arguments given after x13as can be either flags or filenames, depending on the situation.
Table 2.1 gives a summary of the flags available in X-13ARIMA-SEATS the remainder of this section will
describe what each flag means in more detail. These flags can be specified in any order on the command line.
(Some must be followed by appropriate filenames).
The -m and -d flags were described in the previous section. Note that one cannot specify both of these flags
in the same run.
The -i flag indicates that the next argument is the path and filename of the input specification file. This
flag does not need to be specified as long as the input specification file is the first argument; therefore, x13as
test and x13as -i test are equivalent. The -i and -m flags cannot be specified in the same run.
Similar to -i, the -o flag indicates that the next argument is the path and filename for the output. The
output extensions described earlier (.out and .err) as well as extensions associated with the save command
will be appended to this filename. This flag also does not need to be specified as long as the input specification
file is the first argument and the output filename is the second argument (as in Equation 2.1). So any of the
following commands are equivalent:
Table 2.1: X-13ARIMA-SEATS Program Flags
flags description of flags

-c Sum each of the components of a composite adjustment, but
only perform modeling or seasonal adjustment on the total
-d filename Filename (without extension) for data metafile
-g dirname Directory where graphics metafile and related files for input
to external graphics programs are stored
-i filename Filename (without extension) for input specification file
-m filename Filename (without extension) for input metafile
-n (No tables) Only tables specifically requested in the input
specification file will be printed out
-o filename Filename (without extension) used for all output files gener-
ated during this run of the program
-p No pagination is used in main output file
-q Run X-13ARIMA-SEATS in quiet-mode (warning messages not
sent to the console)
-r Produce reduced X-13ARIMA-SEATS output (as in GiveWin
version of X-12-ARIMA)
-s Store seasonal adjustment and regARIMA model diagnostics
in a file
-v Only check input specification file(s) for errors; no other pro-
cessing
-w Wide (132 character) format is used in main output file
x13as test test2

x13as -i test -o test2
x13as -o test2 -i test
However, x13as -i test test2 will generate an error, since the first argument is the flag -i, not the spec
file. The -o flags cannot be specified in the same run as the -m or -d flags. The -o and -m flags cannot be
specified in the same run.
For operating systems that allow blank spaces in file names, the convention for specifying a file name as a
flag is similar to that specified in Section 2.5.3. All filenames with at least one space in the filename or path
should be enclosed in quotation marks (").
So any of the following commands should execute correctly:
x13as "c:\My Spec Files\test" "c:\My Output\test2"

x13as -i "c:\My Spec Files\test" -o "c:\My Output\test2"
x13as -o "c:\My Output\test2" -i "c:\My Spec Files\test"
x13as -m "c:\My Spec Files\alltest"
x13as "c:\My Spec Files\testsrs" -d "c:\My Data Files\testsrs"
2.7. FLAGS 13
The -s flag specifies that certain seasonal adjustment and regARIMA modeling diagnostics that appear in
the main output be saved in file(s) separate from the main output. These include tables in the main output file
that are not tables of time series. Such tables cannot be stored in the format used for individual time series
tables. When the -s flag is used, X-13ARIMA-SEATS automatically stores the most important of these diagnostics
in a separate file that can be used to generate diagnostic summaries. This file (called the diagnostics summary
file) will have the same path and filename as the main output, with the extension .udg. So for
x13as test -s
the diagnostics summary file will be stored in test.udg, and for
x13as test -s -o testout
the diagnostics summary file will be stored in testout.udg.

The diagnostics summary file is an ASCII database file. Within the diagnostic file, each diagnostic has a
unique key to access it’s value. The key is separated from the diagnostic value by a colon (’:’), followed by white
space. So in the entry below:
freq: 12
The key for this entry would be freq, and the value for the key would be 12. Each record in the file provides
a value for a unique key found at the beginning of the line.
User-defined metadata can be stored in the diagnostics summary file (for more details, see the description
of the metadata spec in Section 7.9).
A program is available via the Internet at http://www.census.gov/srd/www/x13as/ that reads the seasonal
adjustment diagnostics file and produces a summary of the seasonal adjustment diagnostics. This program is
written in the Icon programming language (see Griswold and Griswold 1997).
The -g flag indicates that the next argument is the complete path name of a directory into which output
will be stored that is intended as input for a separate graphics program. This output consists of the following
files:
(1) files of diagnostic data to be graphed, which are produced by the options specified in the .spc file;
(2) a graphics metafile containing the names of these files;
(3) a diagnostics summary file containing information about the time series being processed, about the
regARIMA model fit to the series (if any), and about the seasonal adjustment requested (if any);
The graphics metafile carries the extension .gmt and the diagnostics summary file carries the extension
.udg; these files carry the filename used for the main program output. For example, if a user enters
x13as test -g c:\sagraph

the graphics metafile will be stored in c:\sagraph\test.gmt and the diagnostics summary file will be stored
in c:\sagraph\test.udg. For
x13as test -g c:\sagraph -o testout
the graphics metafile will be stored in c:\sagraph\testout.gmt and the diagnostics summary file will be stored
in c:\sagraph\testout.udg. In both cases, related files needed to generate seasonal adjustment graphics will
be also be stored in the c:\sagraph subdirectory. (NOTE: The directory entered after the -g flag must already
have been created and should be different from the directory used for the output files; it can be a subdirectory
of the latter.)
The Census Bureau website (http://www.census.gov/srd/www/x13graph/) distributes two versions of X-
13-Graph (Lytras 2012a, 2012b), the successor to X-12-Graph (Hood 2002a, 2002c). X-13-Graph is a program
that produces graphs from the graphics mode output of X-13ARIMA-SEATS. For examples of the use of X-13-
Graph, see Findley and Hood (1999). For a list of the codes that X-13ARIMA-SEATS stores when run in graphics
mode, along with the abbreviations used in the graphics metafile to denote these codes, see Appendix A. One
version of the program is X-13-Graph Batch (Lytras 2012a), which uses SAS/GRAPH R
(SAS Institute Inc.
1990) to produce the graphs; this version has a companion Windows interface to simplify selection of graph
options. The second version of the program is X-13-Graph Java, which is written entirely in Java (Lytras
2012b).
If both the -g and -s options are used in the same X-13ARIMA-SEATS run, the complete version of the seasonal
adjustment diagnostics file will be stored in the directory specified by the -g option (and not in the directory
of the main output file). If a model diagnostics file is also generated, that file will be stored in the graphics
directory as well. A warning message is written to the screen and to the log file telling the user that the seasonal
adjustment diagnostics file (and the model diagnostics file, if it is produced) is in the graphics directory.
The -n, -w, -p, and -r flags all affect the format of program output. The -n option allows the user to
restrict the number of tables appearing in the main output file. The X-13ARIMA-SEATS program produces a
large number of tables in the main output file. While X-13ARIMA-SEATS is flexible in allowing users to determine
which tables are to be printed out, it is sometimes convenient to restrict the output to only a few tables. To
facilitate this, the -n flag specifies that, as the default, no tables will be written to the main output file. Then
only those tables specified by the user in the spec file are written.
The -w flag specifies that a wide (132 character) format is used to print out tables in the main output file.
The default is an 80 character tabular format. The exact format of the output tables is determined by the
magnitude of the series values and by what degree of precision is requested in the series spec.
The -p flag specifies that page breaks and headers will be suppressed in the main output file. If this option
is not specified, then page breaks will be inserted at the beginning of each table of output, along with a title for
the run, series name, and page number.
The -r flag specifies that output tables and headers will be written in a format that will reduce the amount
of output printed out to the main output file. The tables printed out are consolidated, and some blank lines in
the printout are suppressed. This output option was first utilized in the version of X-13ARIMA-SEATS developed
for use with the GiveWin econometrics package (see Doornik and Hendry 2001).
The -q flag specifies that X-13ARIMA-SEATS will be run in “quiet mode”. Warning messages that are normally
printed to the console are suppressed, although error messages shall still be printed to the console. All warning
messages not printed to the screen will be stored in the error file (see Section 2.3).
2.8. PROGRAM LIMITS 15
The -c flag is used only to restrict a composite seasonal adjustment run done with an input metafile (-
m). In a composite seasonal adjustment, X-13ARIMA-SEATS usually seasonally adjusts a set of component time
series, as well as their composite (also called aggregate), which is usually their sum (for more details, see the
description of the composite spec in Section 7.4). An input specification file is needed for each series. When
-c is invoked, the seasonal adjustment and modeling options specified in the input spec files for the component
series are ignored; the component series are only used to form the composite series. This option is useful when
identifying a regARIMA model for the composite series.
Finally, the -v flag specifies that X-13ARIMA-SEATS will be run in an input verification mode to enable the
user to see if there are errors in one or more input spec files. This allows the user to check the program options
for errors without doing the complete X-13ARIMA-SEATS runs for all the series. The -v flag cannot be used with
the -s, -c, -n, -w, or -p flags.
2.8 Program limits

The X-13ARIMA-SEATS Fortran source code contains limits on the maximum length of series, maximum number
of regression variables in a model, etc. These limits are set at values believed to be sufficiently large for the
great majority of applications, without being so large as to cause memory problems or to significantly lengthen
program execution times.
Table 2.2 details those parameter variables in the model.prm and srslen.prm files corresponding to X-13-
ARIMA-SEATS program limits that are subject to user modification.
The limits may be modified if required, but the Fortran source code of the program must then be recompiled
and relinked to put the new limits into effect. The limits potentially requiring modification for this purpose
occur in parameter statements in the files model.prm and srslen.prm. We suggest keeping a backup copy of
the original files, in case problems arise from an attempt to modify program limits.
Table 2.2: X-13ARIMA-SEATS Program Limits
parameter value
variable (limit) description of parameter
pobs 780 maximum length of the series on input. The number,
pobs + pfcst (see below), is the maximum length
of input series of user-defined regression variables and
user-defined prior adjustment factors—the additional
pfcst values are allowed to accommodate values of
regression variables or adjustment factors in a possible
forecast period
pyrs 85 maximum number of years in the forecast and backcast
extended series
psp 12 maximum seasonal period, i.e., observations more fre-
quent than psp times per year are not allowed
pfcst 120 maximum number of forecasts
pb 80 maximum number of regression variables in a model
(including predefined and user-defined regression vari-
ables specified, plus any regression variables generated
by automatic outlier detection or an AIC test)
pureg 52 maximum number of user-defined regression variables
porder 36 maximum lag corresponding to any AR or MA param-
eter
pdflg 3 maximum number of differences in any ARIMA factor
(nonseasonal or seasonal)
psrs 5000 maximum number of files that can be processed by a
metafile
3 The Specification File and Its Syntax
Contents
3.1 Examples of Input Specification Files . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2 Print and save . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Dates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 General rules of input syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Tables
3.1 X-13ARIMA-SEATS Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
The main input to X-13ARIMA-SEATS comes from a special input file called an input specification file. This
file contains a set of specifications or specs that give X-13ARIMA-SEATS various information about the data and
the desired seasonal adjustment options and output, the time series model to be used, if any, etc. Table 3.1
describes the different specs that are currently available in the X-13ARIMA-SEATS program.
Each spec is defined in the spec file by its name, which is followed by braces { } containing arguments
and their assigned values. The arguments and their value assignments take the form argument = value, or, if
multiple values are required, argument = (value1 , value2 , . . . ). There are various types of values: titles, variable
names, keywords, numerical values, and dates. These are defined and illustrated in the documentation of the
individual specs in Chapter 7. Because of their occurrence in several specs, detailed discussions of the print
and save arguments (Section 3.2), and date argument values (Section 3.3) are given below.
There are no required arguments for any spec other than either series or composite (see below). Most
arguments have default values; these are given in the documentation of each spec. Default values for all
arguments are used if no arguments are specified.
Typically, not all specs are included in any one spec file. In fact, for most X-13ARIMA-SEATS runs (any that
is not a composite run) there is only one required spec in the specification file—the series spec. This spec must
include either the data or file argument. (The only exception is when a data metafile is used; see Section 2.5.2
for more details.) Thus, X-13ARIMA-SEATS will accept the minimal spec file series {data=( data values )} .
However, this spec file produces no useful output.
For seasonal adjustment runs, the x11 spec is needed, unless one or more of the force, x11regression,
slidingspans, or history (with the estimates argument set to perform a seasonal adjustment history) specs are
present. In this case, X-13ARIMA-SEATS behaves exactly as if the x11 spec were present with default arguments,
which is equivalent to including x11{} in the spec file.
For model identification runs, the identify spec is needed. For model estimation, the arima and/or re-
gression specs, and the estimate spec are ordinarily included. If the estimate spec is absent, but one or more
of the outlier, automdl, pickmdl, check, forecast, x11, slidingspans and history specs is present, this
17
18 CHAPTER 3. THE SPECIFICATION FILE AND ITS SYNTAX
Table 3.1: X-13ARIMA-SEATS Specifications
series a required spec except when composite adjustment is done. It specifies the
time series data, start date, seasonal period, span to use in the analysis,
and series title,
composite specifies that both a direct and an indirect adjustment of a composite
series be performed; it is used instead of the series spec,
transform specifies a transformation and/or prior adjustment of the data,
x11 specifies seasonal adjustment options, including mode of adjustment, sea-
sonal and trend filters, an Easter holiday adjustment option, and some
seasonal adjustment diagnostics,
x11regression specifies irregular regression options, including which regressors are used
and what type of extreme value adjustments will be made to robustify the
regression on the irregular component,
seats specifies options to perform an ARIMA model-based seasonal adjustment
as in SEATS, a seasonal adjustment methodology developed by Victor
Gómez and Agustin Maravall (see Gómez and Maravall (1996)),
force specifies options to force the totals of the seasonally adjusted series to be
the same as the original series,
automdl specifies an automatic model selection procedure based on TRAMO (see
Gómez and Maravall 1996 and Gómez and Maravall 2001a,
pickmdl specifies an automatic model selection procedure based on X-11-ARIMA
(see Dagum 1988),
arima specifies the ARIMA part of the regARIMA model,
regression specifies regression variables used to form the regression part of the reg-
ARIMA model, and to determine the regression effects removed by the
identify spec,
estimate requests estimation or likelihood evaluation of the model specified by the
regression and arima specs, and also specifies estimation options,
check produces statistics useful for diagnostic checking of the estimated model,
forecast specifies forecasting with the estimated model,
outlier specifies automatic detection of additive outliers and/or level shifts using
the estimated model. There is an optional test for temporary level shifts,
identify produces autocorrelations and partial autocorrelations for specified orders
of differencing of the data with regression effects (specified by the regres-
sion spec) removed for ARIMA model identification,
slidingspans specifies that a sliding spans analysis of seasonal adjustment stability be
performed,
history requests the calculation of a historical record of seasonal adjustment revi-
sions and/or regARIMA model performance statistics.
metadata allows users to specify metadata keys and values for storage in the diag-
nostics summary file.
spectrum allows users to specify options related to the spectral plots generated by
the program.
3.1. EXAMPLES OF INPUT SPECIFICATION FILES 19
forces estimation of the specified model. In this case, X-13ARIMA-SEATS behaves exactly as if the estimate spec
were present with default arguments, which is equivalent to including estimate{} in the spec file. If the arima
spec is absent, estimation proceeds with the default ARIMA(0 0 0) model (white noise). This is equivalent to
including arima{} in the spec file.
The order of the specification statements in the spec file (with one exception), and the order of arguments
within the braces of any spec do not matter. The only requirement is that series or composite must be the
first spec. The spec file is free format, and blank spaces, tabs, and blank lines may be used as desired to make
the spec file more readable. Comments can also be included. The use of comments and other general rules
governing input syntax are discussed in Section 3.4. Important: There must be a carriage return at the end
of the last line of the spec file, otherwise, this line will not be read. This is a Fortran requirement.
3.1 Examples of Input Specification Files
A very simple spec file producing a default X-11 run is given in Example 3.1. The spectrum diagnostics in the
output file of this run indicated the presence of a trading day component, and a message saying this was written
in the output. A regARIMA model can be used to both estimate the trading effect and to extend the series by
forecasts prior to seasonal adjustment.
Example 3.1: X-13ARIMA-SEATS spec file for a default X-11 run
series { title = "Monthly Retail Sales of Household Appliance Stores"

data = ( 530 529 526 532 568 785 543 510 554 523 540 599
574 619 619 600 652 877 597 540 594 572 592 590
632 644 621 604 613 828 578 533 582 605 660 677
682 684 700 705 747 1065 692 654 719 690 706 759
769 730 740 765 791 1114 695 680 788 778 780 805
852 823 831 836 913 1265 726 711 823 780 844 870
865 915 920 935 1030 1361 859 852 954 895 993 1109
1094 1173 1120 1159 1189 1539 1022 987 1024 1005 1054 1098
1191 1191 1161 1201 1294 1782 1154 1059 1178 1126 1120 1233
1260 1311 1302 1365 1395 1899 1123 1087 1210 1157 1159 1260
1357 1265 1231 1287 1452 2186 1309 1242 1388 1400 1397 1527
1654 1650 1555 1560 1836 2762 1541 1480 1619 1455 1510 1698
1651 1749 1783 1863 2074 3051 1836 1690 1856 1796 1904 1927
1978 2055 1976 2204 2423 3502 1977 1767 1935 1900 2073 2143
2299 2247 2162 2274 2529 3731 2184 1901 2058 1974 2018 2091
2239 2253 2157 2190 2397 3659 2170 2086 2297 2251 2311 2520
)
start = 1972.jul }
x11{ }
Examples 3.2 and 3.3 illustrate spec files that might be used to identify the ARIMA part of the model before
the final seasonal and trading day adjustment is achieved in Example 3.4. Alternatively, the X-11 trading day
adjustment procedures described in Section 7.20 could be used.
It is customary to make at least two runs of X-13ARIMA-SEATS when modeling a time series. The first run
is usually done to permit identification of the ARIMA part of the model; the second run is done to estimate
and check the regARIMA model, and possibly to use it in forecasting the series. The spec file for the first run
requires the series and identify specs, and may also include the transform and regression specs. The spec
file for the second run includes the series, arima, and estimate specs; possibly the transform and regression
specs; and the outlier, check, and forecast specs as desired. The two runs of X-13ARIMA-SEATS require two
different spec files, or, more conveniently, the spec file from the first run can be modified for use in the second
run. If diagnostic checking suggests changes need to be made to the estimated model, then the spec file can be
modified again to change the model for a third run of the program.
The contents of a typical spec file for the model identification run might follow the same format as Example
3.2.
Example 3.2: X-13ARIMA-SEATS spec file for regARIMA model identification
series{title = "Monthly Retail Sales of Household Appliance Stores"

data = ( 530 529 526 532 568 785 543 510 554 523 540 599
574 619 619 600 652 877 597 540 594 572 592 590
.
.
.
2239 2253 2157 2190 2397 3659 2170 2086 2297 2251 2311 2520)
start = 1972.jul}
transform{function = log}
regression{variables = td} # Comment: Series has trading-day effects
identify{diff=(0, 1) sdiff = (0, 1)}
This spec file includes the series, transform, regression, and identify specs. It provides X-13ARIMA-SEATS
with the data given in the series spec, takes the logarithm of the series (transform spec), and specifies regression
variables (regression spec) known or suspected to affect the series. Here, variables = td includes the six
trading-day contrast variables (td6) in the model and also adjusts the series for leap year effects. (See Section
4.3 and the documentation of the regression spec in Section 7.13.) The identify spec performs a regression
of the differenced transformed series (also adjusted for length-of-month effects) on the differenced regression
variables (the six trading-day variables). The regression uses the highest order of seasonal and nonseasonal
differencing specified, (1 − B)(1 − B 12 ). The identify spec then computes a regression residual series for the
undifferenced data from which it produces tables and line printer plots of the sample autocorrelation and partial
autocorrelation functions for all combinations of seasonal and nonseasonal differencing specified (here, four sets
of ACFs and PACFs).
After studying the output from the first run and identifying the ARIMA part of the model as, for example,
(0 1 1)(0 1 1)12 , the identify spec is commented out and the arima and estimate specs are added to the spec
file. The resulting spec file is given in Example 3.3 (the data are not reproduced in full).
This spec file includes the series, transform, regression, arima, and estimate specs. It specifies (re-
gression and arima specs) and fits (estimate spec) the following model:
6
X
12
(1 − B)(1 − B ) yt − βi Tit = (1 − θB)(1 − ΘB 12 )at ,
i=1
3.1. EXAMPLES OF INPUT SPECIFICATION FILES 21
Example 3.3: X-13ARIMA-SEATS spec file for regARIMA model estimation

data = ( 530 529 526 532 568 785 543 510 554 523 540 599
574 619 619 600 652 877 597 540 594 572 592 590
.
.
.
2239 2253 2157 2190 2397 3659 2170 2086 2297 2251 2311 2520)
start = 1972.jul}
# identify{diff=(0, 1) sdiff = (0, 1)}
arima{model = (0,1,1)(0,1,1)}
estimate{print = iterations}
where the Tit are the six trading-day regression variables. The series yt being modelled consists of the logarithms
of the original data adjusted for leap-year effects. If diagnostic checking of residuals, outlier detection, or
forecasting were desired, the appropriate specs would need to be added to the spec file.
Assuming this is a satisfactory model, a seasonal adjustment utilizing forecast extension can be performed
by adding the x11 and forecast to the input specification file. Such a spec file appears in Example 3.4 (the
data are not reproduced in full).
Example 3.4: X-13ARIMA-SEATS spec file for seasonal adjustment

data = ( 530 529 526 532 568 785 543 510 554 523 540 599
574 619 619 600 652 877 597 540 594 572 592 590
.
.
.
2239 2253 2157 2190 2397 3659 2170 2086 2297 2251 2311 2520)
start = 1972.jul}
# identify{diff=(0, 1) sdiff = (0, 1)}
arima{model = (0,1,1)(0,1,1)}
estimate{print = iterations}
forecast{maxlead = 60}
x11{seasonalma = s3x9}
The spec file now generates seasonal adjustments from 3 × 9 seasonal filters (x11) for the trading day pre-
adjusted series. The pre-adjusted series is extended by 60 forecasts (forecast) prior to seasonal adjustment.
The main output file will contain some diagnostics concerning the quality of the seasonal adjustment. Additional
diagnostics can be specified by including the appropriate specs described in Chapter 7.
3.2 Print and save

Control of the output from X-13ARIMA-SEATS is achieved within individual specs by using the print and save
arguments. The print argument controls the given spec’s output to the main output file, while the save
argument allows certain output tables to be written to files. For ease of reference we refer to all the individual
parts of the output subject to control through print and save as “tables”, even though some of this output
(e.g., line printer plots of an ACF) is not in a form that is ordinarily thought of as a table. The tables subject to
control through print and save are listed with their default print status and file extensions (for savable tables)
under the documentation of the print and save arguments for each spec. Tables output to files using save are
written in a format with high numerical precision and with minimal or no labelling information to facilitate
their use for further analysis utilizing other software. Saved tables are also given a consistent format—a single
tab separates fields.
Default output from a spec is written to the main output file if the print argument is absent, or if
print=default or print=() appears in the spec. To stop a spec from writing output to the main output
file, set print=none. (Note: A few specs write some minor labelling information to the screen even with
print=none.) To have all the available output tables and plots for a spec written to the main output file,
set print=all. To have all the available output tables (no plots) for a spec written to the main output file,
set print=alltables. To have a small subset of the available output tables for a spec written to the screen,
set print=brief. Individual tables may be added to the default, brief, and none print levels by including
their names as print argument values. These may (but need not) be preceded by a +. For example, in the
estimate spec, print=(+iterations +residuals), which is equivalent to print = (default +iterations
+residuals), requests printing of results from the estimation iterations and the residuals from the estimated
model, in addition to the default output. Using print=(none estimates) requests printing of only the param-
eter estimates. Individual tables may be suppressed from the default and all print levels by including their
names preceded by a - as print argument values, e.g., print=brief -acf or print=(all -iterations).
If the user wishes to save any output tables to files, these must be specifically listed in the save arguments
of the appropriate specs, e.g., save=(mdl estimates) in the estimate spec. Those tables that are savable
may be specified in the print and save arguments using either a “long” name, the name listed in the spec’s
description, or a “short” 3-letter name, which is the same as the file extension used if the table is saved. For
example, the optional table regcmatrix in the estimate spec can also be specified as rcm. The keywords none,
all, alltables, default, and brief defined above are not available for use in the save argument. Also, names
of tables to be saved should not be preceded with a + or -. Not all tables are savable, and not all specs produce
savable tables.
The save argument writes the specified tables to individual files. A saved file will be placed in the same
directory as the output, and will be given the filename of the main output file, but with a distinct 3-letter
extension. If a file with this name already exists, it will be overwritten. The extensions used are listed under
the documentation of the print and save arguments for each spec. For example, suppose X-13ARIMA-SEATS is
run (on a DOS machine) from the directory C:\TSERIES using as input a spec file stored in SALES.SPC in that
directory. If the estimate spec contains save = (mdl estimates), the resulting saved tables of the model and
parameter estimates will be written to the files C:\TSERIES\SALES.MDL and C:\TSERIES\SALES.EST. If files
with these names already exist, they will be overwritten. Although the extensions used by X-13ARIMA-SEATS
have been chosen to avoid obvious conflicts (examples of extensions not used are .dat, .exe, .com, .for, .spc),
users should still exercise caution to prevent unintended overwriting of files by X-13ARIMA-SEATS saves. A list
of the files saved, with an * indicating those overwriting existing files, appears at the beginning of the program’s
3.3. DATES 23
output. If there are errors in the spec file or the program terminates prematurely for other reasons, some or all
of the saved files may not be written.
3.3 Dates
Date arguments occur in several specs, and their values are always specified in the same format. Dates for
monthly data are written year.month; this format generalizes to other seasonal periods (e.g., year.quarter). It
is necessary to include all four digits when specifying a year. Thus, 67 means the year AD (or CE) 67, not AD
1967.
For monthly data the months can be denoted by either the integers 1–12 or by three-letter month abbrevi-
ations (jan, feb, mar, apr, may, jun, jul, aug, sep, oct, nov, and dec). Thus, 1967.12 and 1967.dec are
equivalent. For quarterly data or data with other seasonal periods, only integers are allowed, e.g., 1967.1 and
1967.4.
For data of any periodicity, a zero can be placed in front of integers from 1 to 9 for padding (for example,
2002.02 is an acceptable date specification for February 2002).
Dates are used to define the starting time point of a series, and when defining a subset (span) of a time series
to analyze. They are also used when defining outlier regression variables. For example, to specify regression
variables for an additive outlier in April of 1978 and a level shift beginning in September of 1982, we use the
following:
regression { variables=(ao1978.apr ls1982.sep) }
The seasonality of the dates used must match the seasonality specified for the data in the series spec, e.g.,
ao1972.jan is valid for monthly data but is not permitted for quarterly data.
3.4 General rules of input syntax
Allowable input characters

The allowable input characters, excepting characters that appear within quotes, are letters, numbers,
spaces, tabs, newline characters, and the following:
= . , { } ( ) [ ] + - #
The program will ignore any other ASCII characters in the spec file, but will flag them and generate a
warning message. The following additional characters are allowed within quotes:
! % ’ * / : ; < > ? @ \ _ / ~ ^
Also, double quotes are allowed within statements delimited by single quotes and vice-versa.
Braces, parentheses, and brackets

The { }, ( ), and [ ] characters serve different functions and cannot be used interchangeably. { } is used
to contain arguments in a spec, ( ) is used to contain a list of multiple values for an argument, and [ ] is
used (i) to contain values used in defining certain special arguments, such as the duration of an Easter
holiday regression variable, e.g., regression {variables = (td Easter[14])}, and (ii) to enclose the
lags present in an ARIMA model with missing lags, e.g., arima {model = (0 1 [1,3])}.
Case sensitivity
Spec names, arguments, dates, keywords (such as none and all), and predefined regression variable names
(such as td and seasonal) are not case sensitive. Thus, TD and td are the same; both are recognized by
the variables argument of the regression and x11regression specs.
Comments
Anything on a line after the # character, unless the # character is in quotes, is taken to be a comment.
If parts of a spec are commented out, what remains must still have balanced parentheses, brackets, and
braces.
Equals sign
The equals sign, = , is used when assigning values to arguments, e.g., print = none, or title =
"Monthly Retail Sales of Household Appliance Stores".
Line length in the spec file
Lines in the spec file are limited to 132 characters—any characters appearing beyond column 132 are
ignored. In particular, note that if a data set with lines exceeding 132 characters is placed in a spec
file this will result in data truncation on input. The 132 characters per line limitation does not apply,
however, to data read from a separate file (not the spec file) using the file argument. (The latter would
be governed by Fortran input line length restrictions, which may be system specific.)
Multiple argument values
Multiple argument values must be enclosed together in parentheses, e.g., variables=(td seasonal
const). If an argument accepts only a single value or it accepts multiple values but only one value is given,
then parentheses are optional. For example, the following are all valid; variables=td, variables=(td),
variables = (td seasonal), start=1967.4, and start=(1967.4).
Null list
A null list of arguments is allowed, e.g., outlier{ }. Any implied arguments in the null list then take
on their default values.
Numerical values
Numerical values can be specified in free format, including the use of exponential notation (e.g., 400,
400.0, 400., and 4.e+2 all denote the same real value). Integer notation must be used when an integer is
required (e.g., 2, not 2.0 or 2.e+0).
Ordering
The only restriction on the ordering of specs is that either series or composite must be the first spec.
Except for the b argument of the regression and x11regression specs, there are no restrictions on the
ordering of arguments within specs (see Sections 7.13 and 7.20 for more details). The ordering of multiple
values given to arguments matters for certain obvious cases, such as observations in data arguments
(series, transform, regression, and x11regression specs), the ARIMA model specification in the
3.4. GENERAL RULES OF INPUT SYNTAX 25
model argument (arima spec), and dates in span arguments (series and outlier specs).
Separators
Blank spaces, tabs, and blank lines may be used as separators as desired. Within a list of multiple argu-
ment values, single commas may also be used as separators, e.g., data=(0, 1, 2, 3, 4, 5). Commas
must be used to indicate missing argument values that are to be replaced by default values (for arguments
that require a specific number of values). For example, the span argument requires two values. In the
statement span=(1967.4, ), the presence of the comma after 1967.4 indicates that the second span
argument value is missing, so it takes on its default value (the date of the last observation).
Titles and filenames
A title, such as the name of a time series, must consist of at least one allowable input character (see
above), even if blank, and must be enclosed in either single or double quotes (’title’ or "title").
Lower and upper case of characters is preserved within titles. When the # character appears within
quotes, it is considered part of the title and does not denote the start of a comment. Titles must be
completed on one line and contain no more than 79 characters. Filenames, including the path, must
follow the same rules as titles.
4 RegARIMA modeling Capabilities
Contents
4.1 General model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Data input and transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Regression variable specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4 Identification and specification of the ARIMA part of the model . . . . . . . . . 35
4.5 Model estimation and inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.6 Diagnostic checking including outlier detection . . . . . . . . . . . . . . . . . . . . 37
4.7 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Tables
4.1 Predefined Regression Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Section 4.1 describes the general model handled by the X-13ARIMA-SEATS program. Sections 4.2 to 4.7 give
summary descriptions of the capabilities of X-13ARIMA-SEATS for the various stages of regARIMA modeling and
forecasting: data input and transformation, regression variable specification, ARIMA model identification and
specification, model estimation and inference, diagnostic checking including outlier detection, and forecasting.
These sections also mention which input specification statements (specs) are used to control the execution of
the capabilities discussed. Detailed documentation of the specs is given in Chapter 7.
When building a regARIMA model, it is strongly recommended that one examine a high resolution plot of
the time series. Such a plot gives valuable information about seasonal patterns, potential outliers, stochastic
nonstationarity, etc. Additional plots may also be useful for examining the effects of possible transformations on
the series, or of applying various differencing operators to the series. Since X-13ARIMA-SEATS does not possess
such plotting capabilities, other software must be used for this purpose.
4.1 General model
ARIMA models, as discussed by Box and Jenkins (1976), are frequently used for seasonal time series. A general
multiplicative seasonal ARIMA model for a time series zt can be written
φ(B)Φ(B s )(1 − B)d (1 − B s )D zt = θ(B)Θ(B s )at (4.1)
where B is the backshift operator (Bzt = zt−1 ), s is the seasonal period, φ(B) = (1 − φ1 B − · · · − φp B p )
is the nonseasonal autoregressive (AR) operator, Φ(B s ) = (1 − Φ1 B s − · · · − ΦP B P s ) is the seasonal AR
operator, θ(B) = (1 − θ1 B − · · · − θq B q ) is the nonseasonal moving average (MA) operator, Θ(B s ) =
(1 − Θ1 B s − · · · − ΘQ B Qs ) is the seasonal MA operator, and the at s are i.i.d. with mean zero and variance
26
4.1. GENERAL MODEL 27
σ 2 (white noise). The (1 − B)d (1 − B s )D implies nonseasonal differencing of order d and seasonal differencing
of order D. If d = D = 0 (no differencing), it is common to replace zt in (4.1) by deviations from its mean, that
is, by zt − µ where µ = E[zt ].
A useful extension of ARIMA models results from the use of a time-varying mean function modelled via
linear regression effects. More explicitly, suppose we write a linear regression equation for a time series yt as
X
yt = βi xit + zt (4.2)
i
where yt is the (dependent) time series,

P the xit are regression variables observed concurrently with yt , the βi are
regression parameters, and zt = yt − βi xit , the time series of regression errors, is assumed to follow the ARIMA
model in (4.1). modeling zt as ARIMA addresses the fundamental problem with applying standard regression
methodology to time series data, which is that standard regression assumes that the regression errors (zt in
(4.2)) are uncorrelated over time. In fact, for time series data, the errors in (4.2) will usually be autocorrelated,
and, moreover, will often require differencing. Assuming zt is uncorrelated in such cases will typically lead to
grossly invalid results.
The expressions (4.1)) and (4.2) taken together define the general regARIMA model allowed by the X-13-
ARIMA-SEATS program. Combining (4.1)) and (4.2), the model can be written in a single equation as
X
φ(B)Φ(B s )(1 − B)d (1 − B s )D yt − βi xit = θ(B)Θ(B s )at . (4.3)
i
The regARIMA model (4.3) Pcan be thought of either as generalizing the pure ARIMA model (4.1) to allow for a
regression mean function ( βi xit ), or as generalizing the regression model (4.2) to allow the errors zt to follow
the ARIMA model (4.1). In any case, notice that the regARIMA model implies that first the regression effects
are subtracted from yt to get the zero mean series zt , then the error series zt is differenced to get a stationary
series, say wt , and wt is then assumed to follow the stationary ARMA model, φ(B)Φ(B s )wt = θ(B)Θ(B s )at .
Another way to write the regARIMA model (4.3) is
X
(1 − B)d (1 − B s )D yt = βi (1 − B)d (1 − B s )D xit + wt . (4.4)
i
where wt follows the stationary ARMA model just given. Equation (4.4) emphasizes that the regression variables
xit in the regARIMA model, as well as the series yt , are differenced by the ARIMA model differencing operator
(1 − B)d (1 − B s )D .
Notice that the regARIMA model as written in (4.3) assumes that the regression variables xit affect the
dependent series yt only at concurrent time points, i.e., model (4.3) does not explicitly provide for lagged
regression effects such as βxi,t−1 . Lagged effects can be included by the X-13ARIMA-SEATS program, however,
by reading in appropriate user-defined lagged regression variables.
The X-13ARIMA-SEATS program provides additional flexibility in the specification of the ARIMA part of a
regARIMA model by permitting (i) more than two multiplicative ARIMA factors, (ii) missing lags within the
AR and MA polynomials, (iii) the fixing of individual AR and MA parameters at user-specified values when the
model is estimated, and (iv) inclusion of a trend constant, which is a nonzero overall mean for the differenced
series ((1 − B)d (1 − B s )D yt ). These features of regARIMA model specification are discussed and illustrated in
Section 4.6.
Detailed discussions of ARIMA modeling are given in the classic book by Box and Jenkins (1976), and also
in several other time series texts, such as Abraham and Ledolter (1983) and Vandaele (1983).
28 CHAPTER 4. REGARIMA MODELING CAPABILITIES
4.2 Data input and transformation

Observations of the original time series to be analyzed are read into the program with the series spec. The
data may either be included in the series spec or read from a file. The span and modelspan arguments of
the series spec are used to restrict the analysis to a span of the data, omitting data from the beginning and/or
end of the original time series. The series spec is also used to specify the starting date, seasonal period (if
appropriate), and title for the time series.
The transform spec provides nonlinear transformations of the data, as well as modification by prior-
adjustment factors. The nonlinear transformations included are the (Box and Cox 1964) family of power
transformations (such as the logarithm or square root), and the logistic transformation (useful for a time series
of proportions greater than 0 and less than 1). A predefined prior adjustment may be specified that divides
each observation in a monthly series by the corresponding length of month (or length of quarter for quarterly
series) and then re-scales it by the average length of month (or quarter). Similarly, leap year adjustment factors
for February are also available. Finally, a set of user-defined prior-adjustments may be supplied for division into
or subtraction from the original time series. The result of the series and transform specs is the time series
yt , t = 1, . . . , n, used in the regARIMA model 4.3.
4.3 Regression variable specification

Specification of a regARIMA model requires specification of both the regression variables (the xit ’s in (4.2)) and
the ARIMA model (4.1) for the regression errors zt . The former is done using the regression spec, and the latter
using the arima spec (discussed in Section 4.4). Choosing which regression variables to include requires user
knowledge relevant to the time series being modelled. Several regression variables that are frequently used in
modeling seasonal economic time series are built into the X-13ARIMA-SEATS program, and can be easily included
in the model. These are discussed below, and the actual regression variables used are given in Table 4.1 in this
section. Specification and use of these variables is described in the documentation of the regression spec in
Section 4.6. In addition, users may input data for any other regression variables (called user-defined regression
variables) that they wish to include in a model. As part of model estimation (see Section 4.5), X-13ARIMA-SEATS
provides standard t-statistics to assess the statistical significance of individual regression parameters, and χ2 -
statistics to assess the significance of groups of regression parameters corresponding to particular effects (such
as trading-day effects).
The most basic regression variable is the constant term. If the ARIMA model does not involve differencing,
this is the usual regression intercept, which, if there are no other regression variables in the model, represents
the mean of the (stationary) series. If the ARIMA model does involve differencing, X-13ARIMA-SEATS uses
a regression variable such that, when it is differenced according to the ARIMA model (see equation (4.4)), a
column of ones is produced. The corresponding parameter is then called a trend constant, since it provides for a
polynomial trend of the same degree as the number of differences in the model. For example, with nonseasonal
differencing (d > 0) but no seasonal differencing (D = 0), the (undifferenced) trend constant regression variable
is proportional to td . Notice that the lower order polynomial terms, tj for 0 ≤ j < d, are not included among
the regression variables because they would be differenced to zero by (1 − B)d , hence their coefficients cannot
be estimated. With or without the trend constant, the model (4.3) (or (4.4)) implicitly allows for these lower
order polynomial terms through the differencing. If seasonal differencing is requested (D > 0), the nature of the
undifferenced trend constant regression variable is more complicated, though the trend constant can be thought
4.3. REGRESSION VARIABLE SPECIFICATION 29
of as allowing for a polynomial of degree d + D. Without a trend constant, model (4.3) implicitly allows for a
polynomial of degree d + D − 1.
Fixed seasonal effects in a monthly series can be modelled using 12 indicator variables, one for each calendar
month. Since these 12 variables always add to one, however, they are confounded with an overall level effect.
This leads to one of two singularity problems: collinearity with the usual constant term in a model with no
differencing; or a singularity in a model with differencing since the 12 variables, when differenced, always
sum to 0. One appropriate reparameterization instead uses 11 contrasts in the 12 indicator variables. An
alternative reparameterization uses 11 variables taken from the Fourier (trigonometric) series representation
of a fixed monthly pattern. The variables used for both of these parameterizations are given in Table 4.1.
X-13ARIMA-SEATS allows either of these options, and also allows specifying the trigonometric terms only for
selected frequencies. For quarterly series, or for series with other seasonal periods, X-13ARIMA-SEATS constructs
the appropriate versions of these variables. Notice that these variables cannot be used in a model with seasonal
differencing, as they would all be differenced to zero.
Trading-day effects occur when a series is affected by the differing day-of-the-week compositions of the
same calendar month in different years. Trading-day effects can be modelled with 7 variables that represent
(no. of Mondays), . . . , (no. of Sundays) in month t. Bell and Hillmer (1983) note, however, that a better
parameterization of the same effects instead uses 6 contrast variables defined as (no. of Mondays) − (no. of
Sundays), . . . , (no. of Saturdays) − (no. of Sundays), along with a seventh variable for length of month (lom)
or its deseasonalized version, the leap-year regressor (lpyear). In X-13ARIMA-SEATS the 6 contrast variables
are called the tdnolpyear variables. Instead of using a seventh regressor, a simpler and often better way to
handle multiplicative leap-year effects is to re-scale the February values Yt of the original time series before
transformation to m̄F eb Yt /mt , where Yt is the original time series before transformation, mt is the length of
month t (28 or 29), and m̄F eb = 28.25 is the average length of February. (If the regARIMA model includes
seasonal effects, these can account for the length of month effect except in Februaries, so the trading day model
only has to deal with the leap year effect.) When this is done, only the tdnolpyear variables need be included
in the model. X-13ARIMA-SEATS allows explicit choice of either approach, as well as an option (td) that makes
a default choice of how to handle length-of-month effects—see the documentation of the regression spec.
The preceding paragraph assumes the time series being modelled represents the aggregation of some daily
series (typically unobserved) over calendar months. Such series are called monthly flow series. If the series
instead represents the value of some daily series at the end of the month, called a monthly stock series, then
different regression variables are appropriate. Trading-day effects in end-of-month stock series can be modelled
using 7 indicator variables for the day-of-the-week that the months end on. Since the sum of these variables
is always one, this leads to a singularity problem, so 6 contrast variables are used instead. (See Table 4.1.)
X-13ARIMA-SEATS also allows specification of regression variables appropriate for stock series defined as of some
other day of the month, e.g., for beginning of the month stock series.
For a general discussion of stock and flow series, access Wikipedia (2009).
For quarterly flow time series, X-13ARIMA-SEATS allows the same trading-day options as in the monthly
case. Trading-day effects in quarterly series are relatively rare, however, because the calendar composition of
quarters does not vary as much over time, on a percentage basis, as that of months does. Trading-day variables
are not provided for flow time series with seasonal periods other than monthly or quarterly, or for stock series
other than monthly.
Table 4.1: Predefined Regression Variables
Regression effect1 Variable definition(s)

2
Trend Constant (
const −d s −D 1 for t ≥ 1
(1 − B) (1 − B ) I(t ≥ 1), where I(t ≥ 1) =
0 for t < 1
3
Fixed Seasonal  
seasonal 1
 in January 1
 in November
M1,t = −1 in December , . . . , M11,t = −1 in December
 
0 otherwise 0 otherwise
 
2
Fixed Seasonal sin(ωj t), cos(ωj t), where ωj = 2πj/12, 1 ≤ j ≤ 6 (Drop sin(ω6 t) ≡ 0)
sincos[ ]
Trading Day T1,t = (no. of Mondays) − (no. of Sundays), . . . , T6,t = (no. of Saturdays) −
(monthly or quarterly flow) (no. of Sundays)
tdnolpyear, 4 td
One Coefficient (no. of weekdays) − 52 (no. of Saturdays and Sundays)

Trading Day
(monthly or quarterly flow)
td1nolpyear, 5 td1coef
Length-of-Month mt − m̄, where mt = length of month t (in days) and m̄ = 30.4375 (average length
(monthly flow) of month)
lom
Length-of-Quarter qt − q̄, where qt = length of quarter t (in days) and q̄ = 91.3125 (average length
(quarterly flow) of quarter)
loq
Leap Year 
(monthly and quarterly flow) 0.75
 in leap year February (first quarter)
lpyear LYt = −0.25 in other Februaries (first quarter)

0 otherwise

2 Restrictions, if any, are given in parentheses. Each entry also gives the name used to specify the regression effect in the
variables argument of the regression spec, e.g., regression { variables=const }.

3 The variables shown are for monthly series. Corresponding variables are available for any other seasonal period.
4 In addition to these 6 variables, the td option also includes the lpyear regression variable (for untransformed series), or it
re-scales February values of Yt to m̄F eb Yt /mt , where m̄F eb = 28.25 (average length of February) (for an original series Yt that is
transformed). Quarterly td is handled analogously.
5 In addition to this variable, the td1coef option also includes the lpyear regression variable (for untransformed series), or it
re-scales February values of Yt to m̄F eb Yt /mt , where m̄F eb = 28.25 (average length of February) (for an original series Yt that is
transformed). Quarterly td1coef is handled analogously.
Table 4.1: Predefined Regression Variables in X-13ARIMA-SEATS (continued)
Regression effect Variable definition(s)

Stock Trading Day 
(monthly stock) 1
 w̃th day of month t is a Monday
tdstock [w] D(w)1,t = −1 w̃th day of month t is a Sunday ,

0 otherwise


1
 w̃th day of month t is a Saturday
· · · , D(w)6,t = −1 w̃th day of month t is a Sunday ,

0 otherwise

where w̃ is the smaller of w and the length of month t. For end-of-month stock
series, set w to 31, i.e., specify tdstock[31].
6
One Coefficient I(w)t = − 35 D(w)1,t − 51 D(w)2,t + 15 D(w)3,t + 53 D(w)4,t +D(w)1,t where Di,t is the
Stock Trading Day ith stock trading day regressor defined as above for stock day w. For end-of-month
(monthly stock) stock series, set w to 31, i.e., specify tdstock1coef[31].
tdstock1coef [w]
7
E(w, t) = w1 ×[no. of the w days before Easter falling in month (or quarter) t].
Easter Holiday (Note: This variable is 0 except in February, March, and April (or first and second
(monthly or quarterly flow) quarter). It is nonzero in February only for w > 22.)
easter[w]
8
End-of-Month Let Ef (w, t) denote the flow easter[w] regressor with long-term means removed.
Stock Easter Then, in the monthly case:
(monthly or quarterly stock) 
easterstock[w] Ef (w, t)
 in February
Es (w, t) = Ef (w, t) + Ef (w, t − 1) in March .

0 otherwise

(Note: This variable is 0 except in February and March (or first quarter). It is
nonzero in February only for w > 22.)
6 For
details on the derivation of this regressor, see Findley and Monsell (2009).
7
The actual variable used for monthly Easter effects is E(w, t) − Ē(w, t), where the Ē(w, t) are the “long-run” monthly means
of E(w, t) corresponding to a 500 year period of the Gregorian calendar, 1600-2099. This provides a close approximation to the
average calculated over the much longer period of a complete cycle of the dates of Easter. For more details, see Bednarek (2007)
and Montes (2001). (These means are nonzero only for February, March, and April). Analogous deseasonalized variables are used
for Labor Day and Thanksgiving effects, and for quarterly Easter effects.
8 The variable shown is for monthly stock series. Corresponding variables are available for quarterly stock series, with the formula
documented in Section 4.2 of Findley (2009).


Statistics Canada Easter If Easter falls before April w, let nE be the number of the w days on or before
(monthly or quarterly flow) Easter falling in March. Then:
sceaster[w] 
nE /w
 in March
E(w, t) = −nE /w in April .

0 otherwise

If Easter falls on or after April w, then E(w, t) = 0.

(Note: This variable is 0 except in March and April (or first and second quarter).)
6
Labor Day L(w, t) = w1 ×[no. of the w days before Labor Day falling in month t]. (Note:
(monthly flow) This variable is 0 except in August and September.)
labor[w]
6
Thanksgiving ThC (w, t) = proportion of days from w days before Thanksgiving through Decem-
(monthly flow) ber 24 that fall in month t (negative values of w indicate days after Thanksgiving).
thank[w] (Note: This variable is 0 except in November and December.)
Additive Outlier at t0
aodate 0 (
(t ) 1 for t = t0
AOt 0 = (date 0 is the date corresponding to time point t0 )
0 6 t0
for t =
Level Shift at t0 (
lsdate 0 (t ) −1 for t < t0
LSt 0 =
0 for t ≥ t0
Temporary Change at t0 (
tcdate 0 (t ) 0 for t < t0
T Ct 0 = ,
αt−t0 for t ≥ t0
where α is the rate of decay back to the previous level (0 < α < 1).
Seasonal Outlier at t0 
sodate 0 0
 for t ≥ t0
(t0 )
SOt = 1 for t < t0 , t same month/quarter as t0

−1/(s − 1) otherwise


Ramp, t0 to t1 
rpdate 0 -date1 −1
 for t ≤ t0
(t0 ,t1 )
RPt = (t − t0 )/(t1 − t0 ) − 1 for t0 < t < t1

0 for t ≥ t1

Temporary Level Shift, t0 to t1 

tldate 0 -date1 0
 for t < t0
(t0 ,t1 )
T Lt = 1 for t0 ≤ t ≤ t1

0 for t > t1

X-13ARIMA-SEATS also provides a simplified model for trading day variation of monthly or quarterly flow
series that uses only one regressor, a weekday-weekend contrast variable:
5
Tt = (no. of W eekdays) − (no. of Saturdays and Sundays)
2
The underlying assumption for this model is that all weekdays (Monday through Friday) have identical
effects, and Saturday and Sunday have identical effects. In X-13ARIMA-SEATS this model can be estimated in
two ways: by specifying the td1coef option if the user wishes the program to make the choice of how to handle
length of month effects as with the td option mentioned above, or by specifying the td1nolpyear option in
which case the length of month effects model must be specified by the user, as with tdnolpyear.
The daily constraints from the flow series given above can be applied to the case of stock trading day as
well. The one-coefficient stock trading day variable for stock day w is given below:
3 1 1 3
I(w)t = − D(w)1,t − D(w)2,t + D(w)3,t + D(w)4,t + D(w)1,t
5 5 5 5
See Findley and Monsell (2009) for more details, with an application using industrial inventory series.
Holiday effects (in a monthly flow series) arise from holidays whose dates vary over time if (i) the activity
measured by the series regularly increases or decreases around the date of the holiday, and (ii) this differentially
affects two (or more) months depending on the date the holiday occurs each year. (Effects of holidays with
a fixed date, such as Christmas, are indistinguishable from fixed seasonal effects.) Easter effects are the most
frequently found holiday effects in U.S. economic time series, since the date of Easter Sunday varies between
March 22 and April 25. Labor Day and Thanksgiving also are potential, though less common, sources of holiday
effects. The basic model used by X-13ARIMA-SEATS for Easter and Labor Day effects assumes that the level of
activity changes on the w−th day before the holiday for a specified w, and remains at the new level until the
day before the holiday. For Thanksgiving the model used assumes that the level of activity changes on the day
that is a specified number of days before or after Thanksgiving and remains at the new level until December
24. The regression variable constructed for the holiday effect is, for a given month t, the proportion of the
affected time period that falls in month t. (Actually, as noted in Table 4.1, these regressors are deseasonalized
by subtracting off their long-run monthly means.) Essentially the same Easter effect variable applies also to
quarterly flow time series, but Labor Day and Thanksgiving effects are not present in quarterly series.
For stock series, the perspective of stocks as accumulations of monthly flows can be used to derive holiday
regressors for end of month stock series from cumulative sums of holiday regressors for flow series. A similar
approach has been used to obtain useful stock series trading day regressors by Cleveland and Grupe (1983),
Bell (1984, 1995), and Findley and Monsell (2009). While this could be done for any of the moving holidays in
X-13ARIMA-SEATS the only holiday currently with an end of month stock implementation is Easter; regressors
are available for both monthly and quarterly series.
Let E(w)fm,y
low
denote the deseasonalized Easter[w] regressor for a monthly flow series derived for month m
stock
and year y. The end of month stock Easter regressor Em,y is generated as follows:


 0, for m=1
E(w)f low ,

for m=2
2,y
E(w)stock
m,y = f low f low


 E(w)2,y + E(w)2,y , for m=3
0, for 4 ≤ m ≤ 12.

When 1 ≤ w ≤ 21, E(w)fm,y

low
is zero in February, so E(w)stock
m,y is nonzero only in March.
See Findley (2009) for more details and an application for manufacturing inventory series, and Titova and
Monsell (2009) and Chow and Moore (2009) for applications on US and UK inventory series.
X-13ARIMA-SEATS provides four other types of regression variables to deal with abrupt changes in the level
of a series of a temporary or permanent nature: additive outliers (AOs), level shifts (LSs), temporary changes
(TCs), seasonal outliers (SOs), ramps, and temporary level shifts (TLs). AOs affect only one observation in
the time series, LSs increase or decrease all observations from a certain time point onward by some constant
amount, TCs allow for an abrupt increase or decrease in the level of the series that returns to its previous level
exponentially rapidly, SOs allow for an abrupt increase or decrease in the level of the seasonal pattern that is
compensated for in the other months or quarters, ramps allow for a linear increase or decrease in the level of
the series over a specified time interval, and temporary level changes increase or decrease all observations for
a specific time span contained within the series by some constant amount. The specific regression variables
used to model these effects are given in Table 4.1. (LS regression variables are defined as −1 and then 0, in
preference to an equivalent 0 and then 1 definition, to make the overall level of the regression mean function
of any forecasts consistent with the most recent level of the time series. Similar considerations dictate the
definition of ramp variables and seasonal outliers.)
The regression spec allows specification of AOs, LSs, TCs, SOs, TLs, and ramps for cases where prior
knowledge suggests such effects at known time points. Often, however, large seasonal movements make it
difficult to identify where such changes in level have occurred. Determination of the location and nature of
potential outliers is the objective of the outlier detection methodology implemented by the outlier spec—see
Section 4.6 and the outlier spec documentation in Section 7.11. This methodology can be used to detect AOs,
TCs, and LSs (not ramps, seasonal outliers, or temporary level shifts); any that are detected are automatically
added to the model as regression variables.
4.4. IDENTIFICATION AND SPECIFICATION OF THE ARIMA PART OF THE MODEL 35
Prespecified AOs, LSs, TCs, SOs, TLs, and ramps are actually simple forms of interventions as discussed by
Box and Tiao (1975). While X-13ARIMA-SEATS does not provide the full range of dynamic intervention effects
discussed by Box and Tiao, often a short sequence of suitably chosen AO, LS, TC, TL and/or ramp variables
can produce reasonable approximations to more complex dynamic intervention effects, albeit at the cost of an
additional parameter or two. Analogous remarks apply to the relation between regARIMA models containing
(user-defined) regression variables that are themselves stochastic time series, and the dynamic transfer function
models discussed by Box and Jenkins (1976), chapters 10 and 11. Thus, regARIMA models can often be used to
approximate more general dynamic transfer function models, although transfer function models require special
treatment when forecasting, since future values of stochastic explanatory variables are generally unknown. (See
Box and Jenkins 1976, Section 11.5).
4.4 Identification and specification of the ARIMA part of the model

The ARIMA part of a regARIMA model is determined by its orders and structure, e.g., (p d q), (P D Q),
and s for model (4.1). If no regression variables are included in the model, then determination of the orders
for the resulting pure ARIMA model (called ARIMA model identification) can be carried out by following
well-established procedures that rely on examination of the sample autocorrelation function (ACF) and sample
partial autocorrelation function (PACF) of the time series yt and its differences. For regARIMA models, a
modified approach is needed, since the presence of regression effects can distort the appearance of the ACF
and PACF. Typically, the differencing orders can be identified by examining ACFs of the time series yt and its
differences. Then, one should obtain the residuals from a regression of the differenced data on the differenced
regression variables. The ACF and PACF of these residuals can then be examined in the usual way to identify
the AR and MA orders of the regression error term in the regARIMA model. This approach to regARIMA
model identification is discussed and illustrated in Bell and Hillmer (1983).
The key spec used to implement this approach to regARIMA model identification is the identify spec. For
illustration, consider a monthly seasonal time series. The usual ACFs examined to determine the differencing
needed are those of yt , (1 − B)yt , (1 − B 12 )yt , and (1 − B)(1 − B 12 )yt . The identify spec can produce all
these ACFs in a single run. Once the differencing has been determined, another run of X-13ARIMA-SEATS can be
made using the identify and regression specs together to (i) regress the differenced yt series on the differenced
regression variables, and (ii) produce the ACF and PACF of the regression residuals for use in identifying the
AR and MA orders of the model. For example, if one nonseasonal and one seasonal difference are specified
(d = 1 and D = 1), the identify and regression specs will fit the model
X
(1 − B)(1 − B 12 )yt = βi (1 − B)(1 − B 12 )xit + wt (4.5)
i
by ordinary least squares (OLS), and will produce the ACF and PACF of the regression residuals wt in (4.5).
An alternative approach that does not require two runs of the X-13ARIMA-SEATS program can be used if
the maximum differencing orders (nonseasonal and seasonal) that may be required are assumed known. For
example, suppose that these maximum differencing orders are d = 1 and D = 1. Then the identify and
regression specs can be used to (i) perform OLS regression on (4.5) toPproduce parameter estimates β̃i , (ii)
compute an estimated (undifferenced) regression error series z̃t = yt − i β̃i xit , and (iii) produce ACFs and
PACFs of z̃t , (1 − B)z̃t , (1 − B 12 )z̃t , and (1 − B)(1 − B 12 )z̃t . These ACFs and PACFs can be examined to
determine the orders of differencing required, as well as the orders of the AR and MA operators for the model.
There is one exception to the above remarks. If a constant term is specified in the regression spec, then
it will be included when the OLS regression is done on (4.5), but not when the regression effects are removed
from the data. Thus, actually, z̃t = yt − Σi≥2 β̃i xit if β̃1 x1t is the trend constant term. To explain why this
is done, we consider (4.5). From remarks in Section 4.3, a trend constant variable in model (4.5) allows for
a polynomial of degree 2, though the constant and linear terms (for t0 ≡ 1 and t) are implicitly allowed for
through the differencing by (1 − B)(1 − B 12 ). Since the constant and linear coefficients cannot be estimated, the
full polynomial effect cannot be subtracted from the undifferenced series yt . Rather than subtract only the t2
term of the polynomial, X-13ARIMA-SEATS ignores the estimated trend constant when creating the undifferenced
regression error series z̃t . Similar remarks apply to the general model (4.4). The only effect that inclusion of a
trend constant has on the computations of the identify spec is that its inclusion in (4.4) will affect the regression
estimates β̃i for i ≥ 2.
4.5 Model estimation and inference
The regression and arima specs specify a regARIMA model. The estimate spec then estimates the model
parameters by exact maximum likelihood, or by a variant known as conditional maximum likelihood (Box and
Jenkins 1976, pp. 209–212), which is sometimes called conditional least squares. Users may specify maximization
of the fully exact likelihood, or of the likelihood conditional for the AR but exact for the MA parameters, or
of the likelihood conditional for both the AR and MA parameters. Differences in AR parameter estimation
between exact and conditional likelihood maximization are generally small, and there are situations where
each approach is appropriate. (See Chapter 5.) Differences between exact and conditional likelihood for MA
parameter estimation are more fundamental, with exact likelihood being the recommended approach. The
option of choosing the conditional likelihood for MA parameters is provided in X-13ARIMA-SEATS mainly for
comparison of results with other software, and for occasional use to produce initial estimates for exact maximum
likelihood estimation when convergence problems arise. (See Section 5.1.) The default option is exact maximum
likelihood estimation for both the AR and MA parameters.
Whichever choice of estimation method is made, the resulting log-likelihood for a pure ARIMA model is
reduced to a sum of squares function that is then minimized by a nonlinear least squares routine (MINPACK,
discussed by More, Garbow, and Hillstrom (1980). To maximize the likelihood for a full regARIMA model, an
iterative generalized least squares (IGLS) algorithm (Otto, Bell, and Burman 1987) is used. This algorithm
involves two general steps: (i) for given values of the AR and MA parameters, the regression parameters that
maximize the likelihood are obtained by a generalized least squares (GLS) regression (using the covariance
structure of the regression errors, which is determined by their ARIMA model); and (ii) for given values, βi ,
of the regressionPparameters, the ARIMA model is fit by maximum likelihood to the time series of regression
errors, zt = yt − βi xit . IGLS iterates between these two general steps until convergence is achieved. (Output
options in the estimate spec allow for display of intermediate results from the estimation iterations, if desired.)
The likelihood function (exact or conditional) is evaluated using an approach derived from those suggested by
Box and Jenkins (1976), Chapter 7, Ljung and Box (1979), Hillmer and Tiao (1979), and Wilson (1983). Section
4.4 discusses certain problems that may arise in model estimation that all users should be aware of.
Statistical inferences about regARIMA model parameters may be made using asymptotic results for max-
imum likelihood estimation of ARIMA models (Box and Jenkins 1976, chapter 7; Brockwell and Davis 1991,
chapter 8 and regARIMA models (Pierce 1971)). These results state that, under suitable assumptions, the
parameter estimates are approximately normally distributed with means equal to the true parameter values
4.6. DIAGNOSTIC CHECKING INCLUDING OUTLIER DETECTION 37
and with a certain covariance matrix that can be estimated. (The “suitable assumptions” include that the true
model form is used, that the model’s AR operators are all stationary and its MA operators are all invertible, and
that the series is sufficiently long for the asymptotic results to apply.) Using these results, X-13ARIMA-SEATS
provides standard errors for the ARMA and regression parameter estimates, and, optionally, correlation (or
covariance) matrices for the estimates of both the ARMA and the regression parameters. (The regression pa-
rameter estimates are asymptotically uncorrelated with the ARMA parameter estimates.) These results may
be used in the usual way to make normal theory inferences about model parameters, including, as mentioned
in Section 4.3, use of t-statistics and χ2 -statistics produced by X-13ARIMA-SEATS to assess the statistical signif-
icance of individual regression parameters and of groups of regression parameters corresponding to particular
regression effects. Also, since X-13ARIMA-SEATS prints out the value of the maximized log-likelihood function,
various likelihood ratio tests are possible by making multiple runs of the program with different models.
X-13ARIMA-SEATS uses the maximum likelihood estimate of the residual variance σ 2 , which is σ̂ 2 = SS/(n −
d − s · D), where SS is the residual sum-of-squares and n − d − s · D is the effective number of observations after
differencing. (If the likelihood function that is conditional with respect to the AR parameters is used, replace
n − d − s · D by n − p − d − s · P − s · D.) Notice there is no “degrees of freedom” adjustment for model parameters
being estimated. For this reason, if X-13ARIMA-SEATS is used to fit a pure regression model—a model whose
regression errors follow the ARIMA(0 0 0) model—σ̂ 2 will differ from the usual unbiased regression variance
estimate. Consequently, the resulting standard errors, t-statistics, and χ2 -statistics for the regression parameter
estimates will also differ slightly from those that would be obtained from a standard regression program.
An alternative approach to inference is to use the likelihood-based model selection criteria produced by
X-13ARIMA-SEATS: AIC, AICC (also known as the F-adjusted AIC), Hannan-Quinn, and BIC. For each of these
statistics, the model producing the lower value is preferred. One advantage to these criteria over standard t-sta-
tistics, χ2 -statistics, and likelihood ratio tests is that they may be used to compare non-nested models—models
that differ from each other in such a way that one model cannot be obtained simply by removing parameters
from another model. (E.g., AR(1) versus MA(1) is a non-nested comparison.) Some caution must be exercised
in use of the model selection criteria. Section 5.5 discusses certain situations that arise in regARIMA modeling
for which the use of these criteria, as well as standard likelihood ratio tests, is invalid.
4.6 Diagnostic checking including outlier detection

Diagnostic checking of a regARIMA model is performed through various analysis of the residuals from model
estimation, the objective being to check if the true residuals (at in (4.3)) appear to be white noise—i.i.d. N (0, σ 2 ).
(Note: Normality of the at s is not needed for the large sample estimation and inference results; it is most
important for validity of prediction intervals produced in forecasting.) The check spec is used to produce
various diagnostic statistics using the residuals from the fitted model. To check for autocorrelation, X-13-
ARIMA-SEATS can produce ACFs and PACFs of the residuals (with standard errors), along with Ljung and Box
(1978) summary Q-statistics. X-13ARIMA-SEATS can also produce basic descriptive statistics of the residuals
and a histogram of the standardized residuals. The residuals can be written to a file for further analysis (such
as high resolution plotting) by other software.
The residuals produced by the algorithm used for exact maximum likelihood estimation in X-13ARIMA-SEATS
are the conditional expectations of the innovations at given the data {yt ; t = 1, . . . , n}, that is, E[at |{yt }]. (See
Box and Jenkins 1976, Appendix A7.4.) They are computed using the estimated values of the model parameters.
For a nonseasonal ARIMA model of order (p, d, q), the residuals are computed for time points t = d+1−q, . . . , n,
which gives n − d + q residuals. For a seasonal ARIMA model with seasonal period s, the first residual would
be at t = d + Ds − q − Qs + 1. For the airline model d = q = 1 and D = Q = 1, so there will be n residuals
going from t = 1, . . . , n. Other algorithms that can be used to evaluate or approximate the likelihood of an
ARIMA model – e.g., the conditional likelihood option in X-13ARIMA-SEATS , or the Kalman filter (used in
other software) – can produce different residuals, even different numbers of residuals (even when the approaches
yield the same likelihood value). If the estimated model is invertible, residuals computed in these different ways
will eventually converge to each other as t increases, and so will typically be very close in the latter part of the
time series, and may be very similar for most of the time series.
An important aspect of diagnostic checking of time series models is outlier detection. The outlier spec of
X-13ARIMA-SEATS provides for automatic detection of additive outliers (AOs), temporary change outliers (TCs)
and level shifts (LSs). These outlier types (referring to AOs, TCs, and LSs as “outliers”) and their associated
regression variables are defined in Section 4.3. X-13ARIMA-SEATS’ approach to outlier detection is based on
that of Chang and Tiao (1983)—see also Chang, Tiao, and Chen (1988)—with extensions and modifications as
discussed in Bell (1983) and Otto and Bell (1990). The general approach is similar to stepwise (GLS) regression,
where the candidate regression variables are AO, LS, and/or TC variables for all time points at which outlier
detection is being performed—3n variables for detection of AOs, LSs, and TCs over an entire time series of
length n. (Actually, slightly fewer than 3n variables are used in this case for reasons discussed in the DETAILS
section of the outlier spec documentation in Section 7.11.) In brief, this approach involves computing t-statis-
tics for the significance of each outlier type at each time point, searching through these t-statistics for significant
outlier(s), and adding the corresponding AO, LS, or TC regression variable(s) to the model. Overly burdensome
computation is avoided by holding the AR and MA parameters fixed as the outlier t-statistics are computed
for each time point and outlier type. X-13ARIMA-SEATS provides two variations on this general theme. The
addone method provides full model re-estimation after each single outlier is added to the model, while the
addall method re-estimates the model only after a set of detected outliers is added. A description of both these
methods is given in the documentation of the outlier spec in Section 7.11, with more details in Appendix B of
Findley, Monsell, Bell, Otto, and Chen (1998).
During outlier detection a robust estimate of the residual standard deviation, 1.48 × the median absolute
deviation of the residuals (Hampel, Ronchetti, Rousseeuw, and Stahel 1986, p. 105), is used. Because outlier
detection involves searching over all (or a specified set of) time points for the most significant outliers, the usual
normal distribution critical values (e.g., 2.0) are too low for judging significance in outlier detection. The default
critical value is determined by the number of observations in the interval searched for outliers (see Table 7.22),
but this can be changed by the user.
When a model contains two or more level shifts, including those obtained from outlier detection as well as
any level shifts specified in the regression spec, X-13ARIMA-SEATS will optionally produce t-statistics for testing
null hypotheses that each run of two, three, etc. successive level shifts actually cancels to form a temporary
level shift. Two successive level shifts cancel to form a temporary level shift if the effect of one offsets the effect
of the other, which implies that the sum of the two corresponding regression parameters is zero. Similarly,
three successive level shifts cancel to a temporary level shift if the sum of their three regression parameters
is zero, etc. (There is a user-specified limit on the number of successive level shifts in the runs tested.) The
t-statistics produced are the sums of the estimated parameters for each run of successive level shifts divided
by the appropriate standard error. An insignificant temporary level shift t-statistic (say, one less than 2 in
magnitude) fails to reject the null hypothesis that the corresponding level shifts cancel to form a temporary
level shift. These tests are provided primarily as diagnostics to help users assess the impacts of level shifts in a
model. Of course, if one or more of these t-statistics are significant, the user may wish to re-specify the model
4.7. FORECASTING 39
with the relevant level shift regression variables replaced by appropriate temporary level shift variables. (These
can be specified as user-defined regression variables, or can be obtained by fixing the coefficient of the level shift
regressors so that they sum to zero.) The choice between using level shifts (which correspond to permanent
changes in the level of a series) versus temporary level shifts could be important for forecasting a series with
level shifts near the end of the data.
4.7 Forecasting
For a given regARIMA model with parameters estimated by the X-13ARIMA-SEATS program, the forecast spec
will use the model to compute point forecasts, and associated forecast standard errors and prediction intervals.
The point forecasts are minimum mean squared error (MMSE) linear predictions of future yt s based on the
present and past yt s assuming that the true model is used—which means we assume the regARIMA model
form is correct, that the correct regression variables have been included, that no additive outliers or level shifts
will occur in the forecast period, that the specified ARIMA orders are correct, and that the parameter values
used (typically estimated parameters) are equal to the true values. These are standard assumptions, though
obviously unrealistic in practical applications. What is more realistically hoped is that the regARIMA model
will be a close enough approximation to the true, unknown model for the results to be approximately valid.
Two sets of forecast standard errors are produced. One assumes that all parameters are known. The other
allows for additional forecast error that comes from estimating the regression parameters, while still assuming
that the AR and MA parameters are known. For a reasonably long time series, (Box and Jenkins 1976, pp.
267–269) observe that the contribution to forecast error of the error in estimating the AR and MA parameters
is generally small, thus providing a justification for ignoring this source of error when computing the forecast
standard errors.
If the series has been transformed, then forecasting results are first obtained in the transformed scale, and
then transformed back to the original scale. For example, if one specifies a model of form (4.3) for yt = log(Yt ),
where Yt is the original time series, then yt is forecasted first, and the resulting point forecasts and prediction
interval limits are exponentiated to produce point and interval forecasts in the original (Yt ) scale. The resulting
point forecasts are MMSE for yt = log(Yt ), but not for Yt under the “standard” assumptions mentioned above.
Analogous procedures are followed for other allowable transformations. If any prior adjustments are made, these
will also be inverted in the process of transforming the point forecasts and prediction interval limits back to the
original scale.
If there are any user-defined regression variables in the model, X-13ARIMA-SEATS requires that the user supply
data for these variables for the forecast period. For the X-13ARIMA-SEATS predefined regression variables, the
program will generate the future values required. If user-defined prior adjustment factors are specified, values
for these should also be supplied for the forecast period.
User-defined regressors are usually deterministic (i.e. nonstochastic) variables, such as special calendar-
effect or outlier variables. But when forecasts are of particular interest, it can be appropriate to use stochastic
variables, e.g. economic time series values, as regressors, see Findley (2007).
5 Points Related to regARIMA Model
Estimation
Contents
5.1 Initial values for parameters and dealing with convergence problems . . . . . . . 40
5.2 Invertibility (of MA operators) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.3 Stationarity (of AR operators) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.4 Cancellation (of AR and MA factors) and overdifferencing . . . . . . . . . . . . . 42
5.5 Use of model selection criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.5.1 Avoid using the criteria to compare models with different sets of outlier regressors when
possible . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.5.2 Model comparisons for transformed data . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.5.3 Do not use the criteria to compare models with different differencing operators . . . . 48
Tables
5.1 Probability that a Chi-Square Variate with ν Degrees of Freedom Exceeds 2ν +
∆AIC for ∆AIC = 0, 1, 2, 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
While the IGLS algorithm and nonlinear least squares routine used by the X-13ARIMA-SEATS program
are quite reliable at finding maximum likelihood estimates for regARIMA models, problems in estimation
occasionally do occur. Some problems that can arise in model estimation are discussed below, along with
possible solutions. This is followed by important cautions regarding the use of the model selection criteria
produced by the X-13ARIMA-SEATS program.
5.1 Initial values for parameters and dealing with convergence prob-
lems
Users may supply initial values for AR and MA parameters that are then used to start the iterative likelihood
maximization. This is rarely necessary, however, and is not generally recommended. The default choice of initial
parameter values in X-13ARIMA-SEATS is 0.1 for all AR and MA parameters. (Initial values are not needed for
the regression parameters, which are determined in the GLS regressions.) This default choice of initial values
appears to be adequate in the great majority of cases. Supplying better initial values (as might be obtained,
e.g., by first fitting the model using conditional likelihood) does not seem to speed up convergence enough to
make obtaining the initial estimates generally worth the effort. A possible exception to this occurs if initial
estimates that are likely to be extremely accurate are already available, such as when one is re-estimating a
40
5.2. INVERTIBILITY (OF MA OPERATORS) 41
model with a small amount of new data added to a time series. However, the main reason for specifying initial
parameter values is to deal with convergence problems that may arise in difficult estimation situations.
When X-13ARIMA-SEATS’ iterative estimation scheme fails to converge, several remedies are available. If
the program stopped short of convergence because it reached the maximum number of iterations (indicated by
a warning message to this effect and the printing of parameter values at the last iteration), then rerunning the
program with initial parameter values set at the values obtained at the last iteration may produce convergence.
An easier, though computationally slower, alternative is to simply increase the number of iterations allowed and
rerun the program. If the program crashed before converging or reaching the maximum number of iterations,
then it may help to first fit the model by conditional likelihood, and then use the resulting parameter estimates
as initial values for exact maximum likelihood estimation. On the other hand, it has been our experience
that convergence problems are often due to the use of a model that is complicated (e.g., high order), or poorly
conditioned. In such cases, the appropriate action is to examine the results and specify a simpler model. Sections
5.2 through 5.4 discuss some particular situations that can lead to estimation problems and that suggest specific
model modifications.
5.2 Invertibility (of MA operators)

An MA polynomial, θ(B) = 1−θ1 B − · · · −θq B q , is invertible if all the roots, G1 , . . . , Gq , of θ(B) = 0 lie outside
the unit circle (|Gj | > 1 for all j). As shown in Brockwell and Davis (1991), pp. 123–125, for any invertible MA
operator in an ARIMA model there are one or more corresponding noninvertible MA operators that produce
the same autocovariance structure, and hence the same unconditional likelihood function. Although the data
thus cannot discriminate between the invertible and corresponding noninvertible models, the preferred choice
is the invertible model. This is essential for forecasting—grossly incorrect forecasting results can be obtained
with noninvertible models. There is one important exception. MA polynomials with roots on the unit circle
(|Gj | = 1), the boundary of the invertibility region, do not cause problems for forecasting when handled
appropriately (by exact maximum likelihood for MA models).
Estimation in X-13ARIMA-SEATS enforces invertibility constraints on the MA parameters in the iterative non-
linear maximization of the likelihood function. Strictly speaking, then, models estimated by X-13ARIMA-SEATS
are invertible. If the maximum likelihood estimates (MLEs) for a given model are actually on the boundary
of the invertibility region, i.e., the model at the MLEs contains an MA operator with zeroes exactly on the
unit circle, then X-13ARIMA-SEATS’ nonlinear search will approach the boundary of the invertibility region from
within, and will generally get as close to the boundary as the convergence tolerance dictates or the maximum
number of iterations allows. X-13ARIMA-SEATS can thus effectively produce estimated models on the boundary
of the invertibility region. Convergence of the estimation iterations in such cases can be slow, since finding the
maximum of the likelihood function on the boundary of the constrained parameter space is a difficult optimiza-
tion problem. More importantly, convergence of the estimation to the invertibility boundary often indicates
that the model is poorly conditioned, and should alert users to examine the results (and possibly detailed
output of the estimation iterations) for signs of this. Section 5.4 discusses the most important causes of poor
conditioning—cancellation of factors and overdifferencing—and the appropriate remedies.
Estimation seems most likely to produce a noninvertible model when the model contains a seasonal difference
and a seasonal MA polynomial, e.g., 1 − ΘB s when the MLE of Θ is 1. As such models are commonly used for
seasonal economic time series, users should be alert to this possibility and be aware of the appropriate action
to take as discussed in Section 5.4.
42 CHAPTER 5. POINTS RELATED TO REGARIMA MODEL ESTIMATION
5.3 Stationarity (of AR operators)
An AR polynomial, φ(B) = 1 − φ1 B − · · · − φp B p , is stationary if all roots of φ(B) = 0 lie outside the

unit circle; otherwise, it is nonstationary. (More accurately, the series wt = (1 − B)d (1 − B s )D zt following the
model φ(B)Φ(B s )wt = θ(B)Θ(B s )at (derived from equation (1)) is stationary if and only if the zeroes of all the
AR polynomials lie outside the unit circle.) The exact (for AR) likelihood function assumes all AR operators
are stationary. Hence, the exact (for AR) likelihood can be evaluated, and estimation and other analysis (e.g.,
forecasting) performed, only if the AR parameters satisfy stationarity constraints. Thus, when the exact (for
AR) likelihood function is used, X-13ARIMA-SEATS enforces stationarity constraints on the estimation. Unless
cancellation of factors is present (see the next section), it is unlikely for X-13ARIMA-SEATS’ nonlinear estimation
to approach the boundary of the stationary region, since the log-likelihood approaches −∞ as this boundary is
approached.
If the likelihood is defined conditionally with respect to the AR parameters, stationarity is neither assumed
nor enforced by the X-13ARIMA-SEATS software. Model estimation, forecasting, etc., are not compromised by
parameter values outside the stationary region in this case. Inference results, however, are affected, as noted
in Section 4.5. Special techniques (as in, e.g., Fuller 1976, Section 8.5) are required for inference about AR
parameters outside the stationary region.
5.4 Cancellation (of AR and MA factors) and overdifferencing
Cancellation of AR and MA factors is possible when a model with a mixed ARMA structure is estimated. A
model as in (1) or (3) is said to have a mixed ARMA structure if either p > 0 and q > 0, or P > 0 and
Q > 0. (Technically, a model with p > 0 and Q > 0, or with P > 0 and q > 0, is also mixed, but such mixed
models are unlikely to lead to cancellation problems.) The simplest example of cancellation occurs with the
ARMA(1,1) model, (1 − φB)zt = (1 − θB)at , when φ = θ. Cancelling the (1 − φB) factor on both sides of the
model (1 − φB)zt = (1 − φB)at leaves the simplified model, zt = at . Because of this, the likelihood function
will be nearly constant along the line φ = θ. This can lead to difficulties with convergence of the nonlinear
estimation if the MLEs for the ARMA(1,1) model approximately satisfy φ̂ = θ̂. Analogous problems occur in
more complicated mixed models when an AR polynomial and an MA polynomial have a common zero (e.g., the
ARIMA (2,1,2)(0,1,1) model that is used as a candidate model for the automdl spec). For a fuller discussion
of this topic, see Box and Jenkins (1976), pp. 248-250.
If the X-13ARIMA-SEATS program has difficulty in converging when estimating a mixed model, cancellation
of AR and MA factors may be responsible. In any case, possible cancellation can be checked by computing
zeroes of the AR and MA polynomials (setting print=roots in the estimate spec), and examining these for
zeroes common to an AR and an MA polynomial. If a common zero (or zeroes) is found, then the model
should be simplified by cancelling the common factor(s) (reducing the order of the corresponding AR and MA
polynomials), and the model should be re-estimated. Cancellation need not be exact, but may be indicated by
zeroes of an AR and an MA polynomial that are approximately the same.
It is also possible for estimated MA polynomials to have factors that cancel with differencing operators.
This occurs when a model has a nonseasonal difference and an estimated nonseasonal MA polynomial contains
a (1 − B) factor, or the model has a seasonal difference and an estimated seasonal MA polynomial contains a
(1 − B s ) factor. For example, the model (1 − B)(1 − B s )zt = (1 − θB)(1 − ΘB s )at involves such cancellation
5.5. USE OF MODEL SELECTION CRITERIA 43
if either θ or Θ is estimated to be one. Such cancellation is called “overdifferencing”, since it implies that
the series was differenced more times than necessary to achieve stationarity. When overdifferencing occurs the
corresponding difference and MA factor may be cancelled to simplify the model, but the user must then also add
to the model regression term(s) to account for the deterministic function of time that was previously annihilated
by the cancelled differencing operator. This means that if a nonseasonal difference is cancelled with a (1 − θB)
MA factor with θ̂ = 1, then the simplified model should include a trend constant (or overall mean, if the model
had only this one difference). If a seasonal difference is cancelled with a (1 − ΘB s ) seasonal MA factor with
Θ̂ = 1, then the simplified model should include both a trend constant (or overall mean) and fixed seasonal
effects. Overdifferencing is discussed by Abraham and Box (1978) and Bell (1987).
If estimation converges to an overdifferenced model, modifying the model by removing the differencing
operator and MA factor that cancel as well as including the appropriate regression terms, and then re-estimating
the model, is somewhat optional, because this cancellation does not necessarily lead to problems with model
estimation and other results (assuming use of the likelihood function that is exact for the MA parameters).
In particular, forecasting results should be the same for both the overdifferenced model and the corresponding
modified model, and regression and ARMA parameter estimates and standard errors under the two models
should be approximately the same. (However, log-likelihood values and the corresponding model selection
criteria will be different for the two models—see the next section.) This contrasts with the situation regarding
cancellation of AR and MA factors. Since cancellation of AR and MA factors is more likely to lead to convergence
problems in estimation, common AR and MA factors should always be removed from the model, and the model
re-estimated.
5.5 Use of model selection criteria
The X-13ARIMA-SEATS program provides the following model selection criteria: AIC (Akaike 1973, see also
Findley 1985, 1999, and Findley and Wei 2002), AICC (Hurvich and Tsai 1989), a criterion due to Hannan and
Quinn (1979), and BIC (Schwarz 1978). Suppose the number of estimated parameters in the model, including
the white noise variance, is np . If after applying the model’s differencing and seasonal differencing operations,
there are N data, and if the estimated maximum value of the exact log likelihood function of the model for the
untransformed data is denoted LN , then the formulas for these criteria are:
AICN = −2LN + 2np

−1
n +1
AICCN = −2LN + 2np 1 − pN
HannanQuinnN = −2LN + 2np log log N
BICN = −2LN + np log N.
If a function f of the transform spec is applied before regARIMA model estimation, then the maximized log
likelihood LN of the untransformed data Yt in the formulas above is obtained as follows. Given the regARIMA
model’s differencing operator as (1 − B)d (1 − B s )D , let the transformed data yt = f (Yt ) used for modeling be
y−(d+sD)+1 , . . . , y0 , y1 , . . . , yN . The number N is called the effective number of observations. Let LyN denote
the regARIMA model’s maximized log likelihood for y1 , . . . , yN conditional on y−(d+sD)+1 , . . . , y0 , which is
D
calculated as the maximized log likelihood of the regARIMA model for (1 − B)d (1 − B s ) yt , 1 ≤ t ≤ N . Then
N
X df (Yt )
LN = LyN + log
.
t=1
dYt
The second term on the right is called the transformation adjustment. (It is the Jacobian of the data
transformation Yt = f −1 (yt ) , 1 ≤ t ≤ N ; see Chapter 6 of Mood, Graybill, and Boes, 1974.) In the case of
PN
the (natural) log transformation f (Yt ) = log Yt , for example, it is − t=1 log Yt . Defining the model selection
criteria in terms of the untransformed data Yt makes it possible to compare competing transformations for this
data, for example the log transformation and no transformation, see Section 7.18.
Akaike’s Minimum AIC criterion (MAIC) states that, between any two models, the one with the smaller
AIC is preferred; see Akaike (1973) and Findley (1999) for example. Similarly, for each of the other model
selection criteria above, the model with the smaller value is preferred. This property is determined by the sign
of the difference of the criterion values. Focussing on AIC, given two models, designated model 1 and model 2,
(1) (2) (1)
with log maximum likelihood values and numbers of estimated parameters denoted by LN and LN and np
(2)
and np , respectively, we consider the AIC difference
(1) (2) (1) (2)

AICN − AICN = −2{LN − LN } − 2(n(2) (1)
p − np ). (5.1)
When model 1 is of the correct type and is a special case of (is ”nested in”) model 2, then for long enough
(2) (1)
time series, the AIC difference (5.1) varies approximately like a chi-square variate with np − np degrees of
freedom, i.e. asymptotically
(1) (2)
−2{ln LN − ln LN } ∼ χ2n(2) −n(1) , (5.2)
p p
holds, under standard assumptions, including the requirement that the true model is invertible, i.e. without
unit magnitude roots in the MA
polynomial
(see Taniguchi
and Kakizawa 2000, p. 61). The same result applies
(1) (2)
to AICC differences because np + 1 /N and np + 1 /N tend to zero as N increases.
Under (5.2), the asymptotic probability that model 2 will have a smaller AIC and thus incorrectly be
preferred over model 1 by the MAIC criterion is, from (5.2),
(1) (2)
P (AICN − AICN > 0) = P (χ2n(2) −n(1) > 2(n(2) (1)
p − np )). (5.3)
p p
Thus, the right hand side of (5.2) gives the asymptotic probability of a Type I error by the Minimum AIC
criterion, i.e. of rejecting model 1 in favor of model 2. Some relevant values when the aictest variable of the
regression spec (see Section 7.13) is used with trading day and holiday effect regression models of Table 4.1
are given in Table 5.1.
Table 5.1 shows the effect on the asymptotic Type I error probability of using certain values ∆AIC of the
aicdiff argument in conjunction with the aictest argument to bias the decision toward the model without the
regression effect tested. (The default is aicdiff = 0.) The degrees of freedom values ν for which probabilities
P (χ2ν > 2ν + ∆AIC ) are given apply to certain trading day models defined in Table 4.1, e.g., tdnolpyear
and td1nolpyear, with and without lpyear. However, with td and td1coef in the multiplicative adjustment
Table 5.1: Probability that a Chi-Square Variate with ν Degrees of Freedom Exceeds 2ν + ∆AIC for
∆AIC = 0, 1, 2, 3.
ν P (χ2ν > 2ν) P (χ2ν > 2ν + 1) P (χ2ν > 2ν + 2) P (χ2ν > 2ν + 3)

1 .157 .083 .046 .025
2 .135 .082 .050 .030
5 .075 .051 .035 .025
6 .062 .043 .030 .020
7 .051 .036 .025 .017
∞ 0 0 0 0
case, when fixed leap year ratio preadjustment factors are used with tdnolpyear and td1nolpyear regressors,
instead of estimating a coefficient of LYt , the model with no trading day effects is not nested in the td and
td1coef models, see the Details of Section 7.13. In these two cases, the use of aictest can be shown to have
an asymptotic probability of a Type I error equal to zero, because the incorrect use of the fixed leap year ratio
preadjustment factors cause the models with them to be asymptotically worse than the model with no trading
day effects, so the discussion below leading to (5.4) applies.
When aictest = easter is used, the Type I error probabilities are slightly higher than those given in
Table 5.1 because, instead of single model, three different models, with easter[1], easter[8] and easter[15]
regressors, respectively, are being compared to a model with no Easter regressor.
Type I error probabilities may provide some helpful insights into properties of MAIC, but it must be kept
in mind that they arise from a different modeling paradigm. The minimum AIC criterion is based on a deep
approximation property rather than on conventional significance tests: under assumptions that encompass those
used to calculate Type I error probabilities, an AIC difference is an asymptotically unbiased estimate of the
difference between the Kullback-Leibler quasi-distances from the true model to the estimated models; see Akaike
(1973), Findley (1999) and Findley and Wei (2002) for example. The Minimum AIC criterion seeks to indicate
which model is closer to the truth in this sense. This property can justify the use of MAIC for some nonnested
model comparisons where likelihood ratio tests based a chi-square distribution don’t exist.
Also, regardless of whether the models are nested or nonnested, if model 2 is asymptotically worse than
model 1 (specifically, farther from the true model in the Kullback-Leibler sense), then it can be shown that
1 n (1) (2)
o 2 (2) (1)
lim AICN − AICN = lim {LN − LN } = C1,2 < 0, (5.4)
N →∞ N N →∞ N
(1) (2)
(in probability) with the result that AICN − AICN tends to −∞ effectively linearly in N . Hence MAIC will
(1) (2)
strongly prefer model 1 for large enough N . The same result holds for AICCN − AICCN (and also for the
other criteria above). This property further helps to explain why AIC and AICC have often been found to be
(1) (2)
effective with nonnested model comparisons. For such comparisons, AICCN − AICCN is often rather large,

e.g. greater than three, with series of average lengths unless the models being compared are quite close for the
modeled series (as can happen with the easter[w] regressors of Table 4.1).
In situations in which multiple models are compared (more than two or three, perhaps substantially more),
it is worthwhile to consider the model with the second smallest AICC value as well as the minimum AICC
model, and perhaps other models whose AICC value is close to the minimum value, especially when the model
comparisons are non-nested. These alternative model sometimes have more desirable features, e.g. several
fewer parameters, better interpretability, greater consistency with the model chosen for several closely related
series, etc. Burnham and Anderson (2004) on page 271 offers rough rules of thumb for this situation, that
we formulate with AICC instead of AIC. With AICC min denoting the minimum AICC value and AICC alt
denoting the second smallest AICC value or the AICC value of some similarly competitive alternative model,
set ∆ = AICC alt − AICC min . If ∆ ≤ 2, there is substantial support for the alternative model, considerably
less if 4 ≤ ∆ ≤ 7, and essentially no support if ∆ > 10.
We now turn to some situations that require special consideration.
5.5.1 Avoid using the criteria to compare models with different sets of outlier
regressors when possible
Critical values near 4.0 or larger are usually used to select outlier regressors with a given ARIMA model, see
Table 7.22. (This is done to compensate for level of significance distortions and loss of power resulting from the
large number of tests done by the automatic procedure of the outlier spec.) Outliers that enter the model with
large critical values usually cause the maximum log likelihood to increase quite substantially and AIC and the
other criteria to decrease correspondingly.
As a consequence, unless the models being compared have the same outliers (which often have similar effects
on both log likelihoods and therefore have effects that almost cancel in differences of criterion values), the
outliers can largely determine the model selection, rather than more relevant data properties. In particular, the
model with the most outliers will often be the one with the smallest criterion value. Therefore, with automatic
model selection using the model selection criteria, when the outlier sets are not automatically the same, it can be
important to find out if differences in outlier sets have determined the outcome. This can be done by changing
the specifications of the most competitive models so that all these models have same outlier regressors and then
estimating the modified models and comparing their model selection criteria.
5.5.2 Model comparisons for transformed data
Often a log transformation, or other Box-Cox power transformation,

λ2 + Ytλ − 1 /λ, λ 6= 0
,
ln Yt , λ=0
is applied to the original data Yt prior to regARIMA modeling, see Section 7.18. (Note that the power λ = 1
yields Yt , i.e., no transformation.) Frequently this transformation is preceded by division of the series Yt by
positive prior ratio-adjustment factors ct .
For monthly data, an important example is the leap year preadjustment factors defined by
 28
 28.25 , 28-day months
29
ct = 28.25 , 29-day months , (5.5)
1, other months

(see Section 7.13). When both kinds of transformations are used, then
(
nln Yt − ln ct , o λ=0
yt = 2 λ (5.6)
λ + (Yt /ct ) − 1 /λ, λ 6= 0
is the series for which a regARIMA model is sought.

d
With δ (B) = (1 − B) (1 − B s )D , suppose the observed series is indexed as Yt , −(d + sD) + 1 ≤ t ≤ N , so the
transformed series to which regARIMA model (4.3) is fit is yt , −(d + sD) + 1 ≤ t ≤ N . Thus zt = δ (B) yt , 1 ≤
t ≤ N is the series from which regression and ARMA parameters are estimated by maximizing the regARMA
model’s Gaussian-form log likelihood function LzN (β, φ, Φ, θ, Θ, σa2 ; z1 , . . . zN ). Denoting its maximum value by
LyN as above, a log likelihood for the untransformed data Y1 , . . . YN (conditional1 on the initial observations
Yt , −(d+sD)+1 ≤ t ≤ 0) is obtained by adding to the log likelihood the log Jacobian transformation adjustment 2
PN
t=1 ln |dyt /dYt |. This yields the maximum log likelihood value
N
X
LN = LyN + ln |dyt /dYt | ,
t=1
for the model for Yt , whose definition now includes any data transformation (and/or preadjustments) as well as
the regARIMA specification. N is called the effective number of observations.
For example, for yt given by (5.6), the transformation adjustment is
N
X
ln c−1 λ−1

t (Yt /ct ) ,
t=1
PN −1 PN
which reduces to = − t=1 ln Yt when λ = 0. For the logistic transformation of data Yt preadjusted
t=1 ln Yt
so that 0 < Yt /ct < 1 always holds, we have
Yt /ct Yt
yt = ln= ln ,
1 − Yt /ct ct − Yt
PN
and the transformation adjustment is − t=1 ln c−1
2

t (ct Yt − Yt ) .
To compare different ratio preadjustments and/or different transformations (and perhaps different regression
and ARMA specifications at the same time), we replace LyN by LN in the criterion function formulas for AIC,
−1
n +1
AICC, Hannan-Quinn and BIC above, e.g. AICCN = −2LN + 2np 1 − pN .
1 To provide this interpretation and other properties desirable for signal extraction, the initial d + sD values of the series Y are
t
assumed to be statistically independent of the δ (B) yt , see Bell (1984) and Bell and Hillmer (1988). This is the only statistical
assumption made for these initial variates.
2 Because δ (B) y is a function of Y , s ≤ t, the Jacobian matrix [∂z /∂Y ]
t s t s 1≤s,t≤N is a triangular matrix. Consequently,
det [∂zt /∂Ys ]1≤s,t≤N = ΠN N
t=1 ∂zt /∂Yt = Πt=1 dyt /dYt .
5.5.3 Do not use the criteria to compare models with different differencing oper-
ators
d∗
The preceding discussion shows that a model with a different differencing operator δ ∗ (B) = (1 − B) (1−B s )D∗
such that d∗ +sD∗ 6= d+sD, yields a log likelihood function that is for a set of Yt values different from Y1 , . . . YN
(and that is conditional on a different set of initial values). Therefore its log likelihood function, and hence also
the value of any of the model selection criteria, is not comparable3 to the values of the same criterion obtained
d
with the differencing operator δ (B) = (1 − B) (1 − B s )D . To compare models with different differencing
operators, the out-of-sample forecast error output of the history spec (Section 7.8) can be used, with the
graphical diagnostics discussed in Sections 3 and 4 of Findley, Monsell, Bell, Otto, and Chen (1998) when the
series is long enough that regARIMA models can be estimated reliably without the final two years of data,
which are withheld for forecasting.
3 Ozaki (1977) proposed a rescaling of AIC

N to compare different orders of differencings of nonseasonal ARIMA models. In the
seasonal case, the analogue would be to multiply AICN by (N + d + sD)/N and use the resulting value in all model comparisons.
There is neither theoretical nor systematic emprical support for such a rescaling of any of the criteria, so rescaling this way is not
an accepted practice for model selection.
6 Points Related to Seasonal
Adjustment and Modeling Diagnostics
Contents
6.1 Spectral Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.2 Sliding Spans Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.3 Revisions History Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Tables
6.1 Revision Measure Calculated for Revision Lag Analysis . . . . . . . . . . . . . . . . . 55
The X-13ARIMA-SEATS seasonal adjustment program contains several new diagnostics for modeling, model
selection, adjustment stability, and for judging the quality of indirect as well as direct seasonal adjustments.
This chapter deals specifically with three diagnostics that can be generated by the X-13ARIMA-SEATS program.
• Section 6.1 describes the spectral plots that X-13ARIMA-SEATS produces of the original series, the reg-
ARIMA residuals, the final seasonal adjustment and the final irregular component. The plots are marked
at frequencies commonly associated with seasonal and trading day variation, so the user can easily check
for residual effects in the model residuals or seasonal adjustment. For more information, see Section 2.1
of Findley, Monsell, Bell, Otto, and Chen (1998) and Soukup and Findley (1999).
• Section 6.2 describes the sliding spans diagnostics, which compare seasonal adjustments from overlapping
spans of a given time series. This provides an indication of the stability of the seasonal adjustment.
• Section 6.3 describes revisions history diagnostics, another stability diagnostic. The basic revision is the
difference between the initial seasonal adjustment (often referred to as the concurrent adjustment) and
the seasonal adjustment with all the data available at the time of the analysis (often referred to as the
final adjustment).
6.1 Spectral Plots
X-13ARIMA-SEATS provides spectral plots and associated interpretative messages to alert the user to the presence
of seasonal and trading day effects. Spectral output is available for the original series and as many as three
series resulting from modeling or seasonal adjustment, namely the model residuals, when modeling in specified,
and the adjusted series and irregulars series, when adjustment for seasonal (and possibly also trading day or
holiday) effects is specified.
49
50 CHAPTER 6. POINTS RELATED TO SEASONAL ADJUSTMENT AND MODELING DIAGNOSTICS
For a stationary time series xt with mean µ and autocovariances γk = E (xt − µ) (xt+k − µ), k = 0, ±1, . . .,
the spectral density (spectrum for short) is a non-negative function g (λ), 0 ≤ λ ≤ 1/2, which reformulates the
content of the autocovariances in terms of amplitudes at frequencies of half a cycle per sampling period (month
for our purposes) or less, in such a way that
Z 1/2
γk = g (λ) cos 2πkλ g (λ) dλ, k = 0, ±1, . . . .
0
When xt is a stationary ARMA process with the backshift operator polynomial formula ϕ (B) (xt − µ) =
ϑ (B) at , then its spectrum can be shown to be given by
i2πλ 2

2 ϑ e

g (λ) = σ 2,
|ϕ (ei2πλ )|
where σ 2 is the variance of the white noise at ;√see Priestley (1981). Here ei2πλ = cos 2πλ + i sin 2πλ, and for a
complex number u + iv (u, v real), |u + iv| = u2 + v 2 .
For the first-differenced original series of the series or composite spec (transformed in accord with the
transform spec), the program’s warning message about ”visually significant” seasonal peaks, or the associated
plot, can alert the user to the possibility that the series has a seasonal effect that is predictable (stable) enough
that X-13ARIMA-SEATS can estimate it with reasonable success. (If there are seasonal peaks in the spectrum but
none that meet the criteria for visual significance, see below, then it is likely that any ”seasonal” effects in the
series change too rapidly from year to year or are too obscured by ”noise” to be estimated reliably or stably.)
For the regARIMA model residuals (when a regARIMA model is estimated), and for the first-differenced,
transformed seasonally adjusted series and the irregulars series (when the x11 or seats spec is used), the
messages indicate that the model or adjustment procedure for seasonal or trading day effects has either failed
to capture such effects or, worse, has induced such effects in the series over the time interval used for spectrum
estimation. Because seasonal and trading day patterns can change over time, and because adequate modeling
or adjustment is usually most important for recent data, the time interval of the most recent 96 observations is
the default interval for spectrum estimation (or the time interval specified by the applicable modelspan or span
argument when the latter interval has length less than 96). In the case of trading day peaks a peak (especially
one at the lower of the two trading day frequencies) shows the need for trading day estimation if this was not
done, and otherwise shows that the trading day regression model used is inadequate for the time interval used
for spectrum estimation.
At seasonal frequencies, a peak in the model residuals indicates the need for a better fitting model for the
time interval used for spectrum estimation. A peak in the spectrum from the seasonally adjusted series or
irregulars reveals inadequacy of the seasonal adjustment filters for this interval, thereby indicating that different
filters and/or a shorter data span should be considered. Usually, the spectrum estimator requires 72 data points
to produce peaks sharply defined enough to trigger warning messages for seasonal or trading day effects.
The default spectrum estimator used to detect seasonal and trading day effects is an autoregressive spectral
estimator. For the series xt (for example, the model residuals) whose spectrum is being estimated from data
x1 , . . . , xN , autoregressive log-spectrum estimates (in decibel units) have the form
 
2
 
 σ̂m 
ŝ (λ) = 10 log10 2 , 0 ≤ λ ≤ 0.5, (6.1)
 2π 1 − Pm φ̂j ei2πjλ 
 
j=1
6.1. SPECTRAL PLOTS 51
where the coefficient estimates φ̂j are those of the linear regression of xt − x̄ on xt−j − x̄, 1 ≤ j ≤ m for the
PN
data, with x̄ = N −1 t=1 xt , and where σ̂m 2
is the sample variance of the resulting regression residuals. For
large enough m (and N ), a strong component with period 1/λ0 results in a near-zero value of the denominator
of (6.1) at λ0 and therefore in a peak at λ0 in the graph of (6.1) - unless there is a stronger periodic component
at a nearby frequency. For a discussion of this estimator, see pp. 600–612 of Priestley (1981). Application of
the proof of Corollary 5.6.3 of Brillinger (1975) to the results of Theorem 6 of Berk (1974) shows that, under
Berk’s assumptions, which include invertibility of xt , the log transformation in (6.1) stabilizes the large sample
2
variance of ŝ(λ), i.e. the limiting value of E (ŝ(λ) − g (λ)) , as m → ∞ in case g (λ) is not an AR process.
However, the constant value for the end point frequencies λ = 0, 1/2 is twice as large as the constant value for
the intermediate frequencies. 0 < λ < 1/2.
X-13ARIMA-SEATS uses m = 30 for monthly series, which yields high resolution of strong components,
meaning peaks that are sharply defined in the main output file’s plot of ŝ (λ) (Recall that for the spectra providing
information about the original series and the seasonal adjustment, the series xt results from suppression of a
trend component by differencing or detrending. Trends produce peaks at and near λ = 0 that are so dominant
that they diminish the resolution of all other peaks.) The spectrum plots of X-13ARIMA-SEATS show values of
ŝ (λ) at 61 frequencies that have the form λk = k/120, 0 ≤ k ≤ 60, with two exceptions: for the values k/120
closest to the trading day frequencies (0.348, and 0.432 cycles per month for monthly series), λk is assigned
the value of the trading day frequency instead of the value k/120. At trading day frequencies, values of ŝ (λk )
are plotted with a column of T’s. At seasonal frequencies (1/12, 2/12,...,6/12 cycles per month for monthly
series) values of ŝ (λk ) are plotted with a column of S’s. At all other frequencies, columns of asterisks (”stars”)
are used. These plots are very similar to those of the BAYSEA seasonal adjustment program (Akaike 1980 and
Akaike and Ishiguro 1980) and are produced by a modified version of BAYSEA’s Fortran code.
The monthly trading day frequency 0.348 can be derived by noting that a daily component which repeats
.
every seven days goes through 4.348 = 30.4375/7 cycles in a month of average length, 365.25/12 = 30.4375 days.
It is therefore seen to advance 0.348 cycles per month when the data are obtained at twelve equally spaced times
in an 365.25 days, the average length of a year. The connection of peaks at 0.432 cycles/month with trading
day components is weaker, see Cleveland and Devlin (1980), and not as reliable.
Note that in X-13ARIMA-SEATS there needs to be at least 60 observations in the time series to display the
trading day frequencies in the plots and attempt to identify peaks for trading day frequencies.
Because of difficulties associated with statistical significance tests for periodic components in autocorrelated
data, see Chapter 8 of Priestley (1981), such tests are not used. The warning messages of X-13ARIMA-SEATS
are based on an empirically obtained criterion of ”visual significance” determined as follows from the range
ŝmax − ŝmin of the ŝ (λk ) values, where ŝmax = maxk ŝ (λk ) and ŝmin = mink ŝ (λk ). To be ”visually significant”,
the value ŝ (λk ) at a trading day or seasonal frequency λk (other than the seasonal frequency λ60 = 0.5) must
be above the median of the plotted values of ŝ (λ) and must be larger than both neighboring values ŝ (λk−1 )
and ŝ (λk+1 ) by at least 6/52 times the range ŝmax − ŝmin . In the main output file’s line printer plots of spectra,
ŝmax is plotted 52 lines above ŝmin , so a visually significant peak must be at least six lines (six ”stars”) high.
Peak of any size at λ60 = 1/2 are ignored, and when there are no other v.s. peaks at other seasonal
frequencies, so are peaks at λ50 = 5/12 for cycles of 2.4 repetitions a year. The theoretical results from Berk
(1974) and Brillinger (1975) mentioned above describe how ŝ(λ) will be more randomly variable at λ60 = 1/2
than at other seasonal frequencies. Worse behavior is to be expected when the invertibility assumption required
for this precise result fails, as it may in practice. (For example, the bi-infinite sample models for all model-based
seasonal decomposition components from a seasonal ARIMA model always have non-invertible moving average
factors arising from the ARIMA model’s differencing polynomial.) The empirical finding from practice is that
v.s. peaks at λ60 = 1/2 occur too often in the spectra of seasonal adjusted and irregulars series that have few
or no other v.s. seasonal peaks. Also, v.s. peaks at λ50 = 5/12 rather frequently occur when there are no other
v.s. peaks. There is no economic explanation for such peaks. The most recent versions of TRAMO-SEATS and
TSW ignore peaks at these two frequencies.
6.2 Sliding Spans Diagnostics

The sliding spans diagnostics are described in detail and compared with other quality diagnostics in the articles
Findley, Monsell, Shulman, and Pugh (1990) and Findley and Monsell (1986). An abbreviated presentation
will be given here. The basic diagnostics are descriptive statistics of how the seasonal adjustments and their
month-to-month changes vary when the span of data used to calculate them is altered in a systematic way: any
two neighboring spans differ to the extent that one starts and ends a year later than the other. The span length
is determined by the length of the seasonal filter utilized for the adjustment. The ending date of the last span is
usually the date of the most recent datum in the time series. Four spans are used if enough data are available.
The index value j = 1 is assigned to the span with the earliest starting date, j = 2 to the span with the next
earliest starting date, . . . .
For series whose seasonally adjusted values are all positive, the two most important sliding spans statistics,
A(%) and M M (%), are calculated as follows. For a month t belonging to at least two spans, one of which is the
j-th span, let At denote its seasonally (and, if applicable, trading day and holiday) adjusted value obtained from
the complete series, and let Ajt denote the adjusted value obtained when the seasonal adjustment procedure
being considered (the procedure determined by the software options selected) is applied to only data in the j-th
span. The seasonal adjustment At is called (unacceptably) unstable if
maxj Ajt − minj Ajt

> .03. (6.2)
minj Ajt
Further, for months t such that both t and t − 1 belong to at least two spans, the “seasonally adjusted month-
to-month percent change” 100 × (At − At−1 ) /At−1 is called unstable if
Ajt Ajt
max − min > .03. (6.3)
j Ajt−1 j Ajt−1
In (6.2), the index j ranges over all spans containing month t; in (6.3) the j-th span must contain month t − 1
also.
A (%) is used to denote the percent of months with unstable adjustments calculated with respect to the
number of month for which the left hand side of (6.2) is defined (the number of months common to at least
two spans). The analogous quantity for (6.3) is denoted M M (%). We recommend that, except in special
circumstances of the sort discussed below, the seasonal adjustment produced by the procedure chosen should
not be used if A (%) > 25.0 (> 15.0 is considered problematic) or if M M (%) > 40.0.
There is a similarly defined statistic Y Y (%) for year-to-year percent changes in the seasonally adjusted data,
100 × (At − At−12 ) /At−12 , based on the same threshold used to define unstable adjustments and month–to-
month changes, usually the default .03 shown in (6.2) and (6.3). Because these year-to-year changes in the
6.3. REVISIONS HISTORY DIAGNOSTICS 53
adjusted series can be misleading indicators of trend direction when turning points occur between months t
and t − 12, they are rather less important than the adjusted values themselves and month-to-month changes in
the adjusted values. The statistic Y Y (%) is correspondingly less important than the others, but it is included
in the output of X-13ARIMA-SEATS because of the interest some data users have in year-to-year changes. The
output text describes values of Y Y (%) greater than 10.0 as extreme, but this information is usually redundant
in the sense that series with such a value have, in our experience, usually also had excessive values of A(%) or
M M (%). In any case, we would not reject an adjustment based solely on the value of Y Y (%).
Sometimes, the causes of large values of A(%) or M M (%) can be identified and seen to be not very prob-
lematic. For example, this could be the case when the months with unstable adjustments or changes are heavily
concentrated in a known problem period several years back from the current year, or in one or two fixed calendar
months each year that all data users can be expected to regard as quite problematic, such as winter months in
series known to be very sensitive to differences in winter weather conditions. The sliding spans output makes it
easy to identify such concentrations.
The output can show when a “mild” increase in the threshold beyond 0.03 will dramatically decrease the
values of A(%) and M M (%) to quite acceptable levels: we have identified a few series for which increasing the
threshold to .05 seemed justifiable, because most of the months for which the left hand sides of (6.2) and (6.3)
were between .03 and .05 were months with very large seasonal movements, where users would be tolerant of
more uncertainty, and not many months had values of these statistics substantially larger than .05.
This experience stimulated us to carry out a limited exploratory study with a variety of Census Bureau
series focussed on the goal of finding a statistical relationship between appropriate threshold values and seasonal
factor size, a relationship we could use to increase or decrease the threshold according to the size of the seasonal
movements. However, within the set of series considered, we found no correlation between appropriate threshold
values and the size of the seasonal movements. For example, there were relatively many series with quite large
seasonal movements for which good values of A(%) and M M (%) obtained with the .03 threshold and there
were a number of series with only moderately large seasonal movements for which the use of the .05 threshold
did not lead to acceptable values of A(%) and M M (%). In fact, simulation experiments readily show that in a
series with fixed seasonal effects (every January has the same seasonal factor, etc.), the values of the seasonal
adjustment are quite sensitive to the variability of the irregulars component and quite insensitive to the size of
the seasonal movements.
More often than not, when a choice of adjustment options for a series produces an adjustment that sliding
spans diagnostics classify as unacceptable, there will be a different choice of options, perhaps with different
seasonal filter lengths, or different trading day adjustment or forecast extension options, that will result in an
adjustment that is classified as acceptable. When no choice of options produces an acceptable adjustment, the
issue is not whether the series is “seasonal” in some sense, but whether its seasonal behavior is repetitive enough,
or revealed clearly enough in the available time series data, that it can be estimated with adequate reliably by
X-13ARIMA-SEATS under any of the options considered.
6.3 Revisions History Diagnostics

X-13ARIMA-SEATS generates revisions between the initial estimate and the most recent estimate, for several
quantities derived from seasonally adjusting a time series (see Table 7.17). X-13ARIMA-SEATS can also generate
historical out-of-sample forecast errors and likelihood statistics derived from regARIMA model estimation. For
some supporting theory for out-of-sample squared forecast error diagnostic output, see Findley (2005). These
revisions and historical values are obtained as follows.
For a given series yt where t = 1, . . . , T , we define At|n to be the seasonal adjustment of yt calculated from
the series y1 , y2 , . . . , yn , where t ≤ n ≤ T . The concurrent seasonal adjustment of observation t is At|t and the
most recent or ”final” adjustment of observation t is At|T . The percent revision of the seasonally adjusted series
is defined to be
At|T − At|t
Rt =
At|t
and this is what is reported by the program. The revisions of the trend component and of seasonal factors
derived from multiplicative or log-additive seasonal adjustment are also reported as percent revisions.
With additive seasonal adjustments, Rt is calculated the same way if all values At|t have the same sign (the
analogous statement holds for trends). Otherwise, differences are calculated:
Rt = At|T − At|t
In the additive adjustment case, revisions of seasonal factors are always calculated as differences, St|T − St|t ,
or, with projected seasonal factors St|T − St|t∗ , where t∗ denotes ending date of the series used to obtain the
projected factor for month t.
Let Ct|n denote the month-to-month (or quarter-to-quarter) change in the seasonally adjusted series at time
t calculated from the series y1 , y2 , . . . , yn , or
At|n − At−1|n
Ct|n =
At−1|n
The revision of these changes is defined to be
Rt = Ct|T − Ct|t
Revisions for the month-to-month changes in the trend component are computed in the same manner.
The sadjlags and trendlags arguments produce an analysis of the revisions history for different lags past
the concurrent observation. The target for this revisions analysis depends on the value of the target argument.
Table 6.1 shows how the lagged revisions are calculated for for the different values of target.
If lags corresponding to one and two years (12 and 24 for monthly data, 4 and 8 for quarterly data) are
included in sadjlags, then the revision between the seasonal adjustment calculated one year after time t and
the adjustment 2 years after time t is also calculated, or:
At|t+24 − At|t+12
RYt =
At|t+12
for monthly series. This is done only for the seasonally adjusted series and the month-to-month (quarter-to-
quarter) change of the seasonally adjusted series.
The analysis of the lagged revisions can give a useful picture of the behavior of the revisions over time. Using
the concurrent estimate as the target shows how much a given adjustment changes as you add more data; using
the final estimate as the target shows how quickly a given estimate converges to the final value.
6.3. REVISIONS HISTORY DIAGNOSTICS 55
Table 6.1: Revision Measure Calculated for Revision Lag Analysis
Estimate Concurrent Target Final Target

Seasonally Adjusted Series (At|t+lagi − At|t )/At|t (At|T − At|t+lagi )/At|t+lagi
Final Trend Component (Tt|t+lagi − Tt|t )/Tt|t (Tt|T − Tt|t+lagi )/Tt|t+lagi
Change in Seasonally Adjusted Series (or Trend) Ct|t+lagi − Ct|t Ct|T − Ct|t+lagi
Estimate gives the estimate from the seasonal adjustment.
Concurrent Target gives the formula for the lagged revision history where the target is assumed to be the
concurrent estimate.
Final Target gives the formula for the lagged revision history where the target is assumed to be the final estimate.
At|i is the value of the seasonally adjusted series at time t calculated from the series up to time i.
Tt|i is the value of the trend component at time t calculated from the series up to time i.
Ct|i is the value of the change in the seasonally adjusted series at time t calculated for the series up to time i.
Another motivation for the sadjlags and trendlags options is the fact that concurrent estimates are often
based on preliminary data for the current month (or quarter). If the final data for the month are not available
until two additional months have passed, then it would be appropriate to set sadjlags = 2 in order to study
the revisions to the adjustment based on the final datum for each month. For trends, there is the additional
motivation that concurrent trend estimates are often unstable. For this reason, some analysts wait until several
subsequent months of data are available for trend estimation before examining the X-13ARIMA-SEATS trend for
a recent month. For an analyst who waits three months, trendlags = 3 will provide the revisions of the trend
estimates of interest.
7 Documentation for Individual Specs
Contents
7.1 ARIMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.2 AUTOMDL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3 CHECK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7.4 COMPOSITE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7.5 ESTIMATE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.6 FORCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.7 FORECAST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.8 HISTORY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.9 METADATA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.10 IDENTIFY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.11 OUTLIER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.12 PICKMDL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.13 REGRESSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
7.14 SEATS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.15 SERIES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.16 SLIDINGSPANS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.17 SPECTRUM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.18 TRANSFORM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.19 X11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.20 X11REGRESSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
Tables
7.1 Available Output Tables for Automdl . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.2 Available Log File Diagnostics for Automdl . . . . . . . . . . . . . . . . . . . . . . . . 66
7.3 Available Output Tables for Check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.4 Available Log File Diagnostics for Check . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.5 Default Output Tables for Composite Spec . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.6 Other Output Tables for Composite Spec . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.7 Tables Saved As Percentages in the save Argument . . . . . . . . . . . . . . . . . . . 83
7.8 Available Log File Diagnostics for Composite . . . . . . . . . . . . . . . . . . . . . . . 84
7.9 Default Output Tables for Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.10 Other Output Tables for Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.11 Available Log File Diagnostics for Estimate . . . . . . . . . . . . . . . . . . . . . . . . 89
7.12 Example of ARMA Roots Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.13 Default Output Tables for Force spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
56
57

7.15 Choices for the target Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.16 Available Output Tables for Forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.17 Choices Available for the estimates Argument . . . . . . . . . . . . . . . . . . . . . . . 107
7.18 Default Output Tables for History Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.19 Other Output Tables for History Spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.20 Available Log File Diagnostics for History . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.21 Available Output Tables for Identify . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.22 Default Critical Values for Outlier Identification . . . . . . . . . . . . . . . . . . . . . 125
7.23 Available Output Tables for Outlier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.24 Available Output Tables for Pickmdl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.25 Available Output Tables for Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.26 Available Log File Diagnostics for Regression . . . . . . . . . . . . . . . . . . . . . . . 138
7.27 Predefined regression variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.28 Change of Regime Regressor Types and Syntax . . . . . . . . . . . . . . . . . . . . . . 145
7.29 500 Year (1600-2099) means for Easter regressors of different window length w. . . . . . . . . 148
7.30 Output Tables for Seats spec that can be Printed and Saved . . . . . . . . . . . . . 158
7.31 Output Tables for Seats spec that can Only be Saved . . . . . . . . . . . . . . . . . . 159
7.33 Available Log File Diagnostics for Seats . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.34 X-13ARIMA-SEATS File Extensions for Special SEATS Saved Output . . . . . . . . . . 162
7.35 Available Output Tables for Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
7.36 Default Formats for Each X-11 Format Code . . . . . . . . . . . . . . . . . . . . . . . . 172
7.37 Default Output Tables for Slidingspans Spec . . . . . . . . . . . . . . . . . . . . . . . . 181
7.38 Other Output Tables for Slidingspans Spec . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.39 Sliding span lengths for different seasonal filters chosen . . . . . . . . . . . . . . . . . . . . . . 183
7.40 Seasonal MA parameter greater than 0 at which the span length increases to the value indicated.184
7.41 Available Output Tables for Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.42 Available Log File Diagnostics for Spectrum . . . . . . . . . . . . . . . . . . . . . . . . 190
7.43 Choices Available for the series Argument . . . . . . . . . . . . . . . . . . . . . . . . 190
7.44 Transformations Available Using the function Argument . . . . . . . . . . . . . . . . 195
7.45 Available Output Tables for Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
7.46 Default Output Tables for X11 spec . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
7.47 Other Output Tables for X11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.48 Plots Specified by the print Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
7.50 Available Log File Diagnostics for X11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
7.51 X-13ARIMA-SEATS Seasonal Filter Options and Descriptions . . . . . . . . . . . . . . . . 209
7.52 Modes of Seasonal Adjustment and Their Models . . . . . . . . . . . . . . . . . . . . 212
7.53 Number of Surrounding SI-ratios in Table D 8.B Assumed Affected by a Level Shift215
7.54 Default Output Tables for X11regression . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.55 Other Output Tables for X11regression . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
7.56 Predefined regression variables for x11regression . . . . . . . . . . . . . . . . . . . . . 229
58 CHAPTER 7. DOCUMENTATION FOR INDIVIDUAL SPECS
The following pages provide detailed documentation on all of the specs, with discussions of the available
arguments and their default values. Each spec’s documentation also includes several examples illustrating its
use. For the series and transform specs the examples are intended only to illustrate the capabilities of these
specs. They do not show complete spec files in the sense that if these examples were used as input to the
X-13ARIMA-SEATS program, they would produce no useful output. For the remaining specs (composite, x11,
identify, regression, arima, estimate, outlier, check, forecast, metadata, spectrum, slidingspans, and
history) the examples all show complete spec files that could be used, except that data sets (e.g., for the input
series appearing in the series spec, or for a user-defined regression variable in the regression spec) are often
abbreviated using the . . . notation.
Readers will notice that the examples for a given spec tend to vary, not only in content, but also in format.
This is done deliberately to illustrate and emphasize the flexibility the user has in formatting the spec file.
The next few paragraphs will give the reader a summary of what specs to include in the input file when
doing general tasks (such as a simple seasonal adjustment or modeling run). Except in certain default situations,
arguments must be specified within each spec to accomplish these tasks. Information about these arguments
can be formed within the sections of this chapter devoted to the individual specs.
For the reader who wants the shortest path to a seasonal adjustment, the essential specs are series and x11.
These will yield a default X-11 seasonal adjustment. If it is not clear whether the seasonal adjustment should
be additive or multiplicative then the transform spec should be added. If an elementary approach to trading
day and moving holiday effect estimation and adjustment is desired, then add x11regression. The sliding
spans and history specs provide diagnostics for the stability of the adjustment when the span of data used to
calculate the adjustment changes.
For the reader wanting the shortest path to modeling a time series, the essential specs are series, automdl
(or pickmdl), and possibly transform. Add the forecast spec if forecasting is desired, add outlier if there
are problematic data values or data movements, and add regression if trading day or holiday components
may be present in the series. The arima spec replaces automdl if custom rather than automatic modeling is
desired. It is supported by identify. The check spec provides standard model-fit diagnostics. The history
spec provides forecasting diagnostics for comparing two models, and estimate offers estimation options and
the ability to reuse stored models.
Time series models (obtained via automdl/pickmdl/arima and transformation can improve seasonal
adjustment by extending the data with forecasts (via forecast), by providing a way of dealing with disruptions
to the level of the series (via outlier) and by providing estimates of trading day and holiday effects (via
regression) that are sometimes better than those obtained from x11regression.
The composite spec is required to obtain the indirect adjustment of an aggregate series from adjustments
of its components and to compare this adjustment with its direct adjustment. of its components and to compare
this adjustment with its direct adjustment. For indirect adjustment the composite spec replaces the series
spec.
7.1. ARIMA 59
7.1 ARIMA
DESCRIPTION
Specifies the ARIMA part of the regARIMA model. This defines a pure ARIMA model if the regression spec
is absent. The ARIMA part of the model may include multiplicative seasonal factors and operators with missing
lags. Using the ar and ma arguments, initial values for the individual AR and MA parameters can be specified
for the iterative estimation. Also, individual parameters can be held fixed at these initial values while the rest
of the parameters are estimated.
USAGE
arima { model = ([2 3] 1 1)(0 1 1)12
title = "ARIMA Model"
ar = (0.3f, -0.14)
ma = (-0.7 0.85f) }
ARGUMENTS
ar Specifies initial values for nonseasonal and seasonal autoregressive parameters in the
order that they appear in the model argument. If present, the ar argument must assign
initial values to all AR parameters in the model. Initial values are assigned to parameters
either by specifying the value in the argument list or by explicitly indicating that it is
missing. Missing values take on their default value of 0.1. For example, for a model
with two AR parameters, ar=(0.7, ) is equivalent to ar=(0.7,0.1), but ar=(0.7) is
not allowed. For a model with three AR parameters, ar=(0.8,,-0.4) is equivalent to
ar=(0.8,0.1,-0.4). To hold a parameter fixed during estimation at its initial value,
immediately follow the value in the ar list with an ‘f ’, e.g., ar=(0.7f, 0.1).
ma Specifies initial values for all moving average parameters in the same way ar does so for
autoregressive parameters.
model Specifies the ARIMA part of the model. The format follows standard Box-Jenkins (1976)
notation. In this notation a nonseasonal ARIMA model is specified as (p d q), where p is
the nonseasonal AR order, d is the number of nonseasonal differences, and q is the nonsea-
sonal MA order. A multiplicative seasonal ARIMA model is specified as (p d q)(P D Q),
where p, d, and q are as before, P is the seasonal AR order, D is the number of seasonal
differences, and Q is the seasonal MA order. Here, the first ARIMA factor, (p d q), is as-
sumed to be nonseasonal (i.e., its period is one) and the second ARIMA factor, (P D Q),
is assumed to be seasonal with the seasonal period set in the series spec. More than two
ARIMA factors can be specified, and ARIMA factors can explicitly be given seasonal
periods that differ from the default choices. See DETAILS for more information.
The operator orders (p d q) in the ARIMA factors may be separated by spaces or commas,
e.g., (0 1 1) is the same as (0,1,1). Operators with missing lags are specified by
enclosing those lags present in brackets, with the lags in ascending order. For example,
model = ([2 3] 0 0) specifies the model (1 − φ2 B 2 − φ3 B 3 )zt = at .
print and save No output tables are available for this spec.
title Specifies a title for the ARIMA model, in quotes. It must be less than 80 characters.
The title appears above the ARIMA model description and the table of estimates. The
default is to print ARIMA Model.
DETAILS
The arima spec cannot be used in the same spec file as the pickmdl or automdl specs; the model, ma, and
ar arguments of the arima spec cannot be used when the file argument is specified in the estimate spec.
The model argument may include as many ARIMA factors as desired. However, there is a limit of 108 total
AR, MA, and differencing coefficients in the model. Also, the maximum lag of any AR or MA parameter is 36,
and the maximum number of differences in any ARIMA factor (nonseasonal or seasonal) is 3. (The latter two
limits can be changed—see Section 2.8.)
In general, ARIMA factors are specified in the standard (p d q)s format, where s is the seasonal period of the
operator. Thus, putting (0 1 1)6 in the model argument includes differencing by 1 − B 6 and a 1 − θ6 B 6 MA
term in the model. However, if the seasonal period s is not specified after an ARIMA factor, it is determined
according to the following default rules. The first ARIMA factor without a specified seasonal period is assumed
to be nonseasonal, i.e., its seasonal period is assumed to be one. The second ARIMA factor without a specified
seasonal period is assumed to be a seasonal factor with the seasonal period set in the series spec. For example,
if period = 12 is specified in the series spec (or if the period is set to 12 because the start date there is given
as year.month), then model = (0 1 1)(0 1 1) and model = (0 1 1)1(0 1 1)12 are equivalent. If additional
ARIMA factors are specified, these are assumed to be nonseasonal unless they are explicitly given a seasonal
period. See Example 7.1 for an illustration of a model with three ARIMA factors. Note that if the seasonal
period is one, then any ARIMA factors without a specified seasonal period have period one.
Users should not specify initial values for MA parameters that yield an MA polynomial with roots inside the
unit circle. (See Section 5.4.) Doing so will cause the program to stop and print an error message asking the
user to re-specify the initial parameters and rerun the program. Initial parameters that yield an MA polynomial
with roots on the unit circle are allowed only if this non-invertible polynomial is not being estimated. That
is, this is allowed if no estimation is being done, or if the parameters in this polynomial are specified as fixed
during estimation. For example, if a model has a first order seasonal MA parameter as the only MA parameter,
then ma=(1.0f) is always allowed, ma=(1.0) is allowed only if no estimation is done, and ma=(1.1) is never
allowed.
If the likelihood function that is exact for AR polynomials is used (exact=arma, which is the default—see
the estimate spec), users should not specify initial values for AR parameters that yield a non-stationary AR
polynomial (one with roots on or inside the unit circle). Doing so will cause the program to stop and print an
error message asking the user to re-specify the initial parameters and rerun the program.
The use of fixed coefficients in the ARIMA model can invalidate AIC and the other model selection statistics
as well as some goodness-of-fit diagnostics, see the Details sections of estimate and check.
7.1. ARIMA 61
EXAMPLES
The following examples show complete spec files.
Example 1 Specify and estimate a nonseasonal ARIMA model with a first difference and an MA pa-
rameter at lag 1, i.e., (1 − B)yt = (1 − θB)at .
series { title = "Quarterly Grape Harvest" start = 1950.1

period = 4
data = (8997 9401 ... 11346) }
arima { model = (0 1 1) }
estimate { }
Example 2 Specify and estimate the following seasonal ARIMA model for yt , the logarithm of an original
time series: (1 − φ1 B − φ2 B 2 )(1 − B)(1 − B 12 )yt = (1 − Θ12 B 12 )at . Note that the start date
in the series spec specifies a month, which sets the seasonal period to 12.
series { title = "Monthly sales" start = 1976.jan

data = (138 128 ... 297) }
transform { function = log }
arima { model = (2 1 0)(0 1 1) }
estimate { }
Example 3 Specify and estimate a regARIMA model with fixed seasonal effects, a trend constant,
P and
the ARIMA (0 1 1) model for the regression errors. The model is then (1−B) yt − βi Mit −
c · t = (1 − θB)at , where the Mit are the fixed seasonal effect regression variables.
Series { Title = "Monthly Sales" Start = 1976.jan

Data = (138 128 ... 297) }
Transform { Function = log }
Regression { Variables= (seasonal const) }
Arima { Model = (0 1 1) }
Estimate { }
Example 4 Specify and estimate a model with one difference and an AR(2) operator with lag one
missing; i.e., the model is (1 − φ2 B 2 )(1 − B)yt = at .
series{title = "Annual Olive Harvest" start = 1950

data = (251 271 ... 240) }
arima{model = ([2] 1 0)}
estimate{ }
Example 5 Specify and estimate a model with a trend constant and with regression errors zt following
an ARIMA model with one seasonal difference and a first order seasonal moving average,
but no nonseasonal factor, i.e., (1 − B 12 )zt = (1 − ΘB 12 )at . Note that the seasonal period
of the ARIMA factor must be given explicitly in the model argument, because, as there is
only one ARIMA factor, it would otherwise be assumed to be nonseasonal.

data = (138 128 ... 297) }
regression { variables = const }
arima { model = (0 1 1)12 }
estimate { }
Example 6 Specify and estimate a model including three ARIMA factors. The ARIMA model for the
regression errors zt is (1 − φ1 B)(1 − φ3 B 3 )(1 − B)zt = (1 − ΘB 12 )at . The 1 − φ3 B 3 operator
might be used to account for quarterly autocorrelation since each quarter is comprised of
three months. Note that only the period of the quarterly factor need be given.

data = (138 128 ... 297) }
regression { variables = (const seasonal)}
arima { model = (1 1 0)(1 0 0)3(0 0 1)}
estimate { }
Example 7 Specify and estimate a model with regression errors zt following the “airline model”, ARIMA
(0 1 1)(0 1 1)12 , with the seasonal MA parameter fixed at 1.0. The model used for zt is
(1 − B)(1 − B 12 )zt = (1 − θB)(1 − 1.0B 12 )at . The initial value of 0.1 used for θ is indicated
by a missing value in the ma list. This model is actually equivalent to that used in Example
3, since it results from overdifferencing the model specified there by 1 − B 12 . (See Section
5.4 for a discussion of overdifferencing.)

data = (138 128 ... 297) }
transform{ function = log }
arima { model = (0 1 1)(0 1 1)12
ma = ( ,1.0f)}
estimate { }
7.2. AUTOMDL 63
7.2 AUTOMDL
DESCRIPTION
Specifies that the ARIMA part of the regARIMA model will be sought using an automatic model selection
procedure derived from the one used by TRAMO (see Gómez and Maravall (2001a)). The user can specify the
maximum ARMA and differencing orders to use in the model search, and can adjust thresholds for several of
the selection criteria.
USAGE
automdl { maxorder = (3 1)
maxdiff = (1 1) or diff = (1 0)
acceptdefault = no
checkmu = yes
ljungboxlimit = 0.99
mixed = yes
print = (none bestfivemdl autochoice)
savelog = automodel
}
ARGUMENTS
acceptdefault Controls whether the default model is chosen if the Ljung-Box Q statistic for its model
residuals (checked at lag 24 if the series is monthly, 16 if the series is quarterly) is
acceptable (acceptdefault = yes). If the default model is found to be acceptable, no
further attempt will be made to identify a model or differencing order. The default for
acceptdefault is acceptdefault = no.
checkmu Controls whether the automatic model selection procedure will check for the significance
of a constant term (checkmu = yes), or will maintain the choice of the user made by the
user in the regression spec (checkmu = no). The default for checkmu is checkmu =
yes.
diff Fixes the orders of differencing to be used in the automatic ARIMA model identification
procedure. The diff argument has two input values, the regular differencing order and
the seasonal differencing order. Both values must be specified; there is no default value.
Acceptable values for the regular differencing orders are 0, 1 and 2; acceptable values
for the seasonal differencing orders are 0 and 1. If specified in the same spec file as
the maxdiff argument, the values for the diff argument are ignored and the program
performs automatic identification of nonseasonal and seasonal differencing with the limits
specified in maxdiff.
fcstlim Sets the acceptance threshold for the within-sample forecast error test of the final identi-
fied model. The absolute average percentage error of the extrapolated values within the
last three years of data must be less than this value for forecasts to be generated with
the final model. For example, fcstlim=20 sets this threshold to 20 percent. The value
entered for this argument must not be less than zero, or greater than 100. This option
is only active when rejectfcst = yes. The default for fcstlim is 15 percent.
ljungboxlimit Acceptance criterion for confidence coefficient of the Ljung-Box Q statistic. If the Ljung-
Box Q for the residuals of a final model (checked at lag 24 if the series is monthly,
16 if the series is quarterly) is greater than ljungboxlimit, the model is rejected, the
outlier critical value is reduced, and model and outlier identification (if specified) is
redone with a reduced value (see reducecv argument). The default for ljungboxlimit
is ljungboxlimit = 0.95.
maxdiff Specifies the maximum orders of regular and seasonal differencing for the automatic
identification of differencing orders. The maxdiff argument has two input values, the
maximum regular differencing order and the maximum seasonal differencing order. Ac-
ceptable values for the maximum order of regular differencing are 1 or 2, and the accept-
able value for the maximum order of seasonal differencing is 1. If specified in the same
spec file as the maxdiff argument, the values for the diff argument are ignored and the
program performs automatic identification of nonseasonal and seasonal differencing with
the limits specified in maxdiff. The default is maxdiff = (2 1).
maxorder Specifies the maximum orders of the regular and seasonal ARMA polynomials to be
examined during the automatic ARIMA model identification procedure. The maxorder
argument has two input values, the maximum order of regular ARMA model to be tested
and the maximum order of seasonal ARMA model to be tested. The maximum order for
the regular ARMA model must be greater than zero, and can be at most 4; the maximum
order for the seasonal ARMA model can be either 1 or 2. The default is maxorder = (2
1).
mixed Controls whether ARIMA models with nonseasonal AR and MA terms or seasonal AR
and MA terms will be considered in the automatic model identification procedure (mixed
= yes). If mixed = no, mixed models would not be considered. Note that a model with
AR and MA terms in both the seasonal and nonseasonal parts of the model can be ac-
ceptable, provided there are not AR and MA terms in either the seasonal or nonseasonal.
For example, when mixed = no an ARIMA (0 1 1)(1 1 0) model would be considered,
while an ARIMA (1 1 1)(0 1 1) model would not, since there are AR and MA terms in
the nonseasonal part of the model. The default for mixed is mixed = yes.
print The tables available for output are listed in Table 7.1. The save option is not available
for this spec. The header, autochoice, and unitroottest tables are printed out by
default. For a complete listing of the brief and default print levels for this spec, see
Appendix B.
rejectfcst If rejectfcst = yes, then a test of the out-of-sample forecast error of the final three
years of data will be generated with the identified model to determine if forecast extension
will be applied. If the forecast error exceeds the value of fcstlimit, forecasts will not
be generated with the final identified model, but the model will be used to generate
preadjustment factors for calendar and outlier effects. The default is rejectfcst=no.
savelog The diagnostics available for output to the log file (see section 2.6) are listed on Table
7.2. AUTOMDL 65
Table 7.1: Available Output Tables for Automdl
name short description of table

autochoice ach model choice of automatic model procedure
autochoicemdl amd summary output for models estimated during choice of
ARMA model orders
autodefaulttests adt tests performed on the default model (usually the air-
line model) of the automatic model identification pro-
cedure
autofinaltests aft final tests performed on the model identified by au-
tomdl
autoljungboxtest alb check of the residual Ljung-Box statistic
bestfivemdl b5m summary of best five models found during choice of
ARMA model orders
header hdr header for the automatic modeling output
unitroottest urt choice of differencing
unitroottestmdl urm summary output for models estimated during difference
order identification
Name gives the name of each table for use with the print argument.
Short gives a short name for the tables of the print argument.
7.2.
RARELY USED ARGUMENTS

armalimit Threshold value for t-statistics of ARMA coefficients used for final test of model parsi-
mony. If the highest order ARMA coefficient has a t-value less than this value in magni-
tude, the program will reduce the order of the model. The value given for armalimit is
also used for the final check of the constant term; if the constant term has a t-value less
than armalimit in magnitude, the program will remove the constant term from the set
of regressors. This value should be greater than zero. The default is armalimit = 1.0.
balanced Controls whether the automatic model procedure will have a preference for balanced
models (where the order of the combined AR and differencing operator is equal to the
order of the combined MA operator). Setting balanced = yes yields the same preference
as the TRAMO program. The default is balanced = no.
exactdiff Controls if exact likelihood estimation is used when Hannen-Rissanen fails in automatic
difference identification procedure (exactdiff = yes), or if conditional likelihood es-
timation is used (exactdiff = no). The default is to start with exact likelihood es-
timation, and switch to conditional if the number of iterations for the exact likelihood
procedure exceeds 200 iterations (exactdiff = first).
hrinitial Controls whether Hannan-Rissanen estimation is done before exact maximum likelihood
estimation to provide initial values when generating likelihood statistics for identifying
Table 7.2: Available Log File Diagnostics for Automdl

name short description of diagnostic
alldiagnostics all All modeling diagnostics listed in this table
autodiff adf choice of differencing by automatic model identification
procedure
automodel amd choice of ARIMA model by automatic model identification
procedure
bestfivemdl b5m summary for best five models found during choice of
ARMA model orders
mean mu choice regarding use of constant term with automatically
identified model
Name gives the name of each diagnostic for use with the savelog argument.
Short gives a short name for the diagnostics of the savelog argument.
the ARMA orders (hrinitial = yes). If hrinitial = yes, then models for which
the Hannan-Rissanen estimation yields coefficients that are unacceptable initial values
to the exact maximum likelihood estimation procedure will be rejected. The default is
hrinitial = no.
reducecv The percentage by which the outlier critical value will be reduced when an identified
model is found to have a Ljung-Box Q statistic with an unacceptable confidence coeffi-
cient. This value should be between 0 and 1, and will only be active when automatic out-
lier identification is selected. The reduced critical value will be set to (1−reducecv)×CV ),
where CV is the original critical value. The default is reducecv = 0.14268.
urfinal Threshold value for the final unit root test. If the magnitude of an AR root for the final
model is less than this number, a unit root is assumed, the order of the AR polynomial
is reduced by one, and the appropriate order of differencing (nonseasonal, seasonal) is
increased. This value should be greater than one. The default is urfinal = 1.05.
DETAILS
The automdl spec cannot be used in the same spec file as the pickmdl or arima specs, or when the file
argument is specified in the estimate spec.
The automatic ARIMA model selection procedure implemented into Version 0.3 is based on the procedure
in the TRAMO time series modeling program developed by Victor Gómez and Agustin Maravall (Gómez and
Maravall 1996). It is very similar to TRAMO’s procedure but contains modifications to make use of X-13-
ARIMA-SEATS’ different model estimation procedure, regARIMA model options, transformation and outlier
identification procedures and model diagnostics. Also some additional tests have been added. Consequently,
the model selected can differ from the model TRAMO would select. Extensive testing has shown that the
models selected are usually at least as good as those selected by TRAMO (preliminary results in Hood 2002a).
7.2. AUTOMDL 67
The TRAMO procedure is largely documented in Gómez and Maravall (2001a), but the actual implemen-
tation of the procedure in the current TRAMO program differs somewhat from the description that appears in
the paper.
An overview of the ARIMA model selection procedure is given below, as given in Monsell (2002, 2006). The
procedure can be summarized in five stages:
• default model estimation: a default model is estimated, initial outlier identification and regressor tests
are performed, and residual diagnostics are generated;
• identification of differencing orders: empirical unit root tests are performed to determine the orders
of differencing needed for the model;
• identification of ARMA model orders: an iterative procedure is applied to determine the order of
ARMA parameters;
• comparison of identified model with default model: the identified model is compared to the default
model; and
• final model checks: where the final model is checked for adequacy.
Note that the second stage is optional, as the user can specify the orders of regular and seasonal differencing
using the diff argument.
Default model estimation
The first step of the automatic outlier procedure is to estimate a default model. For monthly and quarterly
series, this is initially an “airline” model: ARIMA (0 1 1)(0 1 1)s .
The default model is used to perform a number of tasks. If tests for trading day, Easter or user-defined
regressors are requested by the user in the regression spec, an initial check for the significance of these effects
is performed using the default model. The X-13ARIMA-SEATS program’s aictest option is used to check the
significance of the regressors using a small sample variant of AIC called AICC (otherwise known as the F-
adjusted Akaike’s Information Criterion, see Hurvich and Tsai 1989). For more details on how AIC tests for
regressors are implemented within X-13ARIMA-SEATS see the DETAILS section of the regression spec.
The procedure then checks the significance of including a constant term in the regARIMA model. A t-
statistic for the mean of the model residuals is generated, and is checked against a critical value of 1.96.
Once these tests are complete, the program performs automatic outlier identification (if specified by the user
in the outlier spec). Details concerning X-13ARIMA-SEATS program’s automatic outlier identification routine
can be found in Appendix B of Findley, Monsell, Bell, Otto, and Chen (1998), or in the DETAILS section of
the outlier spec.
After outlier identification, the trading day, Easter and constant regressors are checked to see if they are still
significant. This test is simpler: t-tests are generated, and a critical value of 1.96 is used to determine if the
regressors are significant (except for the constant regressor, which uses the same value specified in armalimit).
For the trading day regressor, at least one of the regressors needs to have a critical value greater than 1.96.
Note that this test is done for trading day and Easter regressors only if the aictest argument is given in the
regression spec; the constant regressor is always tested.
After the regression part of the default model is determined, the program generates residual diagnostics for
this model. These diagnostics are:
• the Ljung-Box Q statistic for the model residuals (at lag 24 if this is a monthly series, lag 16 for a quarterly
series),
• the confidence coefficient of this Ljung-Box Q statistic ,
• a t-value for the mean of the regARIMA model residuals, and
• an estimate of the residual standard error.
The confidence coefficient is defined to be 1 minus the p-value of the Ljung-Box Q statistic, as in Lehman
(1986). The TRAMO documentation (Gómez and Maravall 1996) refers to the confidence coefficient as the
significance level.
These diagnostics will be compared later to those of the model selected by the automatic model identification
procedure. The model identified by this procedure must show some improvement over the default model in these
residual diagnostics; otherwise, the program will accept the default model.
Just before the model identification phase begins, the program removes the regression effects estimated by
the default model from the original series. It is this series, rather than the original series, that is used in the
model identification routines.
In this way, an attempt is made to robustify the model identification process, to ensure that the choice of
differencing and model orders are not unduly affected by outliers, calendar effects, and other regression effects.
This regression residual series is referred to as the linearized series in the TRAMO documentation.
Identification of differencing orders
Now the program will attempt to identify an appropriate order of differencing for the “linearized” series com-
puted earlier. This is done by performing a series of unit root tests, fitting different ARMA models to the
(sometimes differenced) linearized series. The estimation of these models is done using a technique called the
Hannan-Rissanen method (see Hannan and Rissanen 1982, Gómez and Maravall 2001a). This method com-
putes the estimates of the ARMA parameters by setting up a linear regression using lagged values of the original
series (to estimate the AR parameters) and lagged estimate of the innovations generated recursively from the
autocovariances (to estimate the MA parameters). Biases in the MA parameters are corrected with a technique
provided by Chen (1985), and the MA parameter estimates are improved when AR parameters are present by
applying Chen’s method to the series filtered by the AR filter (see Gómez 1998).
Step 1 : The first stage of the procedure fits a (2 0 0)(1 0 0)s ARIMA model to the linear series using the
Hannan-Rissanen method, and examines the real AR roots of the estimated model. The program considers such
a root a unit root if the modulus of the root is less than 1.042, and the order of differencing that corresponds
to the root (seasonal or nonseasonal) is increased by one.
7.2. AUTOMDL 69
If the Hannan-Rissanen procedure estimates a model with roots inside the unit circle, X-13ARIMA-SEATS
re-estimates the model using exact maximum likelihood estimation and the modulus test described above is
applied to the resulting estimates.
Step 2 : If differencings were found in Step 1, the linearized series is differenced at the start of Step 2. An
ARMA (1 1)(1 1)s model is then fit to the resulting series, and the AR parameters are checked to see if they
are close to one. The criterion for “close to one” depends on whether the program is examining the regular or
seasonal AR coefficients.
If an AR coefficient is found that meets the criterion, the program checks to see if there is a common factor
in the corresponding AR and MA polynomials of the ARMA model that can be cancelled.
If there is no cancellation, the differencing order changes. The linearized series is differenced using this
new set of differencing orders. The ARMA model is fit again, and the program checks to see if any further
differencings can be found. This process repeats until no further differencings are found.
Once the differencing orders are determined, a t-statistic for the mean term of the fully differenced series is
generated based on either the sample mean (if no differencing is identified) or by adding a constant term to the
regARIMA model. The critical value of the test is set based on the number of observations in the series.
This is a simplified overview of the actual process. Other tests may be performed if no differencings are
found in Step 2, and the procedure has checks implemented to avoid going from no differencings after Step 1 to
both regular and seasonal differencing after the first stage of Step 2. For more details, see Gómez and Maravall
(2001a).
Identification of ARMA model orders
Once an appropriate set of differencing orders has been found, the program turns to the identification of
the orders of the ARMA model. The basic procedure involves comparing values of the Bayesian Information
Criterion (see Schwarz 1978) of a number of models, up to a maximum order for the regular and seasonal ARMA
polynomial which can be specified by the user. As with Akaike’s AIC criterion, the model with the lowest BIC
is preferred.
The formula below is the classical formula for BIC that is printed out in the X-13ARIMA-SEATS output.
BICN = −2Lc
N + np log N,
where LcN is the maximized value of the log likelihood evaluated over N observations, np is the number of
estimated parameters in the model, including the white noise variances, and N is the number of observations
remaining after application of the model’s differencing and seasonal differencing operations.
TRAMO uses a variant of this BIC formula in its automatic model identification procedure which divides
the log likelihood and the penalty term by N . In order to be able to use TRAMO’s final selection criteria,
it is necessary that X-13ARIMA-SEATS have a comparable variant of BIC. So X-13ARIMA-SEATS generates the
following BIC which is only used for the automatic modeling procedure:
BIC2N = (−2Lc
N + np log N )/N.
The identification procedure allows the user to specify the maximum order of regular AR and MA polynomial
(mr , can be as high as 3, default of 2) and seasonal AR and MA polynomial (ms , can be as high as 2, default of
1) up to which the program estimates ARIMA models and generates values of BIC2. A three stage procedure
is detailed in Gómez and Maravall (2000) that reduces the number of models estimated.
To get an initial estimate for the seasonal model orders, BIC2 is computed for all ARIMA models of the
form (3 d 0)(P D Q)s , where d and D are the previously determined or specified regular and seasonal orders
of differencing, respectively, and 0 ≤ P, Q ≤ ms . The program then chooses the pair of values P and Q that
minimize BIC2.
Using these values of P and Q, the program now tries to identify the best model orders for the nonseasonal
part of the ARIMA model. BIC2 is computed for all ARIMA (p d q)(P D Q)s models, where d and D are the
regular and seasonal orders of differencing, respectively, and 0 ≤ p, q ≤ mr . The pair of values p and q are
chosen that minimize BIC2.
Using these values of p and q, the selection of seasonal model orders is now refined. The program computes
BIC2 for all ARIMA (p d q)(P D Q)s models, where d and D are the regular and seasonal orders of differencing,
respectively, and 0 ≤ P, Q ≤ ms . The pair of values P and Q are chosen that minimize BIC2.
There is one exception for this third stage of the process: if no seasonal AR was found in the first stage of the
process, and a seasonal differencing is present, then program only computes BIC2 for ARIMA (p d q)(0 D Q)s
models, where d and D are the regular and seasonal orders of differencing, respectively, and 0 ≤ Q ≤ ms . The
values of Q is chosen that minimizes BIC2.
During the ARMA order selection process, X-13ARIMA-SEATS keeps track of the models with the five smallest
BIC2s. Once the identification phase is over, the program will compare the BIC2 for the best model with that
of the other 4 models to see if there are models with BIC2s that are “close” enough that there is no “significant”
difference between the models. The criteria for “close enough” depends on the length of the series, the magnitude
of the difference between the BICs, and other criteria.
If the program finds a model that is “close” enough to the best model, the program also checks to see whether
the model with the higher BIC is more parsimonious (especially in the seasonal operator) than the best model.
If so, the program will accept the more parsimonious model.
The program also checks for model balance. A model is said to be more balanced than a competing model
if the absolute difference between the total orders of the AR plus differencing and MA operators is smaller for
one model than another. While balanced models are useful for model-based seasonal adjustment, it is unclear
whether this criterion is useful for the types of operations X-13ARIMA-SEATS does, as it induces a small bias
toward mixed models and mixed ARMA models can be difficult to estimate due to near cancellation. Therefore,
X-13ARIMA-SEATS makes checking for model balance at this stage optional; the default is not to test for model
balance.
If the identified model is different from the default model, the program redoes many of the steps that
determined the regressors of the default model. Outlier regressors identified for the default model are removed
from the identified model. If the user has specified AIC testing of trading day, Easter, or user defined regressors,
this testing will be redone for the identified model. Then outlier identification is redone for the identified model.
7.2. AUTOMDL 71
Comparison of identified model with default model
At this point, if the identified model is not the default model, the residual diagnostics from the automatically
identified model are compared to those of the default model. Let QA be the confidence coefficient of the Ljung-
Box Q statistic for the automatically identified model (at lag 24 for monthly series, lag 16 for quarterly series),
QD the confidence coefficient of the Ljung-Box Q statistic for the default model, RSEA the residual standard
error for the automatically identified model, and RSED the residual standard error for the default model.
The default model will be preferred over the automatically identified model if
• the number of outliers automatically identified for the default model is less than or equal to the number
of automatically identified outliers for the automatically identified model, AND
• QA < 0.95 and QD < 0.75 and RSED < RSEA , OR
• QA > 0.95 and QD < 0.95 (only on the first pass), OR
• QA < 0.95 and QD < 0.75 and QD < QA and RSED < RSEA ∗ 1.013, OR
• QA ≥ 0.95 and QD < 0.95 and RSED < RSEA × 1.013, OR
• the automatic model is (1 0 1)(0 1 1)s or (1 0 0)(0 1 1)s and φ1 ≥ 0.82, OR
• the automatic model is (0 1 1)(1 0 1)s or (0 1 1)(1 0 0)s and Φ1 ≥ 0.65.
The program then tests to see if the preferred model is acceptable. The confidence coefficient of the Ljung-
Box Q statistic is used as the criterion. If this value is greater than 0.975 (by default), the program will decrease
the critical value of the automatic outlier identification based on the value of reducecv, given the formula below:
CVr = (1 − reducecv) × CV
where CV is the original critical value and CVr is the reduced outlier critical value. The reduced critical value
is not allowed to be smaller than 2.8.
The program will then attempt to redo the automatic modeling procedure and re-identify outliers with the
new critical value. The re-identification of outliers will be done without another automatic model identification
if no outliers were identified earlier.
Diagnostics are then generated for the revised model, and these diagnostics are compared to those of the
previous preferred model. The Ljung-Box Q test is performed again; this time, the test fails if the confidence
coefficient is greater than 0.990. If this does not result in a model with an acceptable Ljung-Box Q, the program
sets the model to be (3 d 1)(0 D 1)s and attempts to identify outliers for this model.
Finally, t-statistics for the trading day, Easter, and constant regressors are checked, as they were after
automatic model identification of the default model. Again, a critical value of 1.96 is used to determine if the
regressors are significant (except for the constant regressor, which uses the same value specified in armalimit).
For the trading day regressor, at least one of the regressors needs to have a critical value greater than 1.96.
Final model checks
Once a final model is selected, a final series of tests for model inadequacy is performed.
First, the model is checked for unit roots in the AR polynomial, to see if the order of regular or seasonal
differencing should be corrected. The program detects a unit AR root if the modulus of a given AR root is
less than or equal to 1.05. If a unit root is detected, the program then reduces the order of the appropriate
AR polynomial, and increases the appropriate order of differencing. The program then estimates the updated
model, and regenerates the model diagnostics.
Next, the model is checked for unit roots in the nonseasonal MA polynomial. Models with nonseasonal MA
unit roots have led to inadmissable decompositions in model based signal extraction procedures such as SEATS.
The program takes the sum of the non-seasonal MA coefficient estimates and checks to see if it is within 0.001
of one. If so, the order of regular differencing is reduced by one, and the order of the MA polynomial is reduced
by one. A constant term is added to the regARIMA model (if one is not already present), and the program
estimates the updated model and regenerate model diagnostics.
If a constant term is not present in the regARIMA model, the program now checks if the t-statistic for the
mean of the model residuals is significant (greater in magnitude than 2.5). If the t-statistic is significant, the
program adds a constant term to the set of regressors.
A test for insignificant ARMA parameters is then performed in an attempt to simplify the identified model.
T -statistics for the ARMA coefficients are generated. The highest order AR, MA, and seasonal AR and MA
coefficients are tested for significance, using the following criteria:
• to avoid model order reduction the t-statistic of the largest order AR, MA, seasonal AR and seasonal MA
coefficients has to be larger in magnitude than the value specified for armalimit, and
• the absolute value of the coefficient estimate itself must be greater than 0.15 (if there is at most 150
observations in the series) or 0.10 (if there is more than 150 observations).
If more than one insignificant coefficient is found for a given type of ARMA parameter (such as nonseasonal
AR, or seasonal MA coefficients) and outlier identification has been specified, the program will reduce the outlier
critical value using the value specified for reducecv. As noted before, the reduced outlier critical value cannot
be less than 2.8. The program will then try to re-identify the model, with a reduced outlier critical value.
If outlier identification was not specified, if the critical value is already 2.8, or if only one insignificant
coefficient is found, the program will reduce the order of the model by setting insignificant coefficients to zero
and estimate the reduced model.
Note that if there is only one ARMA parameter in the model, the program will not remove it, even if it is
insignificant. Also, no ARMA coefficients are eliminated from the model if a unit root is found (that is, if the
magnitude of one of the roots is less than 1.053).
7.2. AUTOMDL 73
EXAMPLES
Example 1 Use the automatic ARIMA modeling procedure to select a model and use it to extend the
series with one year of forecasts. Trading day and stable seasonal regression effects are to
be included in the models. A default seasonal adjustment is to be performed.

file="ussales.dat" }
regression { variables = (td seasonal) }
automdl { }
estimate { }
x11 { }
Example 2 Similar to Example 1, except that the difference orders are preset to a regular and seasonal
difference, and the maximum regular ARMA order to be examined will be 3.

regression { variables = td }
automdl { diff = ( 1 1 )
maxorder = ( 3, ) }
outlier { }
estimate { }
x11 { }
Example 3 The same as Example 1, except that the identified model will be saved in the log file, amd
the program will use AIC to check if trading day regressors are needed.

regression { aictest = td }
automdl { savelog = amd }
estimate { }
x11 { }
7.3 CHECK
DESCRIPTION
Specification to produce statistics for diagnostic checking of residuals from the estimated model. Statistics
available for diagnostic checking include the sample ACF and PACF of the residuals with associated standard
errors, Ljung-Box Q-statistics and their p-values, summary statistics of the residuals, normality test statistics
for the residuals, a spectral plot of the model residuals, and a histogram of the standardized residuals.
USAGE
check { maxlag = 36
print = (none +histogram +acf)
qtype = bp
save = (acf)
savelog = normalitytest
}
ARGUMENTS
maxlag The number of lags requested for the residual sample ACF and PACF for both tables
and plots. The default is 24 for monthly series, 8 for quarterly series.
print and save Table 7.3 gives the available output tables for this spec. The acf, acfplot, histogram,
and normalitytest tables are printed out by default. For a complete listing of the brief
and default print levels for this spec, see Appendix B.
qtype The type of residual diagnostic to be displayed with the sample autocorrelation plots. If
qtype = ljungbox or qtype = lb, the Ljung-Box Q-statistic will be the one produced. If
qtype = boxpierce or qtype = bp, the Box-Pierce Q-statistic will be the one produced.
The Ljung-Box statistic will be produced by default.
7.4.
DETAILS
The check spec uses residuals from the estimated model. If the estimate spec is absent, the check spec forces
estimation of the model (with default estimation options).
Under the null hypothesis that the model is correct, the Ljung-Box or Box-Pierce Q-statistics are asymptot-
ically distributed as χ2 with degrees of freedom equal to the number of lags used in computing them less the
number of AR and MA parameters estimated. The degrees of freedom are shown on the output. Ignore the
Q-statistics and p-values corresponding to zero degrees of freedom.
7.3. CHECK 75
Table 7.3: Available Output Tables for Check
name short save? description of table

acf acf + autocorrelation function of residuals with stan-
dard errors, and Ljung-Box Q-statistics com-
puted through each lag
acfplot acp · plot of residual autocorrelation function with ±
2 standard error limits
pacf pcf + partial autocorrelation function of residuals with
standard errors
pacfplot pcp · plot of residual partial autocorrelation function
with ± 2 standard error limits
acfsquared ac2 + autocorrelation function of squared residuals with
standard errors, and Ljung-Box Q-statistics com-
puted through each lag
acfsquaredplot ap2 · plot of squared residual autocorrelation function
with ± 2 standard error limits
normalitytest nrm · Geary’s a and Kurtosis statistical tests for the
normality of the model residuals, as well as a test
for skewness of the residuals.
histogram hst · histogram of standardized residuals and the fol-
lowing summary statistics of the residuals: min-
imum, maximum, median, standard deviation,
and robust estimate of residual standard devia-
tion (1.48 × the median absolute deviation)
Name gives the name of each table for use with the print and save arguments.
Save? indicates which tables can be saved (+) or not saved (·) into a separate file
with the save argument.
Another diagnostic included in the X-13ARIMA-SEATS software is a model-based F-statistic for determining
if there is stable seasonality in the original series.
This F-test is generated from the chi-square test of groups of regressors used to determine if a particular
group of regression parameters in the regARIMA model are collectively zero. The chi-square test statistic is
given below:
h i
χ̂2 = β̂ 0 V ar(β̂)−1 β̂. (7.1)
One such group of predefined regressors is the fixed seasonal regressors. This type of regressor can be
expressed in two ways - monthly (or quarterly) indicator variables or a trigonometric representation of a fixed
monthly pattern.
When these regression terms are included in the regARIMA model, the chi-square test of the seasonal
regressors is produced, and serves as an indication of the stable seasonality in the series.
Table 7.4: Available Log File Diagnostics for Check

normalitytest nrm Test results from the normality tests on the regARIMA
model residuals (Kurtosis, skewness and Geary’s a statis-
tics)
ljungboxq lbq Significant lags for the Ljung-Box Q statistic
boxpierceq bpq Significant lags for the Box-Pierce Q statistic
seasonalftest sft Model-based f-statistic for seasonality from Lytras, Feld-
pausch, and Bell (2007)
tdftest tft Model-based f-statistic for trading day
When these regression terms are included in the regARIMA model, the chi-square test of the seasonal
regressors is produced, and serves as an indication of the stable seasonality in the series.
The chi square test of the seasonal regressors can be corrected to account for the error in the estimation of
the innovation variance by using the test statistic χ̃2 /k, generated as follows:
χ̃2 χ̂2 n−d−k

= × (7.2)
k k n−d
where χ̂2 is the chi-squared statistic from (7.1), n is the number of observations in the series, d is the degree
of differencing, and k is the number of regressors for the group being tested in (7.1). The test statistic in (7.2)
follows an Fk,n−d−k distribution.
Lytras, Feldpausch, and Bell (2007) compared the performance of the χ̃2 /k statistic to several tests for stable
seasonality that are commonly used by seasonal adjustors but whose statistical properties are unknown. The
simulation studies examined led the authors to recommend the use of the χ̃2 /k statistic over more traditional
diagnostics.
X-13ARIMA-SEATS produces three statistics that test the regARIMA model residuals for deviations from
normality; one tests for skewness using the statistic:
√ Pn
n (Xi − X̄)3
c = Pn i=1
( i=1 (Xi − X̄)2 )1.5
The remaining test statistics both test for kurtosis and can be used when there is no skewness. One is
Geary’s a statistic, defined as in Geary (1936) and Gastwirth and Owens (1977):
1
Pn
| Xi − X̄ |
a = qn Pi=1
1 n 2
n i=1 (Xi − X̄)
7.3. CHECK 77
where X̄ is the sample mean. The other is the sample kurtosis:

Pn
n (Xi − X̄)4
b2 = Pn i=1
( i=1 (Xi − X̄)2 )2
Properties of both are discussed in Section 5.14 of Snedecor and Cochran (1980).
A significant value of one of these statistics indicates that the standardized residuals do not follow a standard
normal distribution. X-13ARIMA-SEATS tests for significance at the one percent level, from values given in tables
from Pearson (1938) and Pearson and Hartley (1954). If the regARIMA model fits the data well, such lack of
normality ordinarily causes no problems.
However, a significant value can occur because certain data effects are not captured well by the model.
Sometimes these effects can be captured by additional or different regressors (e.g. trading day, holiday or
outlier regressors). Thus, significant values can be used as a stimulus to reconsider what regressors to use.
X-13ARIMA-SEATS also produces a measure of the skewness of the residuals
There are other important effects that can cause a significant value, such as random variation of the coeffi-
cients or time-varying conditional variances, which cannot be represented by regARIMA models. These other
effects cause the test statistics and forecast coverage intervals of X-13ARIMA-SEATS to have reduced reliability.
Their presence is often indicated by significant values of the Ljung-Box Q-statistics of the squared residuals.
The number of lags for the ACF of the squared residual is set to be equal to seasonal period of the series
(12 for monthly series, 4 for quarterly series). This value cannot be changed by the maxlag argument.
The use of fixed coefficients in the ARIMA model can invalidate the DF (degrees of freedom) values and
therefore also the associated chi-square P-values in the Ljung-Box or Box-Pierce Q-statistic output of check.
This happens when the fixed values are actually estimated values from a previous model fitting. The P-values
will have the expected (approximate) validity when a statistically insignificant coefficient has been fixed at the
value zero.
EXAMPLES
Example 1 Print all available diagnostic checks of the residuals from the specified model. The sample
autocorrelation and partial autocorrelation function of the residuals is computed through
lag 36 (the default for monthly time series). The check spec forces model estimation to be
performed (with default options) even though the estimate spec is not present.
series { title = "Monthly Retail Sales"

start = 1964.jan
file = "sales1.dat" }
regression { variables = (td ao1967.jun
ls1971.jun easter[14]) }
arima { model = (0 1 1)(0 1 1) }
check { print = (all) }
Example 2 For the same series and model as in Example 1, produce all diagnostic checking statistics
except the except the printed table and plot of the residual PACF. The residual ACF is
computed through lag 36.

start = 1964.jan
file = "sales1.dat" }
regression { variables = (td ao1967.jun
ls1971.jun easter[14]) }
arima { model = (0 1 1)(0 1 1) }
check { print = (all -pacf -pacfplot)
maxlag = 36 }
7.4. COMPOSITE 79
7.4 COMPOSITE
DESCRIPTION
This spec is used as part of the procedure for obtaining both indirect and direct adjustments of a composite
series. For obtaining composite adjustments, it is one of the required spec files referenced in a metafile. Previous
spec files in the metafile must define the component series and how they are combined to form the composite
(see the comptype and compwt arguments of the series spec). This spec is used in place of the series spec.
The user can specify a title for the composite adjustment, a name for the composite series, which tables are
to be printed or stored, and which line-printer plots are to be produced from the indirect adjustment.
USAGE
composite { title = "Total one family housing starts"
name = "hs1ft"
decimals = 2
modelspan = (1985.Jan,)
appendfcst = yes
appendbcst = no
type = stock
print = (brief +indtest)
save = (indseasonal)
savelog = (indtest)
}
ARGUMENTS
appendbcst Determines if backcasts will be included in certain tables selected for storage with the
save option. If appendbcst=yes, then backcasted values will be stored with tables a16,
b1, d10, d16, and h1 of the x11 spec, table s10 of the seats spec, tables a6, a7, a8, a8.tc,
a9, and a10 of the regression spec, and tables c16 and c18 of the x11regression spec.
If appendbcst=no, no backcasts will be stored. The default is to not include backcasts.
appendfcst Determines if forecasts will be included in certain tables selected for storage with the save
option. If appendfcst=yes, then forecasted values will be stored with tables a16, b1,
d10, d16, and h1 of the x11 spec, tables a6, a7, a8, a8.tc, a9, and a10 of the regression
spec, and tables c16 and c18 of the x11regression spec. If appendfcst=no, no forecasts
will be stored. The default is to not include forecasts.
decimals Specifies the number of decimals that will appear in the seasonal adjustment tables of the
main output file. This value must be an integer between 0 and 5, inclusive (for example,
decimals=3). The default number of decimals is zero.
modelspan Specifies the span (data interval) of the composite time series that is to be used to
determine all regARIMA model coefficients. This argument can be utilized when, for
example, the user does not want data early in the series to affect the forecasts, or,
alternatively, data late in the series to affect regression estimates used for preadjustment
before seasonal adjustment. The modelspan argument has two values, the start and end
date of the desired span. A missing value defaults to the corresponding start or end date
of the composite series being analyzed. For example, for monthly data, the statement
modelspan=(1968.1, ) causes whatever regARIMA model is specified in other specs to
be estimated from the time series data starting in January, 1968 and ending at the end
date of the analysis span. A comma is necessary if either the start or end date is missing.
The start and end dates of the model span must both lie within the time span of the
composite series, and the start date must precede the end date.
Another end date specification, with the form 0.per, is available to set the ending date
of modelspan always to be the most recent occurrence of a specific calendar month
(quarter for quarterly data) in the span of data analyzed, where per denotes the calendar
month (quarter). If the span of data considered ends in a month other than December,
modelspan=(,0.dec) will cause the model parameters to stay fixed at the values obtained
from data ending in the next-to-final calendar year of the span.
name The name of the composite time series. The name must be enclosed in quotes and may
contain up to 8 characters. It will be printed as a label on every page of printed output.
print and save The default output tables available for the direct and indirect seasonal adjustments gen-
erated by this spec are given in Table 7.5; other output tables available are given in
Table 7.6. For a complete listing of the brief and default print levels for this spec, see
Appendix B.
Table 7.7 gives table names and abbreviations that can be used with the save argument
to save certain tables as percentages rather than ratios. Specifying these table names
in the print argument will not change the output of the program, and the percentages
are only produced when multiplicative or log-additive seasonal adjustment is specified by
the user in the mode argument of the x11 spec; these quantities will be expressed as
differences if mode = add.
savelog The diagnostics available for output to the log file (see section 2.6) are listed in Table
7.8.
title A title describing the composite time series. The title must be enclosed in quotes and
may contain up to 79 characters - longer text strings will be truncated to the first 79
characters. It will be printed above the data in the output.
type Indicates the type of series being aggregated. If type = flow, the composite series is
assumed to be a flow series; if type = stock, the composite series is assumed to be a
stock series. The default is to not assign a type to the series.
indoutlier If indoutlier=yes, the program will attempt to generate indirect point and level change
outliers from the components of the composite adjustment. This is the default for the
program.
7.4. COMPOSITE 81
Table 7.5: Default Output Tables for Composite Spec

adjcompositesrs b1 + aggregated time series data, prior adjusted, with associ-
ated dates
compositesrs cms + aggregated time series data, with associated dates
header hdr · header for indirect seasonal adjustment
indadjsatot iaa + final indirect seasonally adjusted series, with yearly totals
adjusted to match the original series
indadjustfac iaf + final combined adjustment factors for the indirect seasonal
adjustment
indcalendar ica + final calendar factors for the indirect seasonal adjustment
indftestd8 idf · final unmodified SI-ratios (differences) for the indirect ad-
justment
indirregular iir + final irregular component for the indirect adjustment
indmovseasrat ims · moving seasonality ratios for the indirect seasonal adjust-
ment
indqstat if3 · quality control statistics for the indirect seasonal adjust-
ment
indreplacsi id9 + final replacement values for extreme SI-ratios (differences)
for the indirect adjustment
indresidualseasf irf · F-test for residual seasonality
indrevsachanges i6a + percent changes for indirect seasonally adjusted series
with revised yearly totals
indrndsachanges i6r + percent changes (differences) in the indirect seasonally ad-
justed series
indsachanges ie6 + percent changes (differences) in the indirect seasonally ad-
justed series
indsadjround irn + percent changes for rounded indirect seasonally adjusted
series
indseasadj isa + final indirect seasonally adjusted series
indseasonal isf + final seasonal factors for the indirect seasonal adjustment
indseasonaldiff isd + final seasonal difference for the indirect seasonal adjust-
ment (only for pseudo-additive seasonal adjustment)
indtest itt · test for adequacy of composite adjustment
indtotaladjustment i18 + indirect total adjustment factors
indtrend itn + final trend-cycle for the indirect adjustment
indtrendchanges ie7 + percent changes (differences) in the indirect final trend
component
indunmodsi id8 + final unmodified SI-ratios (differences) for the indirect ad-
justment
indx11diag if2 · summary of seasonal adjustment diagnostics for the indi-
rect seasonal adjustment
origchanges ie5 + percent changes (differences) in the original series
Save? indicates which tables can be saved (+) or not saved (·) into a separate file with the save
argument.
Table 7.6: Other Output Tables for Composite Spec

adjcompositesrsplot b1p · plot of the prior adjusted aggregate series
calendaradjcomposite cac + aggregated time series data, adjusted for regARIMA
calendar effects.
compositeplot cmp · plot of the prior adjusted aggregate series
indaoutlier iao + final indirect AO outliers
indirregularplot iip · plot of the final irregular component from the indirect
seasonal adjustment
indlevelshift ils + final indirect LS outliers
indmcdmovavg if1 + MCD moving average of the final indirect seasonally
adjusted series
indmodoriginal ie1 + original series modified for extreme values from the in-
direct seasonal adjustment
indmodsadj ie2 + seasonally adjusted series modified for extreme values
from the indirect seasonal adjustment
indmodirr ie3 + irregular component modified for extreme values from
the indirect seasonal adjustment
indrobustsa iee + final indirect seasonally adjusted series modified for ex-
treme values
indseasadjplot iap · plot of the final indirect seasonally adjusted series
indseasonalplot isp · indirect seasonal factor plots, grouped by month or
quarter
indtotaladjustment ita + total indirect adjustment factors (only produced if the
original series contains values that are ≤ 0)
indtrendplot itp · plot of the final trend-cycle from the indirect seasonal
adjustment
indyrtotals ie4 · ratio of yearly totals of the original series and the indi-
rect seasonally adjusted series
origwindsaplot ie0 · plot of the aggregate series with the indirect seasonally
adjusted series
outlieradjcomposite oac + aggregated time series data, adjusted for outliers.
prioradjcomposite ia3 + composite series adjusted for user-defined prior adjust-
ments applied at the component level
ratioplotindsa ir2 · month-to-month (or quarter-to-quarter) ratio plots of
the original series
ratioplotorig ir1 · month-to-month (or quarter-to-quarter) ratio plots of
the indirect seasonally adjusted series
argument.
7.4. COMPOSITE 83
Table 7.7: Tables Saved As Percentages in the save Argument

origchangespct ip5 percent changes for composite series
indsachangespct ip6 percent changes for indirect seasonally adjusted series
indrevsachangespct ipf percent changes for indirect seasonally adjusted series with
forced yearly totals
indrndsachangespct ipr percent changes for rounded indirect seasonally adjusted
series
indtrendchangespct ip7 percent changes for indirect trend component
indcalendaradjchangespct ip8 percent changes in original series adjusted for calendar ef-
fects
indseasonalpct ips indirect seasonal component expressed as percentages if ap-
propriate
indirregularpct ipi indirect irregular component expressed as percentages if ap-
propriate
indadjustfacpct ipa indirect combined adjustment factors expressed as percent-
ages if appropriate
Name gives the name of each plot for use with the save arguments.
Short gives a short name for the tables of the save argument.
If indoutlier=no, no indirect outliers will be used in generating components of the

indirect seasonal adjustment.
saveprecision The number of decimals stored when saving a table to a separate file with the save
argument. The default value of saveprecision is 15. Example: saveprecision=10.
yr2000 If yr2000=yes, a “century cutoff” for 2-digit years from data stored in “X-11 formats”
is set at 1945. Years 00-45 are interpreted as 20xx, and years 46-99 are interpreted as
19xx. This is the default for the program.
If yr2000=no, the program assumes all 2-digit years fall in the 20th century and will
convert them to 4-digit years accordingly.
Note: this option is set here to affect program behavior when files are read in other
specs (such as the transform and x11regression spec).
DETAILS
An input specifications file with the composite spec can only be used in conjunction with spec files for com-
ponent series which together define a composite series. The names of these other spec files must be listed in a
metafile in which the name of this spec file appears last. The comptype argument of the series spec of each
component series controls how the components are combined to form the final aggregate (composite) series.
(See Section 2.5 for examples of how to run metafiles).
A composite adjustment run with this metafile produces an indirect seasonal adjustment of the composite
Table 7.8: Available Log File Diagnostics for Composite

indtest itt test for adequacy of composite adjustment
indm1 im1 M1 Quality Control Statistic from indirect adjustment
indm10 imt M10 Quality Control Statistic from indirect adjustment
indm11 ime M11 Quality Control Statistic from indirect adjustment
indq iq Overall index of the quality of the indirect seasonal ad-
justment
indq2 iq2 Indirect Q statistic computed without the M2 Quality
Control statistic
indmsr isr Moving seasonality ratio from indirect adjustment
indicratio iir ¯ C̄ ratio from indirect adjustment
I/
indfstabled8 id8 F-test for stable seasonality, performed on the final SI-
ratios from indirect adjustment
indmovingseasf isf F-test for moving seasonality from indirect adjustment
indidseasonal iid Identifiable seasonality test result for indirect adjustment
alldiagnostics all All seasonal adjustment diagnostics listed in this table
series as well as a direct seasonal adjustment. The indirect adjustment is the combination specified by the
comptype of the components, each adjusted or not adjusted according to the prescriptions of their spec files.
The direct adjustment is done as requested in the spec file of the composite spec. To control the output for the
direct seasonal adjustment, use the print and save arguments of the x11 spec.
To include an unadjusted series as a component of the indirect seasonal adjustment of the aggregate series,
specify the summary measures option by setting type = summary in the x11 spec of this component.
As is mentioned in section 2.7, the -c flag is used only to restrict a composite seasonal adjustment run
done with an input metafile (-m). In a composite run, X-13ARIMA-SEATS usually seasonally adjusts a set of
component series. When -c is invoked, the seasonal adjustment and modeling options specified in the input
spec files for the component series are ignored; the component series are only used to form the composite series.
This option is useful when identifying a regARIMA model for the composite series.
Although none of the tables of seasonal adjustment diagnostics produced in this spec can be saved to its own
file, specifying the diagnostic summary option with the -s flag at runtime allows the user to store information
7.4. COMPOSITE 85
from the composite analysis into a diagnostic summary file (with the file extension .udg). In addition, the
savelog argument can write selected diagnostics into the log file for a given run (with the file extension .log).
For more information, see section 2.6.
If a sliding spans analysis of the direct and indirect adjustments is desired, the sliding spans analysis option
must be specified for each of the component series. If the seasonal filter length is not the same for each
component, then the user must use the length argument of the slidingspans spec to ensure that the spans
stored for the component series are of the same length.
When a revisions history analysis of the seasonally adjusted series is specified for a composite seasonal
adjustment, the revisions of both the direct and indirect seasonal adjustments of the composite series are
produced. The revisions history analysis must be specified for each of the component series.
If a series is designated as a stock or a flow series by using the type argument, then trading day and Easter
regressors specified in regression spec need to agree with this type - one cannot specify stock trading day
regressors for a flow series. If a series type is not specified, then any trading day or holiday regressor may be
used with the series.
EXAMPLES
The following examples illustrating all the steps of a composite adjustment show complete spec files.
Step 1 A spec file must be created for each of the component series. In this example, we process
each of the components (Northeast, Midwest, South and West 1-family housing starts),
using a simple sum to form the composite. An example of the spec file for the Northeast
series (stored in cne1hs.spc), which is seasonally adjusted using 3x9 seasonal filters, is given
below:
series { title="NORTHEAST ONE-FAMILY Housing Starts"

file="cne1hs.ori" name="CNE1HS" format="2R"
comptype=add }
x11 { seasonalma=(s3x9)
title=(
"Component for Composite Adjustment"
"of Total U.S. 1-Family Housing Starts") }
The seasonal adjustment of CNE1HS produced by this spec file will be an addend in the
calculation of the indirect seasonal adjustment of the composite series.
A spec file for a component series that is not seasonally adjusted is given below:
series { title="West ONE-FAMILY Housing Starts"

file="cwt1hs.ori" name="CWT1HS" format="2R"
comptype=add }
x11 { type=summary }
This will cause the unadjusted series stored in cwt1hs.ori to be an addend in the calcula-
tion of the indirect seasonal adjustment of the composite series.
Step 2 Create a spec file for the indirect adjustment of total one-family housing starts, the sums of
four regional series. The direct seasonal adjustment of the series will be multiplicative and
will use a 3x9 seasonal moving average. Both the seasonal factors from the direct adjustment
and the implied factors from the indirect adjustment will be saved. The spec file (stored in
c1fths.spc) appears below:
composite { title="TOTAL ONE-FAMILY Housing Starts"

name="C1FTHS" save=(indseasonal) }
title="Composite adj. of 1-Family housing starts"
save=(D10) }
Step 3 Create a metafile for the input specification files of the component and composite series.
This metafile, stored in hs1ftot.mta appears below:
cne1hs
cmw1hs
cso1hs
cwt1hs
c1fths
Note that the spec file for the composite series is listed last.
Step 4 To run X-13ARIMA-SEATS for this example, enter the following:
x13as -m hs1ftot
and press the return (enter) key.

7.5. ESTIMATE 87
7.5 ESTIMATE
DESCRIPTION
Estimates the regARIMA model specified by the regression and arima specs. Allows the setting of various
estimation options. Estimation output includes point estimates and standard errors for all estimated AR,
MA, and regression parameters; the maximum likelihood estimate of the variance σ 2 ; t-statistics for individual
regression parameters; χ2 -statistics for assessing the joint significance of the parameters associated with certain
regression effects (if included in the model); and likelihood based model selection statistics (if the exact likelihood
function is used). The regression effects for which χ2 -statistics are produced include stable seasonal effects,
trading-day effects, and the set of user-defined regression effects.
USAGE
estimate { tol = 1.0e-5

maxiter = (500)
exact = arma
outofsample = yes
print = (none +model +estimates +lkstats)
save = (model)
savelog = (aic bic) }
ARGUMENTS
exact Specifies use of exact or conditional likelihood for estimation, likelihood evaluation, and
forecasting. The default is exact = arma, which uses the likelihood function that is exact
for both AR and MA parameters. Other options are: exact = ma, use the likelihood
function that is exact for MA, but conditional for AR parameters; and exact = none,
use the likelihood function that is conditional for both AR and MA parameters.
maxiter The maximum number allowed of ARMA iterations (nonlinear iterations for estimating
the AR and MA parameters). For models with regression variables, this limit applies
to the total number of ARMA iterations over all IGLS iterations. For models without
regression variables, this is the maximum number of iterations allowed for the single set
of ARMA iterations. The default is maxiter = 1500.
outofsample Determines the kind of forecast error used in calculating the average magnitude of forecast
errors over the last three years, a diagnostic statistic. If outofsample=yes, out-of-sample
forecasts errors are used; these are obtained by removing the data in the forecast period
from the data set used to estimate the model and produce one year of forecasts (for each
of the last three years of data). If outofsample=no, within-sample forecasts errors are
used. That is, the model parameter estimates for the full series are used to generate
forecasts for each of the last three years of data. The default is outofsample=no.
print and save Table 7.9 gives the default output tables for this spec; the other output tables are given
in Table 7.10. For a complete listing of the brief and default print levels for this spec,
see Appendix B.
7.11.
tol Convergence tolerance for the nonlinear estimation. Absolute changes in the log-likeli-
hood are compared to tol to check convergence of the estimation iterations. For models
with regression variables, tol is used to check convergence of the IGLS iterations (where
the regression parameters are re-estimated for each new set of AR and MA parameters),
see Otto, Bell, and Burman (1987). For models without regression variables there are
no IGLS iterations, and tol is then used to check convergence of the nonlinear iterations
used to estimate the AR and MA parameters. The default value is tol = 1.0e-5.
Table 7.9: Default Output Tables for Estimate

options opt · header for the estimation options
model mdl + if used with the print argument, this controls
printing of a short description of the model; if used
with the save argument, this creates a file con-
taining regression and arima specs correspond-
ing to the model, with the estimation results used
to specify initial values for the ARMA parameters
estimates est + regression and ARMA parameter estimates, with
standard errors
averagefcsterr afc · average magnitude of forecast errors over each of
the last three years of data.
lkstats lks + log-likelihood at final parameter estimates and,
if exact = arma is used (default option), corre-
sponding model selection criteria (AIC, AICC,
Hannan-Quinn, BIC)
file Name of the file containing the model settings of a previous X-13ARIMA-SEATS run. Such
a file is produced by setting save=model or save=mdl in this spec. The filename must be
enclosed in quotes. If the file is not in the current directory, the path must also be given.
If the file argument is used, the model, ma, and ar arguments of the arima spec and
7.5. ESTIMATE 89
Table 7.10: Other Output Tables for Estimate

iterations itr + detailed output for estimation iterations, in-
cluding log-likelihood values and parameters,
and counts of function evaluations and itera-
tions
iterationerrors ite · error messages for estimation iterations, in-
cluding failure to converge
regcmatrix rcm + correlation matrix of regression parameter es-
timates if used with the print argument; co-
variance matrix of same if used with the save
argument
armacmatrix acm + correlation matrix of ARMA parameter esti-
mates if used with the print argument; co-
variance matrix of same if used with the save
argument
lformulas lkf · formulas for computing the log-likelihood and
model selection criteria
roots rts + roots of the autoregressive and moving aver-
age operators in the estimated model
regressioneffects ref + Xβ̂, matrix of regression variables multiplied
by the vector of estimated regression coeffi-
cients
regressionresiduals rrs + residuals from regression effects
residuals rsd + model residuals with associated dates or ob-
servation numbers
Table 7.11: Available Log File Diagnostics for Estimate

aic aic Akaike’s Information Criterion (AIC)
aicc acc Akaike’s Information Criterion (AIC) adjusted for the
length of the series
bic bic Baysean Information Criterion (BIC)
hannanquinn hq Hannan-Quinn Information Criterion
averagefcsterr afc Average forecast error over the last three years of data
the variables, user, and b arguments of the regression spec cannot be used, as well
as the pickmdl and automdl specs.
fix Specifies whether certain coefficients found in the model file specified in the file argument
are to be held fixed instead of being used as initializing values for further estimation. If
fix = all, both the regression and ARMA parameter estimates will be held fixed at
their values in the model file. If fix = arma, only ARMA parameter estimates will be
held fixed at their model file values. If fix = none, none of the parameter estimates
will be held fixed. The default is fix = nochange, which will preserve coefficient values
specified as fixed in the model file and allow re-estimation of all other coefficients.
DETAILS
The inference results provided by X-13ARIMA-SEATS are asymptotically valid (approximately correct for suffi-
ciently long time series) under “standard” assumptions—see Section 4.5. The likelihood based model selection
statistics are provided only if the exact likelihood function is used. See Section 5.5 for comments on the use of
model selection statistics.
If the estimation iterations converge, X-13ARIMA-SEATS prints a message to this effect, and then displays the
estimation results. If the iterations fail to converge, X-13ARIMA-SEATS prints a message indicating this and then
displays the parameter values at the last iteration. These values should not be used as parameter estimates.
Instead, the program should be rerun, possibly starting at the parameter values obtained when the iterations
terminated. Potential causes of convergence problems and suggested remedies are discussed in Section 5.
Tol should not be set either “too large” or “too small”. Setting tol too large can result in estimates too
far from the true MLEs, while setting tol too small can result in an unnecessarily large number of iterations or
lead to a false impression of the precision of the results. What is too large or too small a value for tol depends
on the problem; the default value of tol= 10−5 is offered as a reasonable compromise. Setting tol to a number
less than machine precision for a double precision number (approximately 10−14 for PCs and Sun4 computers)
results in an error, but values for tol that even begin to approach machine precision are certainly too small.
For models with regression variables, a second convergence tolerance is needed to determine convergence of
the ARMA iterations within each IGLS iteration. This tolerance is set by the program to 100 × tol for the first
two IGLS iterations, after which it is reset to tol. (Since relatively large changes can be made to the regression
parameters in the initial IGLS iterations, it is not worth determining the ARMA parameters within tol at the
start.) Thus, when tol takes on its default value of 10−5 , the ARMA convergence tolerance is 10−3 for the first
two IGLS iterations, and thereafter it is 10−5 (= tol). Also, for models with regression variables, a limit is
needed for the maximum number of ARMA iterations allowed within each IGLS iteration. This limit is set to
40.
If the ARMA iterations fail to converge on a particular IGLS iteration, this is generally not a problem. The
program will continue with the next IGLS iteration, and its ARMA iterations may very well converge. In fact,
all that is necessary for overall convergence is that the ARMA iterations of the last IGLS iteration converge,
and that the IGLS iterations themselves converge to the tolerance tol within maxiter total ARMA iterations.
Setting print=roots produces a table of roots of all the AR and MA operators of the estimated model. In
addition to the roots, the table provides the modulus (magnitude) and frequency (on [−.5, .5]) of each root.
7.5. ESTIMATE 91
Table 7.12: Example of ARMA Roots Output
Roots of ARIMA Model

Root Real Imaginary Modulus Frequency
-----------------------------------------------------------
Nonseasonal AR
Root 1 -0.6784 0.8817 1.1125 0.3544
Root 2 -0.6784 -0.8817 1.1125 -0.3544
Nonseasonal MA
Root 1 -7.4107 0.0000 7.4107 0.5000
Seasonal MA
Root 1 1.5583 0.0000 1.5583 0.0000
Roots with modulus greater than one lie outside the unit circle, corresponding to stationary AR or invertible
MA operators. (See Section 5.4.) AR roots on or inside the unit circle (modulus ≤ 1) should occur only when
the likelihood function is defined conditionally for AR parameters (exact = ma or exact = none). MA roots
inside the unit circle (modulus < 1) will never occur, since invertibility is enforced in the estimation. MA roots
on the unit circle (modulus = 1) can be estimated within round-off error, or can occur in an MA operator all
of whose parameters are specified as fixed during estimation.
In sample output shown in Table 7.12, the nonseasonal AR(2) polynomial has a pair of complex conjugate
1/2
roots (zeros), z = x ± iy, with x = −.6784 and y = .8817, whose modulus (magnitude) is r = x2 + y 2 =
1.1125. Because this number is close to unity (1.000), it is worthwhile to examine the nonnegative frequency of
the root, i.e. the number λ ≥ 0 such that z = re±i2πλ to determine if the series may contain a deterministic
periodic component. The reasoning behind this is as follows. Whenever a modelled time series has a periodic
component f (t) with period 1/λ, i.e. f (t + 1/λ) = f (t), then an estimated AR polynomial of sufficiently high
order is likely to have a root near e±i2πλ (unless the differencing operators have e±i2πλ as a root). There are
theoretical results that help to explain why this happens, but a heuristic explanation is that for the simplest
functions with this period,
f (t) = A cos (2πλt + c) ,
−i2πλ
B 1 − ei2πλ B , whose roots are e±i2πλ , has the property that

the AR(2) polynomial φ (B) = 1 − e
φ (B) f (t) = 0.
Thus this AR(2) factor can perfectly predict f (t) from f (t − 1) and f (t − 2). Fitting a model with an AR
operator of order 2 or higher will tend to make the AR parameters take on values so that φ (B) f (t) = 0. (An
AR(1) polynomial suffices when ei2πλ is real, i.e. when λ = 0, 1/2.) . Hence the occurrence of an AR root with
.
modulus r = 1 suggests the presence of an approximately periodic component in the time series.
For monthly series, the frequencies of seasonal effects are λ = 1/12 , 2/12 , 3/12 , . . . , 6/12 (equivalent to
0.0833, 0.1666, 0.2500, . . . , 0.5000, respectively). The frequency λ = 0 is associated with trend movements, and
the frequency λ = 0.3482 with trading day effects. Note that the frequency 0.3544 of the nonseasonal AR roots
in the table above is very close to the trading day frequency. In fact, the time series whose model produced the
table has a strong trading day component, and the automatic modeling procedure added the AR(2) factor to
account for it, because there were no trading day regressors in the regression spec.
In the MA polynomials, near unit roots with seasonal or trend frequencies usually indicate that the MA
polynomials have one or more roots in common with the differencing or seasonal differencing polynomials.
The presence of such a common factor κ (B) indicates that the time series has deterministic trend or seasonal
components. More specifically, in the notation of equation (4.3) of Section 4.1 (but ignoring regressors), it
means that there is a function f (t) satisfying κ (B) f (t) = 0 such that the time series yt can be modelled as
( )
d s
θ (B) Θ (B s )

s (1 − B) (1 − B )
φ (B) Φ (B ) D
yt = f (t) + at .
κ (B) κ (B)

In the example table above, the model’s seasonal moving average polynomial is Θ B 12 = 1 − ΘB 12 with
Θ = 0.6417 so the root is 1/Θ = 1.5583 (the root of Θ (z) = 1 − Θz). Experience suggests that 1/Θ generally
needs to be 1.10 or less before it might be appropriate to replace the model with one having only fixed seasonal
effects (i.e. a model with D = 0 and with. variables=seasonals in the regression spec).
If the nonseasonal MA polynomial has a root close to the number 1 (i.e. modulus near 1, frequency near 0), it
often means that there is overdifferencing. That is, one should consider an alternative model with differencing
order d and nonseasonal MA order q both smaller by one, and a trend constant (i.e. f (t) = C above with
variables=const in the regression spec) should be included in the alternative model if it has a significant
t-statistic.
The use of fixed coefficients in the ARIMA model or the regression model specified in either the regression
or x11regression specs can invalidate the AIC, AICC, Hannan Quinn, and BIC model selection statistics in
the output. This happens when the fixed values are actually estimated values from a previous model fitting.
However, the P-values will have the expected (approximate) validity when a statistically insignificant coefficient
has been fixed at the value zero.
EXAMPLES
Example 1 Estimate by generalized least squares the regression coefficients in the model (1 − B)(yt −
P11
i=1 βi Mit ) = (1 − θB)at , where the Mit are regression variables for monthly fixed seasonal
effects. The MA parameter θ is held fixed at the value 0.25. Model residuals are saved in
a file in the current directory with the same name as the spec file, but with the extension
.rsd.
series { title = "Monthly Sales" start = 1976.1

data = (138 128 ... 297) }
regression { variables = seasonal }
arima { model = (0,1,1) ma = (0.25f) }
estimate { save = residuals }
Example 2 Estimate the seasonal model (1 − φB)(1 − B)(1 − B 12 )zt = (1 − ΘB 12 )at , with tol set to
10−4 , a looser convergence criterion than the default, and decrease the maximum number of
iterations allowed to 100. Since there are no regression parameters in the model, both tol
7.5. ESTIMATE 93
and maxiter apply to the single set of nonlinear ARMA iterations used to estimate φ and
Θ. The likelihood function used in parameter estimation is exact for MA and conditional
for AR parameters. The print argument specifies that the likelihood and parameter values
are printed for each iteration and, following the last iteration, the roots of the estimated
AR and MA operators are printed. The save argument will save the final regARIMA model
into a file.
series { title = "Monthly Inventory" start = 1978.12

data = (1209 834 ... 1002) }
regression { variables = (td ao1999.01) }
arima { model = (1,1,0)(0,1,1) }
estimate { tol = 1e-4 maxiter = 100 exact = ma save = mdl
print = (iterations roots) }
Example 3 Same as Example 2, except the regARIMA model estimates saved in example 2 are used in
this run via the file argument. All parameter estimates are fixed to the values stored in the
model file.

data = (1209 834 ... 1002) }
estimate { file = "Inven.mdl"
fix = all }
Example 4 Same as Example 3, except that three additional data values are available and we wish to
have the program determine if any of them are outliers. The ending date of the data span
in Examples 2 and 3 is December, 1999. The regARIMA model parameters are to be kept
fixed at the values obtained from this data span, which were stored by Example 2.

data = (1209 834 ... 1002 1425 901 1375) }
estimate { file = "Inven.mdl"
fix = all }
outlier { span=(2000.01,) }
7.6 FORCE
DESCRIPTION
An optional spec for invoking options that allow users to force yearly totals of the seasonally adjusted series
to equal those of the original series for convenience. Two forcing methods are available, the original modified
Denton method of X-11-ARIMA and earlier version of X-13ARIMA-SEATS described in Huot (1975) and Cholette
(1978), and a newer method based on the regression benchmarking method of Cholette and Dagum (1994) as
adapted by Quenneville, Cholette, Huot, Chiu, and Fonzo (2004). See also Dagum and Cholette (2006).
USAGE
force{ lambda = 0.0
mode = ratio
rho = 0.85
round = no
start = oct
target = calendaradj
type = regress
usefcst = no
print = (none saa)
save = saa
}
ARGUMENTS
lambda Value of the parameter λ used to determine the weight matrix C for the regression method
of forcing the totals of the seasonally adjusted series. For more details, see Section 2 of
Quenneville et al. (2004). Permissable values of lambda range from −3.0 to 3.0. The
most commonly used values are 1.0, 0.5 and 0.0, while cases could also be made for
using either −2, −1, or 2; other values of lambda are extremely unlikely. The default is
lambda = 0.0.
mode Determines whether the ratios (mode=ratio) or differences (mode=diff) in the annual
totals of the series specified in the argument target and the seasonally adjusted series
are stored, and on what basis the forcing adjustment factors are generated. The default
is mode=ratio.
print and save Table 7.13 gives the available output tables for this spec. All these tables are included
in the default printout. For a complete listing of the brief and default print levels for
this spec, see Appendix B.
to save certain tables as percentages rather than ratios. Specifying these table names
in the print argument will not change the output of the program, and the percentages
are only produced when multiplicative or log-additive seasonal adjustment is specified by
7.6. FORCE 95
Table 7.13: Default Output Tables for Force spec
Name Small Save Description of table

seasadjtot saa + final seasonally adjusted series with constrained yearly totals (if
type = regress or type = denton)
saround rnd + rounded final seasonally adjusted series (if round = yes) or the
rounded final seasonally adjusted series with constrained yearly
totals (if type = regress or type = denton)
revsachanges e6a + percent changes (differences) in seasonally adjusted series with
revised yearly totals
rndsachanges e6r + percent changes (differences) in rounded seasonally adjusted se-
ries
Save indicates which tables can be saved (+) or not saved (·) into a separate file with the save
argument.
the user in the mode argument of the x11 spec; these quantities will be expressed as
differences if mode = add.

revsachangespct p6a percent changes in seasonally adjusted series with forced yearly totals
rndsachangespct p6r percent changes in rounded seasonally adjusted series
rho Value of the AR(1) parameter (ρ) used in the regression method of forcing the totals of
the seasonally adjusted series. Admissable values of ρ must be between 0 and 1, inclusive.
If ρ = 1, the modified Denton method is used. The default for this argument is 0.9 for
monthly series, 0.729 for quarterly series ((0.9)3 ). For more details, see Section 2 of
Quenneville et al. (2004).
round When round=yes, the program will adjust the seasonally adjusted values for each calen-
dar year so that the sum of the rounded seasonally adjusted series for any year will equal
the rounded annual total; otherwise, the seasonally adjusted values will not be rounded.
start This option sets the beginning of the yearly benchmark period over which the seasonally
adjusted series will be forced to sum to the total. Unless start is used, the year is
assumed to be the calendar year for the procedure invoked by setting type=denton or
type=regress, but an alternate starting period can be specified for the year (such as the
start of a fiscal year) by assigning to forcestart the month (either the full name of the
month or the abbreviations shown in Section 3.3) or quarter (q1 for the first quarter, q2
for the second quarter, etc.) of the beginning of the desired yearly benchmarking period.
For example, to specify a fiscal year which starts in October and ends in September, set
start=october or start=oct. To specify a fiscal year which starts in the third quarter
of one year and ends in the second quarter of the next, set start=q3.
target Specifies which series is used as the target for forcing the totals of the seasonally adjusted
series. The choices for this argument are given in Table 7.15.
Table 7.15: Choices for the target Argument
Entry for target argument series

original Original Series
caladjust Calendar Adjusted Series
permprioradj Original Series adjusted for permanent prior adjustment factors
both Original Series adjusted for calendar and permanent prior adjustment
factors
The default for this argument is target=original.

type Specifies options that allow the seasonally adjusted series be modified to force the yearly
totals of the seasonally adjusted series and the series specified in the target argument
to be the same. By default (type=none), the program will not modify the seasonally
adjusted values.
When type=denton, the differences between the annual totals is distributed over the
seasonally adjusted values in a way that approximately preserves the month-to-month (or
quarter-to-quarter) movements of the original series for an additive seasonal adjustment,
and tries to keep the ratio of the forced and unforced values constant for multiplicative
adjustments. For more details see Huot (1975) and Cholette (1978).
When type=regress, a regression-based solution of Cholette and Dagum (1994) to the
problem of benchmarking seasonally adjusted series is used. For more details see Quen-
neville et al. (2004).
These forcing procedures are not recommended if the seasonal pattern is changing or if
trading day adjustment is performed; see DETAILS.
usefcst Determines if forecasts are appended to the series processed by the benchmarking routines
used to force the yearly totals of the seasonally adjusted series. If usefcst = yes, then
forecasts are used to extend the series in the forcing procedure; if usefcst = no, then
forecasts are not used. The default is usefcst = yes.
indforce Determines how the indirect seasonally adjusted series with forced yearly total is gener-
ated. If indforce = yes, the indirect seasonally adjusted series will be modified so that
their yearly totals match those of the target series. If indforce = no, the seasonally
adjusted series with forced yearly totals will be combined for each of the component series
7.6. FORCE 97
to form the indirect seasonally adjusted series with forced yearly totals. The default for
this option is indforce = yes.
DETAILS
0
Let X = (X1 , . . . , XT ) denote the vector of values Xt whose N annual totals within the time span 1 ≤ t ≤ T
0
define the constraints, and let A = (A1 , . . . , AT ) denote the vector of adjusted values At which are to be
modified to have the same annual totals as the Xt . (For example, the Xt could be trading day adjusted values
of an observed time series.) For the method type = regress, with specified values of lambda (λ) and rho (ρ),
0
the vector of forced values Ã = Ã1 , . . . , ÃT satisfying the constraints is given by
−1
Ã = A + CP CJ 0 (JCP CJ 0 ) (JX − JA) , (7.3)
where
C is a T × T diagonal matrix whose diagonal is proportional to
λ λ
|A1 | , . . . , |AT | , where P = ρ|i−j| 1≤i,j≤T , and where J is an N × T matrix of zeros and ones such that

JX and J Ã are the vectors of annual totals defining the forcing constraint,
J Ã = JX, (7.4)
see Quenneville et al. (2004), where it is shown that the right hand side of (7.3) minimizes
0
Ã − A C −1 P −1 C −1 Ã − A
subject to (7.4). As this reference further explains, formulas for Denton’s method are obtained from (7.3) by
letting ρ → 11 . When λ = 0, i.e. when C is the identity matrix with diagonal entries equal to one, this yields a
vector Ã whose entries minimize
X T n o2
fadd Ã = Ãt − At − Ãt−1 − At−1
t=2
T n
X o2
= Ãt − Ãt−1 − (At − At−1 ) ,
t=2

subject to (7.4). fadd Ã is the objective function of Denton’s additive method associated with type = denton
and mode=add. The first expression on the right shows that this method attempts to keep the changes Ãt − At
due to forcing constant over time, whereas the second offers the more appealing interpretation that the method
attempts to have the forced values Ãt preserve the changes in the series At from one observation time to the
next.
1 When ρ = 1, (7.3) cannot be used because P becomes singular and another equation given in Cholette (1984) is used. This
equation involves the inversion of a (T +N )×(T +N ) matrix whereas (7.3) involves the inversion of an N ×N matrix. Consequently,
users might observe an increase in computing time when using ρ = 1. An alternative to using ρ = 1 is to use ρ = 0.9999.
Similarly in part, when the diagonal entries of C coincide with the entries of A, corresponding to the case
λ = 1 when At ≥ 0, 1 ≤ t ≤ T , then letting ρ → 1 in (7.3) yields a vector Ã whose entries minimize
T
!2
X Ãt Ãt−1
fratio Ã = − (7.5)
t=2
At At−1
T
!2 !2
X Ãt−1 Ãt − Ãt−1 At − At−1
= − ,
t=2
At Ãt−1 At−1

subject to (7.4). fratio Ã is the objective function of Denton’s proportional method associated with type
= Denton and mode=ratio. The first expression on the right shows that this method attempts to keep the
ratios of forced to unforced values constant. However, the final expression shows that this method is not one
which attempts to have the forced values Ãt preserve the percent changes 100 (At − At−1 ) /At−1 in the series
At from one observation time to the next in any easily understood sense2 . (These percent changes are often
the most important product of multiplicative seasonal adjustment.) For all times t after the last complete year,
minimization of (7.5) subject to (7.4) yields a ”carry forward” factor c such that Ãt = cAt . As Quenneville
et al. (2004) discuss, this can lead to large revisions in the Ãt at the end when another full year of data Xt
becomes available to provide an additional forcing constraint. The recommended solution is, with λ = 1, to
choose a value of ρ somewhat less than one. This causes the ratios Ãt /At in an incomplete year to decay
effectively geometrically in ρ as t advances beyond the year of the last forcing constraint. This can lead to
Ãt similar to those obtained from Denton’s proportional method within years that have constraints but with
smaller revisions to forced values in incomplete years as additional data become available. Hood (2005) presents
comparison results for the regression method with various choices of λ and ρ and for other forcing methods.
Forcing causes
Xt + Xt+1 + · · · Xt+11 = At + At+1 + · · · At+11 (7.6)
to hold when month t is the first month of a calendar year or specified fiscal year (for which data through Xt+11
are available). The rationale usually given for forcing is the naive idea that ”seasonal adjustment redistributes
the seasonal effects throughout the year”. To indicate the problematic character of this rationale for forcing,
consider the situation in which different seasonal adjusters of a series can have any of the twelve calendar months
as the starting months of their fiscal years, with the result that (7.6) is implicitly assumed to hold for all months
t. (The widely used seasonal adjustment methods, including model-based method that are mean square optimal
if the model is correct, do not specially treat values of t associated with the beginning of the year.) We show
that this assumption can hold for additive or multiplicative seasonal decompositions of the series Xt if and only
if the series has an additive seasonal decomposition
Xt = St + At (7.7)
with perfectly repetitive seasonal effects, e.g. St = St+12 for monthly data. Indeed, in this situation annual
sums of the seasonal effects are constant,
St + St+1 + · · · St+11 = St+1 + · · · St+11 + St+12 = · · · (7.8)

P 2
2 For the latter, the objective function fCT Ã = T
t=2 Ãt /Ãt−1 − At /At−1 of Causey and Trager (1982) would be required.
This function is not a quadratic in the Ãt , so nonlinear methods are required for its minimization.
7.6. FORCE 99
and additive seasonal adjustment procedures produce values of St for which these sums are zero,
St + St+1 + · · · St+11 = 0, (7.9)
for all t (because a nonzero constant component belongs to the level component of the series included in At ).
For additive decompositions (7.7), (7.9) is equivalent to (7.6). Conversely, when (7.9) holds for all t, it is clear
from (7.8) that there is perfect repetition of the seasonal effects, St = St+12 .
Now we show that if (7.6) holds for all t for a multiplicative decomposition
Xt = St∗ At , (7.10)
then Xt has an additive decomposition with perfectly repetitive seasonal effects and therefore is a series for which
the multiplicative decomposition obscures the simplicity of the seasonality. Indeed, if we define St = (St∗ − 1) At ,
then (7.10) can be reexpressed as (7.7), so (7.6) implies that St + St+1 + · · · St+11 = 0 from which St = St+12
follows as before.
Forcing does not produce a perfectly stable seasonal pattern because only calendar or fiscal year totals are
forced, not all twelve month sums. But the preceding discussion shows that the situation of perfectly stable
additive seasonal patterns is the one in which equality of annual totals of adjusted and unadjusted is naturally to
be expected. In particular, there is no conceptual justification for forcing when the seasonal pattern is evolving
from one year to the next. (And since trading day patterns always change from one year to the next, there is no
conceptual justification for forcing to unadjusted annual totals of series that have trading day effects.) There
are practical justifications for forcing in certain contexts such as the complex production situation of national
accounts.
Forcing the seasonally adjusted totals to be the same as the original series annual totals can degrade the
quality of the seasonal adjustment, especially when the seasonal pattern is undergoing change. It is not natural
if trading day adjustment is performed because the aggregate trading day effect over a year is variable and
moderately different from zero.
EXAMPLES
Example 1 A multiplicative monthly seasonal adjustment is to be performed with 3×9 seasonal moving
averages for all months using ARIMA forecast extension of length 12 months, if one of the
default model types is accepted. The fiscal yearly totals for the seasonally adjusted series
will be forced to equal the totals of the original series for a fiscal year starting in October.
SERIES { TITLE="EXPORTS OF TRUCK PARTS" START =1967.1

FILE = "X21109.ORI" }
PICKMDL { }
X11 { SEASONALMA = S3X9 }
FORCE { START = OCTOBER }
Example 2 The same as Example 1, except that the regression-based solution of Cholette and Dagum
(1994) as adapted by Quenneville et al. (2004) is used.
SERIES { TITLE="EXPORTS OF TRUCK PARTS" START =1967.1

FILE = "X21109.ORI" }
PICKMDL { }
FORCE { START = OCTOBER
TYPE = REGRESS
RHO = 0.8
}
Example 3 Revise the seasonally adjusted series so that the sum of the rounded seasonally adjusted
series for any year will equal the rounded annual total; perform no other forcing of the
series.
Series { Title="Imports Of Truck Engines" Start =1967.1

File = "I21110.Ori" }
Pickmdl { }
X11 { Seasonalma = S3X5 }
Force { Type = None
Round = Yes
Save = Rnd
}
7.7. FORECAST 101
7.7 FORECAST
DESCRIPTION
Specification to forecast and/or backcast the time series given in the series spec using the estimated model.
The output contains point forecasts and forecast standard errors for the transformed series, and point forecasts
and prediction intervals for the original series.
USAGE
forecast { maxlead = 12
maxback =12
probability = 0.95
exclude = 10
lognormal = yes
print = (none +transformedbcst +transformed)
save = (variances) }
ARGUMENTS
exclude Number of observations excluded from the end of the series (or from the end of the span
specified by the span argument of the series spec, if present) before forecasting. The
default is to start forecasting from the end of the series (or span), i.e., exclude = 0.
lognormal Determines if an adjustment is made to the forecasts when a log transformation is speci-
fied by the user to reflect that the forecasts age generated from a log-normal distribution
(lognormal = yes). The default is not to make such an adjustment (lognormal = no).
maxback Number of backcasts produced. The default is 0 and 120 is the maximum. (The limit
of 120 can be changed—see Section 2.8.) Note: Backcasts are not produced when Seats
seasonal adjustments are specified, or if the starting date specified in the modelspan
argument of the series spec is not the same as the starting date of the analysis span
specified in the span argument of the series spec.
maxlead Number of forecasts produced. The default is 12 and 120 is the maximum. (The limit of
120 can be changed—see Section 2.8.)
print and save The optional output tables are listed on Table 7.16. The transformed and forecasts
tables are printed out by default. For a complete listing of the default and brief print
levels for this table, see Appendix B.
probability Coverage probability for prediction intervals, assuming normality. The default is proba-
bility=.95, in which case prediction intervals on the transformed scale are point forecast
± 1.96 × forecast standard error.
Table 7.16: Available Output Tables for Forecast

transformed ftr + forecasts on the transformed scale, with corre-
sponding forecast standard errors
variances fvr + forecast error variances on the transformed
scale, showing the contributions of the error as-
suming the model is completely known (stochas-
tic variance) and the error due to estimating any
regression parameters (error in estimating AR
and MA parameters is ignored)
forecasts fct + point forecasts on the original scale, along with
upper and lower prediction interval limits
transformedbcst btr + backcasts on the transformed scale, with corre-
sponding forecast standard errors
backcasts bct + point backcasts on the original scale, along with
upper and lower prediction interval limits
DETAILS
Forecasting is done with the estimated (or evaluated) model. If the estimate spec is not present, the forecast
spec will force estimation (with default options) to be performed before forecasting. The model used for
forecasting is that specified by the regression and arima specs. If the outlier spec is present, the model is
augmented by additional regression variables for any automatically identified outliers. Detected outliers can
affect forecasts indirectly, through their effect on model parameter estimates, as well as directly, when outliers
found near the end of the series affect the computation of the forecasts.
If the model includes one or more moving average operators then the forecasts will depend on the residuals
from the estimated model. The exact argument of the estimate spec determines whether these are computed
corresponding to exact likelihood (the default) or a form of conditional likelihood.
Forecast standard errors include an adjustment for error arising from estimation of any regression parameters
in the model, but do not include an adjustment for error arising from estimation of AR and MA parameters;
see formula (12.42) of Bell (2004) and, for a general approach, Kohn and Ansley (1985).
If the model contains user-defined regression variables, values for these must be provided for all time points
in the forecast period.
Prediction intervals on the transformed scale are defined as
point f orecast ± K × f orecast standard error,
where K denotes the standard error multiplier (from a table of the normal distribution) corresponding to the
specified coverage probability. Point forecasts and prediction interval limits on the original scale are obtained
7.7. FORECAST 103
by inverse transformation of those on the transformed scale, allowing for both transformation (Box-Cox or
logistic) and prior-adjustment factors (including the length-of-month or length-of-quarter adjustment implied
if variables = td is included in the regression spec). If the transform spec includes user-defined prior-
adjustment factors, these must be provided through the forecast period for the results to be inverse transformed.
If they are not provided through the forecast period, then they will be assumed to be 1 in the forecast period.
In this case, effects of the user-defined adjustments on the forecasts will not be (and cannot be) undone.
A reason for using exclude > 0 is to produce forecasts for some time points whose data are withheld for
purposes of evaluating the forecast performance of the model. X-13ARIMA-SEATS facilitates such comparisons by
printing actual forecast errors (observation − point forecast) at all time points in the forecast period for which
corresponding (transformed) observed data exist. Setting exclude > 0 produces within-sample comparisons,
since the data that are withheld from forecasting are not withheld from model estimation. More realistic out-
of-sample forecast comparisons are produced by withholding data from both model estimation and forecasting,
which can be accomplished by using the span argument of the series spec. (See Example 4.)
Whenever forecasts and/or backcasts are generated in an X-13ARIMA-SEATS run in which seasonal adjustment
is performed, they are appended to the original series, and the seasonal adjustment procedures are applied to
the forecast and/or backcast extended series. If a seasonal adjustment is specified in a run in which a regARIMA
model is used but the forecast spec is not, one year of forecasts are generated from the model. The only way
to specify a seasonal adjustment without forecast extension is to set maxlead = 0.
If preadjustments for regARIMA estimated trading day, outlier, holiday or user-defined regression effects
are prior adjusted from the original series, they are also adjusted out of the forecasts and backcasts.
Warning: if seasonal adjustment is specified by the x11 spec, exclude cannot be used to exclude observa-
tions from the end of the series. In case it is used, exclude will be set to zero and a warning message will be
printed.
EXAMPLES
Example 1 Forecast up through 12 steps ahead from the end of a monthly time series, and produce
95 percent prediction intervals. These are all default options. Though the estimate spec
is absent, the presence of the forecast spec forces model estimation with default estima-
tion options. The point forecasts and prediction interval limits for the transformed series
are exponentiated and then multiplied by mt /m̄ (to undo the length-of-month adjustment
produced by variables = td in the regression spec) to convert them back to the original
scale.
SERIES { TITLE = "Monthly sales" START = 1976.JAN

DATA = (138 128 ... 297) }
TRANSFORM { FUNCTION = LOG } REGRESSION { VARIABLES = TD } ARIMA {
MODEL = (0 1 1)(0 1 1)12 } FORECAST { }
Example 2 Forecast up through 24 steps ahead from the end of the same series used in Example 1. Since
the outlier spec is present, the estimated model used in forecasting will include any AO or
LS outliers detected, in addition to the trading-day variables specified by the regression
spec.

Data = (138 128 ... 297) }
Transform { Function = Log}
Regression { Variables = Td }
Arima { Model = (0 1 1)(0 1 1)12 }
Estimate { }
Outlier { }
Forecast { Maxlead = 24 }
Example 3 Exclude 10 data points and forecast up through 15 steps ahead. The entire time series is
used for parameter estimation, including the ten data points excluded at the end of the series
when forecasting. For these last 10 data points, the within-sample forecast errors will be
printed. At each forecast lead the prediction interval limits are obtained by exponentiating
point forecast of the log scale plus and minus 1.645 times the corresponding log forecast
standard error, which corresponds to the requested 90 percent coverage probability.

data = (138 128 ... 297) }
arima { model = (0 1 1)(0 1 1)12 }
estimate { }
forecast { maxlead = 15
probability = .90
exclude = 10 }
Example 4 The series ends in March, 1992, but the last 24 observations are excluded from model
estimation by using a span argument in the series spec. Then, using the model with these
parameter estimates, the last 24 observations are forecast from March, 1990, the end of the
span. The out-of-sample errors in forecasting the last 24 observations will be printed out.
(Contrast this with Example 3.)

data = (138 128 ... 297)
span = ( ,1990.mar) }
transform { function = log}
arima { model = (0 1 1)(0 1 1)12 }
estimate { }
forecast { maxlead = 24 }
7.7. FORECAST 105
Example 5 Forecast up through 12 months ahead from the end of a monthly time series, and produce
95 percent prediction intervals These are all default options. Also produce 12 backcasts of
the series, and perform a default multiplicative seasonal adjustment of the forecast- and
backcast-extended original series, prior-adjusted for trading day effects.
series { title = "monthly sales" start = 2000.jan

file = "ussales.dat" }
arima { model = (0 1 1)(0 1 1)12 }
forecast { maxback=12 }
x11{ }
Example 6 Same as Example 2, except that a log-normal correction will be applied to the forecasts.

Data = (138 128 ... 297) }
Transform { Function = Log}
Regression { Variables = Td }
Arima { Model = (0 1 1)(0 1 1)12 }
Estimate { }
Outlier { }
Forecast { Maxlead = 24 Lognormal = Yes }
7.8 HISTORY
DESCRIPTION
Optional spec for requesting a sequence of runs from a sequence of truncated versions of the time series for the
purpose of creating historical records of (i) revisions from initial (concurrent or projected) seasonal adjustments,
(ii) out-of-sample forecast errors, and (iii) likelihood statistics. The user can specify the beginning date of the
historical record and the choice of records (i) - (iii). If forecast errors are chosen, the user can specify a vector
of forecast leads. Warning: Generating the history analysis can substantially increase the program’s run time.
USAGE
history{ estimates = ( sadj fcst trend )
sadjlags = (1,2,3,12)
trendlags = (1,2,3)
target = final
start = 1975.jan
fstep = ( 1 2 )
fixmdl = no
fixreg = outlier
endtable = 1990.Jan
print = ( all -revvalsa )
save = ( sar trr fcsterrors )
savelog = ( aveabsrevsa aveabsrevtrend )
}
ARGUMENTS
endtable Specifies the final date of the output tables of the revisions history analysis of seasonal
adjustment and trend estimates and their period-to-period changes. This can be used
to ensure that the revisions history analysis summary statistics are based only on final
(or nearly final) seasonal adjustments or trends. If endtable is not assigned a value,
it is set to the date of the observation immediately before the end of the series or to a
value one greater than the largest lag specified in sadjlags or trendlags. This option
has no effect on the historical analysis of forecasts and likelihood estimates. Example:
endtable=1990.jun.
estimates Determines which estimates from the regARIMA modeling and/or the X-11 seasonal
adjustment will be analyzed in the history analysis. Example: estimates=(sadj aic).
The default is the seasonally adjusted series (sadj). Table 7.17 gives a description of the
available estimates.
fixmdl Specifies whether the regARIMA model will be re-estimated during the history analysis.
If fixmdl=yes, the ARIMA parameters and regression coefficients of the regARIMA
model will be fixed throughout the analysis at the values estimated from the entire series
7.8. HISTORY 107
Table 7.17: Choices Available for the estimates Argument
name description of option

sadj Final seasonally adjusted series (and indirect seasonally adjusted
series, if composite seasonal adjustment is performed.
sadjchng Month-to-month (or quarter-to-quarter) changes in the final sea-
sonally adjusted series.
trend Final Henderson trend component.
trendchng Month-to-month (or quarter-to-quarter) changes in the final Hen-
derson trend component.
seasonal Final and projected seasonal factors.
aic AICCs and maximum log likelihoods for the regARIMA model.
fcst Forecasts and evolving mean square forecast errors generated from
the regARIMA model. Warning: This option can be used only
when forecasts are produced, see the forecast spec in Section 7.7.
(or model span, if one is specified via the modelspan argument). If fixmdl=no, the
regARIMA model parameters will be re-estimated each time the end point of the data is
changed. The default is fixmdl=no. This argument is ignored if no regARIMA model is
fit to the series.
fixreg Specifies the fixing of the coefficients of a regressor group, either within a regARIMA
model or an irregular component regression. These coefficients will be fixed at the values
obtained from the model span (implicit or explicitly) indicated in the series or composite
spec. All other coefficients will be re-estimated for each history span. Trading day
(td), holiday (holiday), outlier (outlier), or other user-defined (user) regression effects
can be fixed. This argument is ignored if neither a regARIMA model nor an irregular
component regression model is fit to the series, or if fixmdl=yes.
fstep Specifies a vector of up to four (4) forecast leads that will be analyzed in the history
analysis of forecast errors. Example: fstep=(1 2 12) will produce an error analysis for
the 1-step, 2-step, and 12-step ahead forecasts. The default is (1 12) for monthly series
or (1 4) for quarterly series. Warning: The values given in this vector cannot exceed
the specified value of the maxlead argument of the forecast spec, or be less than one.
Table 7.19. For a complete listing of the brief and default print levels for this spec,
see Appendix B.
sadjlags Specifies a vector of up to 5 revision lags (each greater than zero) that will be analyzed
in the revisions analysis of lagged seasonal adjustments. The calculated revisions for
these revision lags will be those of the seasonal adjustments obtained using this many
observations beyond the time point of the adjustment. That is, for each value revisionlagi
given in sadjlags, series values through time t + revisionlagi will be used to obtain
the adjustment for time t whose revision will be calculated. For more information, see
DETAILS.
Table 7.18: Default Output Tables for History Spec

header hdr · header for history analysis
outlierhistory rot + record of outliers removed and kept for the revisions his-
tory (printed only if automatic outlier identification is
used)
sarevisions sar + revision from concurrent to most recent estimate of the
seasonally adjusted data
sasummary sas · summary statistics for seasonal adjustment revisions
chngrevisions chr + revision from concurrent to most recent estimate of the
month-to-month (or quarter-to-quarter) changes in the
seasonally adjusted data
chngsummary chs + summary statistics for revisions in the month-to-month
(or quarter-to-quarter) changes in the seasonally adjusted
data
indsarevisions iar + revision from concurrent to most recent estimate of the
indirect seasonally adjusted series
indsasummary ias · summary statistics for indirect seasonal adjustment revi-
sions
trendrevisions trr + revision from concurrent to most recent estimate of the
trend component
trendsummary trs · summary statistics for trend component revisions
trendchngrevisions tcr + revision from concurrent to most recent estimate of the
month-to-month (or quarter-to-quarter) changes in the
trend component
trendchngsummary tcs · summary statistics for revisions in the month-to-month
(or quarter-to-quarter) changes in the trend component
sfrevisions sfr + revision from concurrent to most recent estimate of the
seasonal factor, as well as projected seasonal factors.
sfsummary ssm · summary statistics for seasonal factor revisions
lkhdhistory lkh + history of AICC and likelihood values
fcsterrors fce + revision history of the accumulated sum of squared fore-
cast errors
argument.
7.8. HISTORY 109
Table 7.19: Other Output Tables for History Spec

sfilterhistory sfh + record of seasonal filter selection for each observation in
the revisions history (printed only if automatic seasonal
filter selection is used)
saestimates sae + concurrent and most recent estimate of the seasonally ad-
justed data
chngestimates che + concurrent and most recent estimate of the month-to-
month (or quarter-to-quarter) changes in the seasonally
adjusted data
indsaestimates iae + concurrent and most recent estimate of the indirect sea-
sonally adjusted data
trendestimates tre + concurrent and most recent estimate of the trend compo-
nent
trendchngestimates tce + concurrent and most recent estimate of the month-to-
month (or quarter-to-quarter) changes in the trend com-
ponent
sfestimates sfe + concurrent and most recent estimate of the seasonal fac-
tors and projected seasonal factors
fcsthistory fch + listing of the forecast and forecast errors used to generate
accumulated sum of squared forecast errors
argument.
This option is meaningful only if the revisions history of the seasonally adjusted series or
month-to-month (quarter-to-quarter) changes in the seasonally adjusted series is specified
in the estimates argument. The default is no analysis of revisions of lagged seasonal
adjustments.
savelog The diagnostics available for output to the log file (see Section 2.6) are listed on Table
7.20.
start Specifies the starting date of the revisions history analysis. If this argument is not used,
its default setting depends on the length of the longest seasonal filter used, provided that
a seasonal adjustment is being performed (if there is no conflict with the requirement that
60 earlier observations be available when a regARIMA model is estimated and fixmdl=no,
the default for fixmdl). The default starting date is six (6) years after the start of the
series, if the longest filter is either a 3x3 or stable filter, eight (8) years for a 3x5 filter,
and twelve (12) years for a 3x9 filter. If no seasonal adjustment is done, the default is 8
years after the start of the series. Example: start=1990.jun.
target Specifies whether the deviation from the concurrent estimate or the deviation from the
final estimate defines the revisions of the seasonal adjustments and trends calculated at
Table 7.20: Available Log File Diagnostics for History

alldiagnostics all All revisions diagnostics listed in this table
aveabsrevsa asa average absolute revision of the seasonally adjusted
series
aveabsrevchng ach average absolute revision of the month-to-month (or
quarter-to-quarter) changes in the seasonally ad-
justed data
aveabsrevindsa iaa average absolute revision of the indirect seasonally
adjusted series
aveabsrevtrend atr average absolute revision of the final trend compo-
nent
aveabsrevtrendchng atc average absolute revision of the month-to-month (or
quarter-to-quarter) changes in the trend component
aveabsrevsf asf average absolute revision of the final seasonal factors
aveabsrevsfproj asp average absolute revision of the projected seasonal
factors
avesumsqfcsterr afe average sum of squared forecast error for each fore-
cast lag
the lags specified in sadjlags or trendlags. The default is target=final; the alternative
is target=concurrent.
trendlags Similar to sadjlags, this argument prescribes which lags will be used in the revisions
history of the lagged trend components. Up to 5 integer lags greater than zero can be
specified.
This option is meaningful only if the revisions history of the final trend component or
month-to-month (quarter-to-quarter) changes in the final trend component is specified in
the estimates argument. The default is no analysis of revisions lagged trend estimates.
fixx11reg Specifies whether the irregular component regression model specified in the x11regression
spec will be re-estimated during the history analysis. If fixx11reg=yes, the regression
coefficients for the irregular component regression model are fixed throughout the analysis
at the values estimated from the entire series. If fixx11reg=no, the irregular component
regression model parameters will be re-estimated each time the end point of the history
interval is advanced.
7.8. HISTORY 111
The default is fixx11reg=no. This argument is ignored if no irregular component re-

gression model is specified.
outlier Specifies whether automatic outlier detection is to be performed whenever the regARIMA
model is re-estimated during the revisions history analysis. This argument has no effect
if the outlier spec is not used.
If outlier=keep, all outliers automatically identified using the full series are kept in the
regARIMA model during the revisions history analysis. The coefficients estimating the
effects of these outliers are re-estimated whenever the other regARIMA model parameters
are re-estimated. No additional outliers are automatically identified and estimated. This
is the default setting.
If outlier=remove, those outlier regressors that were added to the regression part of
the regARIMA model when automatic outlier identification was performed on the full
series are removed from the regARIMA model during the revisions history analysis.
Consequently, their effects are not estimated and removed from the series. This option
gives the user a way to investigate the consequences of not doing automatic outlier
identification.
If outlier=auto, among outliers automatically identified for the full series, only those
that fall in the time period up to outlierwin observations before the starting date of the
revisions history analysis are automatically included in the regARIMA model. In each
run of the estimation procedure with a truncated version of the original series, automatic
outlier identification is performed only for the last outlierwin+1 observations. An outlier
that is identified is used for the current run, but is only retained for the subsequent runs
of the historical analysis if it is at least outlierwin observations from the end of the
subsequent span of data being analyzed.
outlierwin Specifies how many observations before the end of each span will be used for outlier
identification during the revisions history analysis. The default is 12 for monthly series
and 4 for quarterly series. This argument has an effect only if the outlier spec is used,
and if outlier=auto in the history spec.
refresh Specifies which of two sets of initializing values is used for the regARIMA model parame-
ter estimation. If refresh=yes, the parameter estimates from the last model evaluation
are used as starting values for the current regARIMA model estimation done during
the revisions history. If refresh=no, then the initial values of the regARIMA model
parameters will be set to the estimates derived from the entire series. The default is
refresh=no.
transformfcst Specifies whether the forecast error output of the history analysis is for forecasts of
the original data (transformfcst = no) or for the forecasts of the transformed data
specified by the transform spec (transformfcst = yes). See Details. The default is
transformfcst = no.
x11outlier Specifies whether the AO outlier identification will be performed during the history anal-
ysis for all irregular component regressions that result from the x11regression spec. If
x11outlier=yes, AO outlier identification will be performed for each of the history runs.
Those AO outlier regressors that were added to the irregular component regression model
when automatic AO outlier identification was done for the full series are removed from the
irregular component regression model prior to the history runs. If x11outlier=no, then
the AO outlier regressors automatically identified are kept for each of the history runs. If
the date of an outlier detected for the complete span of data does not occur in the data
span of one of the history runs, the outlier will be dropped from the model for that run.
The coefficients estimating the effects of these AO outliers are re-estimated whenever the
other irregular component regression model parameters are re-estimated. However, no
additional AO outliers are automatically identified and estimated. This option is ignored
if the x11regression spec is not used, if the selection of the aictest argument results
in the program not estimating an irregular component regression model, or if the sigma
argument is used in the x11regression spec. The default is x11outlier=yes.
DETAILS
Section 6.3 gives technical details on revisions history analysis. For some supporting theory for out-of-sample
squared forecast error diagnostic output, see Findley (2005).
When a revision history analysis of the seasonally adjusted series is specified for a composite seasonal
adjustment, the revisions of both the direct and indirect seasonally adjusted series can be produced. The
revision history analysis must be specified for each of the component series, even for those component series
that are not seasonally adjusted, see the Examples section of composite spec documentation in Section 7.4.
The revision history of the indirect seasonally adjusted series (sadj in Table 7.17) is the only revision history
available for indirect seasonal adjustments.
In each input specification file, a starting date for the history analysis must be specified using the start
argument of this spec, and this starting date should be the same for each of the components and the composite
series. If this is not the case, then only the history analysis of the direct seasonal adjustment will be performed.
If the automatic seasonal filter selection option is used, the program will redo the choice of seasonal filter
each time the data span is changed in the revision history analysis. If the seasonal filter should change in the
course of the analysis, a warning message will be printed out, and a table of the seasonal filter lengths chosen
for each data span will be printed out.
The starting date for the forecast revisions depends on the values given for fstep. The starting date for a
history of n-step-ahead forecast errors is n periods after the starting date of the history analysis. Example:
if fstep = (1 12) and start = 1992.jan, the history for the 1-step and 12-step ahead forecasts will start in
February of 1992 and January of 1993, respectively.
In some situations, the program automatically switches to using fixed model coefficients for the history
analysis. This happens when the start of the revisions history analysis (which can be set by the user with the
start argument) causes some truncated data span to have fewer than 60 observations for regARIMA model
estimation, either because of the series length or a span or modelspan argument value (in the series or
composite spec). In this case, the coefficients (ARIMA and regression) of the regARIMA model will be held
fixed throughout the analysis at the values estimated from the entire series (or model span, if one is specified).
Fixing of the coefficients will also occur for every truncated data span that contains data later than the
ending date specified in a modelspan argument. In particular, in the extreme case, when the ending date of
7.8. HISTORY 113
the model span is earlier than the starting date of the history analysis, the coefficients of the regARIMA model
will be fixed throughout the history analysis.
Regression models from the x11regression spec are treated similarly. For example, their coefficients are
fixed if some truncated data span has fewer than 60 observations because of a date assigned to the span
argument of x11regression.
If an outlier specified by the user occurs in the period after the starting date of the revision history, that
outlier will be dropped from the model at the start of the revision history analysis. It will be re-introduced into
the regARIMA model when enough data have been added for the outlier variable to be defined. User-defined
regressors are treated in the same way.
EXAMPLES
The following examples illustrating all the steps of a composite adjustment show complete spec files.
Example 1 A multiplicative monthly seasonal adjustment is to be performed with 3x9 seasonal moving
averages for all months without using regARIMA model forecasts, backcasts, or regression
outlier adjustments. A revision history of just the seasonally adjusted series will be per-
formed (remember, this is the default history) for all data, after a startup period of twelve
years (because 3x9 seasonal factors are used), with an additional analysis on the estimates
made 2 periods after the concurrent observation.
Series { Title = "Sales Of Livestock" Start = 1967.1

File = "cattle.ori" }
X11 { SeasonalMA = S3X9 }
History { sadjlags = 2 }
Example 2 Utilize a seasonal ARIMA model with regression variables for outlier and level shift pread-
justment. The specified regression variables are a constant, trading day effects, and two
level shifts, one in May 1972 and one in Sept. 1976. The ARIMA part of the model is
(0,1,2)(1,1,0)12. Generate a history of the 1-step ahead forecast errors. Start the analysis
in January of 1975; this means the first 1-step ahead forecast error in the analysis is for
February of 1975.
series { title = "Exports of Leather goods"

start = 1969.jul file = "expleth.dat" }
regression { variables = (const td ls1972.may ls1976.oct) }
arima { model = (0 1 2)(1 1 0) }
estimate { }
history { estimates = fcst fstep = 1 start=1975.jan }
Example 3 Using the same regARIMA model and data as in Example 2, generate a history of the 1-step
and 12-step ahead forecast errors. Start the history in January of 1975. Save the history to
a file. In this file, zeros will be printed for the estimates where the 12-step ahead forecast
errors are not defined (in this case, February to December of 1975) in order to maintain a
uniform format for the file.
series { title = "Exports of Leather goods"

start = 1969.jul
file = "expleth.dat" }
regression { variables = (const td ls1972.may ls1976.oct) }
arima { model = (0 1 2)(1 1 0) }
estimate { }
history { estimates = fcst save = r6 start = 1975.jan }
Example 4 A multiplicative monthly seasonal adjustment is to be performed, with 3x3 seasonal moving
averages, using regARIMA model forecasts to extend the series. The regARIMA model will
be fit to the data up to the last December available to the series. A revision history of the
seasonally adjusted series and the trend component will be calculated starting after the sixth
year of the series, with the regARIMA model parameters re-estimated every December. Also,
the history of the seasonal adjustment revisions of this series is integrated into the revision
history calculation of the indirect seasonal adjustment of the composite series of which this
series is a component. (The spec file for the composite series in the metafile must include
an appropriate history spec, see Example 5.)
series { title = "Housing Starts in the Midwest"

start = 1967.1
file = "hsmwtot.ori"
modelspan = (,0.Dec)
comptype=add
}
arima { model = (0 1 2)(0 1 1) }
x11 { seasonalMA = S3X3 }
history { estimates = (sadj trend) }
Example 5 A composite monthly seasonal adjustment is to be performed with 3x3 seasonal moving
averages for all months using regARIMA model forecasts to extend the composite series.
The regARIMA model will be fit to the data up to the last December available to the
series. A revision history of both the direct and indirect seasonally adjusted series and
the trend component from the direct seasonal adjustment will be performed, with the reg-
ARIMA model parameters re-estimated every December. The percent revisions for each of
the estimates will be stored in separate files.
composite{ title = "Total Housing Starts in the US"

modelspan = (,0.Dec)
}
arima { model = (0 1 1)(0 1 1) }
x11 { seasonalMA = S3X3 }
history { estimates = (sadj trend)
save = (sar iar trr) }
7.9. METADATA 115
7.9 METADATA
DESCRIPTION
Specification that allows users to insert metadata into the diagnostic summary file. Users can specify keys and
corresponding values for those keys to insert additional information into the diagnostic summary file stored by
X-13ARIMA-SEATS .
USAGE
metadata { keys = (
"survey"
"analyst"
)
values = (
"United States retails sales"
"Dr. Sigerson"
)
}
ARGUMENTS
keys A list of character strings used as keys for the metadata values specified in the values
list. Up to 20 values can be specified - no single key can be more than 132 characters
long, and all the keys taken together cannot be exceed 2000 characters.
An example with two keys is:
keys = ( "note1"
"note2" )
If a list with more than one entry is used, each key must be on a separate line of the spec
file. The keys should not contain spaces or colons (periods, commas and semicolons can
be used), and should be unique values - each key must be different. Missing values and
blank lines are not allowed.
values A list of character strings used as values associated with the keys provided in the keys
argument. Up to 20 values can be specified - no single entry can be more than 132
characters long, and all the entries taken together cannot exceed 2000 characters.
An example with two arguments is:
values= ( "Special sale caused outlier in October 2005"

"Analysis as of November 2006" )
If a list with more than one entry is used, each value must be on a separate line of the
spec file. Missing values and blank lines are not allowed.
DETAILS
The metadata spec allows users to insert their own metadata into the summary diagnostics file. Users can
specify unique keys and corresponding values for those keys, and these values are then entered as records into
the summary diagnostics file.
These records are formatted as
metadata.key: value
where key is a unique key specified by the user, and value is the corresponding value for that key. The text
"metadata." signifies that this is user-defined metadata.
For example, when the user includes the following metadata spec into an input specification file:
metadata {
keys = (
"analyst"
"date.reviewed"
"units.of.measure"
)
values = (
"Allen Smithee"
"June 15, 2006"
"Millions of Dollars"
)
}
the following records will be written to the summary diagnostics file:
metadata.analyst: Allen Smithee

metadata.date.reviewed: June 15, 2006
metadata.units.of.measure: Millions of Dollars
In previous versions of X-13ARIMA-SEATS , the summary diagnostics file was generated only when the -s or
-g runtime flags are used; now the summary diagnostics file will also be generated whenever the metadata spec
is used.
If there are fewer keys than there are values, a warning message is produced and the program will generate
unique keys based on the position of the value in the array.
7.9. METADATA 117
For example, the following metadata spec:
metadata {
keys = (
"analyst"
"date.reviewed"
)
values = (
"Allen Smithee"
"June 15, 2006"
)
}
produces the following records in the summary diagnostics file:
metadata.analyst: Allen Smithee

metadata.date.reviewed: June 15, 2006
metadata.key3: Millions of Dollars
Not specifying a key argument at all will force the program to generate unique keys for all the values specified.
metadata {
values = (
"Allen Smithee"
"June 15, 2006"
)
}
produces the following records in the summary diagnostics file:
metadata.key1: Allen Smithee

metadata.key2: June 15, 2006
metadata.key3: Millions of Dollars
If more keys are specified than values, execution will cease and an error message will be produced.
Note that the metadata spec can appear in any order relative to the the other specs - it can be the first
spec in the spec file, etc.
EXAMPLES
Example 1 Print all available diagnostic checks of the residuals from the specified model. The check
spec forces model estimation to be performed (with default options) even though the esti-
mate spec is not present. The metadata spec documents the analyst that developed the
spec file.

start = 1964.jan file = "sales1.dat" }
regression { variables = td aictest = ( td easter ) }
arima { model = (0 1 1)(0 1 1) }
check { print = (all) }
outlier { types = all }
metadata {
key = "analyst"
value = "John J. J. Smith"
}
The record stored in the summary diagnostic file is
metadata.analyst: John J. J. Smith
Example 2 For the same series and model as in Example 1, produce all diagnostic checking statistics
except the except the printed table and plot of the residual PACF. The residual ACF is
computed through lag 24.

regression {
variables = ( td ao1967.jun ls1971.jun easter[8] )
)
}
arima { model = (0 1 1)(0 1 1) }
check { print = (all -pacf -pacfplot) }
metadata {
key = (
"analyst"
"spec.updated"
)
value = (
"John J. J. Smith"
"October 31, 2006"
)
}
7.9. METADATA 119

metadata.spec.updated: October 31, 2006
Example 3 For the same series and model as in Example 2, add metadata text to describe the outliers
found by the automatic outlier procedure.

regression {
variables = (
td ao1967.jun ls1971.jun easter[15]
)
}
arima { model = (0 1 1)(0 1 1) }
check { print = (all -pacf -pacfplot) }
x11 { save = d11 }
metadata {
key = (
"analyst"
"spec.final"
)
value = (
"John J. J. Smith"
"November 10, 2006"
"AO caused by strike, LS caused by survey change"
)
}

metadata.spec.updated: November 10, 2006
metadata.key3: AO caused by strike, LS caused by survey change
7.10 IDENTIFY
DESCRIPTION
Specification to produce tables and line printer plots of sample ACFs and PACFs for identifying the ARIMA
part of a regARIMA model. Sample ACFs and PACFs are produced for all combinations of the nonseasonal and
seasonal differences of the data specified by the diff and sdiff arguments. If the regression spec is present, the
ACFs and PACFs are calculated for the specified differences of a series of regression residuals. If the regression
spec is not present, the ACFs and PACFs are calculated for the specified differences of the original data.
USAGE
identify { diff = (0, 1)
sdiff = (0, 1)
maxlag = 36
print = (none +acf +acfplot +pacf +pacfplot) }
ARGUMENTS
diff Orders of nonseasonal differencing specified. The value 0 specifies no differencing, the
value 1 specifies one nonseasonal difference (1 − B), the value 2 specifies two nonseasonal
differences (1−B)2 , etc. The specified ACFs and PACFs will be produced for all orders of
nonseasonal differencing specified, in combination with all orders of seasonal differencing
specified in sdiff. The default is diff=(0).
maxlag The number of lags specified for the ACFs and PACFs for both tables and plots. The
default is 36 for monthly series, 12 for quarterly series.
print and save Table 7.21 gives the available output tables for this spec. All these tables are included in
the default printout, except regcoefficients. For a complete listing of the brief and
default print levels for this spec, see Appendix B.
sdiff Orders of seasonal differencing specified. The value 0 specifies no seasonal differencing,
the value 1 specifies one seasonal difference (1−B s ), etc. The specified ACFs and PACFs
will be produced for all orders of seasonal differencing specified, in combination with all
orders of nonseasonal differencing specified in diff. The default is sdiff=(0).
DETAILS
If the regression spec is present, the program differences the series (after processing by the transform spec) and
the regression variables using the maximum order of differencing specified by the diff and sdiff arguments. The
differenced series is then regressed on the differenced regression Pvariables. The resulting regression coefficients
(β̃i ) are then used to calculate undifferenced regression effects ( i β̃i xit ), which are then subtracted from the
P
undifferenced data (yt ) to produce a time series of undifferenced regression errors (z̃t = yt − i β̃i xit ). This
7.10. IDENTIFY 121
Table 7.21: Available Output Tables for Identify

acf iac + sample autocorrelation function(s), with standard
errors and Ljung-Box Q-statistics for each lag
acfplot acp · line printer plot of sample autocorrelation func-
tion(s) with ± 2 standard error limits shown on
the plot
pacf ipc + sample partial autocorrelation function(s) with
standard errors for each lag
pacfplot pcp · line printer plot of sample partial autocorrelation
function(s) with ± 2 standard error limits shown
on the plot
regcoefficients rgc · Regression coefficients removed from the trans-
formed series before ACFs and PACFs were gen-
erated.
regression error time series and its differences as specified by diff and sdiff, are then used to produce the ACFs
and PACFs.
There is one exception to the above. If a constant term is specified in the regression spec (variables =
(const ...)), it is included when the regression is done but not when the regression effects are subtracted
from the series. See Section 4.4 for more discussion.
ACFs and PACFs are produced for all combinations of nonseasonal and seasonal differencing orders specified
in diff and sdiff. For example, if diff = (0, 1) and sdiff = 1 are specified, then ACFs and PACFs are
computed for (1 − B s )z̃t and (1 − B)(1 − B s )z̃t , where z̃t is the series of regression errors, as discussed above, and
s is the seasonal period specified in the series spec. If diff = (0, 1, 2) and sdiff = (0, 1) are specified,
then ACFs and PACFs are computed for six series: z̃t , (1 − B)z̃t , (1 − B)2 z̃t , (1 − B s )z̃t , (1 − B)(1 − B s )z̃t ,
and (1 − B)2 (1 − B s )z̃t .
If both the identify and estimate specs are present, the identify spec is processed first. Note that the
identify spec uses information from the regression spec, if present, but ignores the arima spec.
Users should make sure that differencing does not produce a singularity among the regression variables,
including any user-defined regression variables, as singularities will cause a fatal error. One way this would
arise is if sdiff was assigned a positive value (e.g., 1), while variables = (seasonal) was included in the
regression spec.
If the number of lags requested for ACFs and PACFs equals or exceeds the length of the series (or the
differenced series), the ACF and PACF will be computed only through the highest lag possible.
EXAMPLES
Example 1 Produce ACF tables useful for identifying the degree of differencing required for the monthly
series yt = log(Yt ), where Yt is the original data input in the series spec. The ACFs are
calculated for yt , (1 − B)yt , (1 − B 12 )yt , and (1 − B)(1 − B 12 )yt . The regression spec is
absent so no regression effects are removed. ACFs are calculated through lag 36, the default
for a monthly time series.
series { title = "Monthly Sales" start = 1976.jan

data = (138 128 ... 297) }
identify { diff = (0, 1)
sdiff = (0, 1)
print = (none +acf) }
Example 2 Remove fixed seasonal effects before computing sample ACFs and sample PACFs. The
regression spec includes a trend constant as well as the fixed seasonal variables. The
identify spec removes the fixed seasonal effects by regressing (1 − B)yt on the differenced
regression variables (1 − B)xit , and computing undifferenced regression residuals z̃t = yt −
P12
i=2 β̃i xit (not subtracting out the trend constant term β̃1 x1t ). It then computes ACFs
and PACFs of z̃t and (1 − B)z̃t . The constant term allows for an overall nonzero mean in
(1 − B)yt , so it is a linear trend constant, i.e., x1t = t.
SERIES { TITLE = "MONTHLY SALES" START = 1976.JAN

DATA = (138 128 ... 297) }
REGRESSION { VARIABLES = (CONST SEASONAL) }
IDENTIFY { DIFF = (0,1) }
Example 3 Produce ACF and PACF plots to identify the AR and MA parts of a regARIMA model. Do
not print ACF and PACF tables. Suppose Yt is the same series as in Example 1, that one
nonseasonal and one seasonal difference are chosen, and that the model will include trading-
day and Easter holiday effects. Because the regression spec is present, the identify spec
first regresses (1 − B)(1 − B 12 )yt on (1 − B)(1 − B 12 )xit , where the xit are the regression
variables for the trading-day and Easter holiday effects, and yt consists of the logarithms
of the original data Yt adjusted for length-of-month effects. (See the description of td in
the regression spec.) If β̃i denote the estimated regression coefficients, then this identify
spec produces ACF and PACF plots for the regression residual series (1 − B)(1 − B 12 ) yt −
P
i β̃i xit . The ACFs and PACFs are computed through lag 30.
Series { Title = "Monthly Sales" Start = 1976.Jan

Data = (138 128 ... 297) }
Transform { Function = Log }
Regression { Variables = (Td Easter[14])}
Identify { Diff = (1) Sdiff = (1) Maxlag = 30
Print = (None +ACFplot +PACFplot) }
7.10. IDENTIFY 123
Example 4 Produce ACFs and PACFs (through lag 16) for model identification, and also estimate a
tentative model for a quarterly series. There is a known level shift in the first quarter of
1971. Its effect is estimated by regressing (1 − B)(1 − B 4 )yt on the differenced level shift
variable. This regression effect is then removed to produce the (undifferenced) regression
residual series, z̃t = yt − β̃LS1971.1t , and ACFs and PACFs are calculated for z̃t , (1 − B)z̃t ,
(1 − B 4 )z̃t , and (1 − B)(1 − B 4 )z̃t . The identify spec ignores the information in the arima
spec.
The spec file below also specifies estimation and standard diagnostic checks of the regARIMA
model, (1 − B)(1 − B 4 ) yt − βLS1971.1t = (1 − θB)(1 − ΘB 4 )at . Such an estimation of a
tentative model on the same run that produces ACFs and PACFs for model identification
is sometimes useful, if one has a prior idea what ARIMA model might be appropriate. This
might be the case if the series had been modelled previously, but new data has since extended
the series. If the diagnostic checks suggest that the tentative model is inadequate, the user
will have information from both the diagnostic checks and the identify spec output to use
in selecting a new model.
series { title = "Quarterly Sales" start = 1963.1 period = 4

data = (56.7 57.7 ... 68.0) }
regression { variables = (ls1971.1) }
arima { model = (0 1 1)(0 1 1) }
identify { diff = (0, 1) sdiff = (0, 1) maxlag = 16 }
estimate { }
check { }
7.11 OUTLIER
DESCRIPTION
Specification to perform automatic detection of additive (point) outliers, temporary change outliers, level shifts,
or any combination of the four using the specified model. After outliers (referring to any of the outlier types
mentioned above) have been identified, the appropriate regression variables are incorporated into the model
as “Automatically Identified Outliers”, and the model is re-estimated. This procedure is repeated until no
additional outliers are found. If two or more level shifts are detected (or are present in the model due to the
specification of level shift(s) in the regression spec), t-statistics can be computed to test null hypotheses that
each run of two or more successive level shifts cancels to form a temporary level shift.
USAGE
outlier{ types = all
critical = 3.75
method = addall
span = (1983.may, 1992.sep)
lsrun = 0
print = (none +header)
save = tests
savelog = id
}
ARGUMENTS
critical Sets the value to which the absolute values of the outlier t-statistics are compared to
detect outliers. The default critical value is determined by the number of observations
in the interval searched for outliers (see the span argument below). It is obtained by a
modification of the asymptotic formula of Ljung (1993) that interpolates critical values
for numbers of observations between 3 and 99. Table 7.22 gives default critical values for
various outlier span lengths.
If only one value is given for this argument (critical = 3.5), then this critical value is
used for all types of outliers. If a list of up to three values is given (critical = (3.5,
4.0, 4.0)), then the critical value for additive outliers is set to the first list entry (3.5
in this case), the critical value for level shift outliers is set to the second list entry
(4.0), and the critical value for temporary change outliers is set to the third list entry
(4.0). A missing value, as in critical = (3.25,,3.25), is set to the default critical
value. Raising the critical value decreases the sensitivity of the outlier detection routine,
possibly decreasing the number of observations treated as outliers.
lsrun Compute t-statistics to test null hypotheses that each run of 2, . . . , lsrun successive level
shifts cancels to form a temporary level shift. The t-statistics are computed as the sum
of the estimated parameters for the level shifts in each run divided by the appropriate
7.11. OUTLIER 125
Table 7.22: Default Critical Values for Outlier Identification
Number of Outlier Number of Outlier

Observations Critical Observations Critical
Tested Value Tested Value
1 1.96 48 3.63
2 2.24 72 3.73
3 2.44 96 3.80
4 2.62 120 3.85
5 2.74 144 3.89
6 2.84 168 3.92
7 2.92 192 3.95
8 2.99 216 3.97
9 3.04 240 3.99
10 3.09 264 4.01
11 3.13 288 4.03
12 3.16 312 4.04
24 3.42 336 4.05
36 3.55 360 4.07
standard error. (See Otto and Bell 1993). Both automatically identified level shifts and
level shifts specified in the regression spec are used in the tests. Lsrun may be given
values from 0 to 5; 0 and 1 request no computation of temporary level shift t-statistics.
If the value specified for lsrun exceeds the total number of level shifts in the model
following outlier detection, then lsrun is reset to this total. The default value for lsrun
is 0, i.e., no temporary level shift t-statistics are computed.
For details on handling temporary level shifts, see DETAILS.
method Determines how the program successively adds detected outliers to the model. The
choices are method = addone or method = addall. See DETAILS for a description of
these two methods. The default is method = addone.
print and save Table 7.23 gives the available output tables for this spec. The header and temporaryls
tables are printed out by default. For a complete listing of the default and brief print
levels for this table, see Appendix B.
Note: The entry for an outlier t-statistic in the finaltests table is set to zero whenever
testing for the outlier (regressor) causes the regression matrix to be singular, and for
any outliers specified in the variables argument of the regression spec. Also, when the
finaltests table is saved, the t-statistics for all automatically identified outliers are also
set to zero. This table cannot be saved when automatic model selection is invoked using
either the automdl or pickmdl specs.
savelog Setting savelog=identified or savelog=id causes a list of automatically identified
outliers to be output to the session log file (see section 2.6 for more information on the
log file).
span Specifies start and end dates of a span of the time series to be searched for outliers. The
Table 7.23: Available Output Tables for Outlier

header hdr · options specified for outlier detection including criti-
cal value, outlier span, and types of outliers searched
for
iterations oit + detailed results for each iteration of outlier detection
including outliers detected, outliers deleted, model
parameter estimates, and robust and non-robust es-
timates of the residual standard deviation
tests ots · t-statistics for every time point and outlier type on
each outlier detection iteration
temporaryls tls · summary of t-statistics for temporary level shift
tests
finaltests fts + t-statistics for every time point and outlier type gen-
erated during the final outlier detection iteration
start and end dates of the span must both lie within the series and within the model
span if one is specified by the modelspan argument of the series spec, and the start
date must precede the end date. A missing value, e.g., span = (1976.jan, ), defaults
to the start date or end date of the series, as appropriate. (If there is a span argument
in the series spec, then, in the above remarks, replace the start and end dates of the
series by the start and end dates of the span given in the series spec.)
types Specifies the types of outliers to detect. The choices are: types = ao, detect additive
outliers only; types = ls, detect level shifts only; types = tc, detect temporary change
outliers only; types = all, detect additive outliers, temporary change outliers, and level
shifts simultaneously; or types = none, turn off outlier detection (but not t-statistics for
temporary level shifts). The default is types = (ao ls).
almost Differential used to determine the critical value used for a set of ”almost” outliers -
outliers with t-statistics near the outlier critical value that are not incorporated into the
regARIMA model. After outlier identification, any outlier with a t-statistic larger than
Critical − almost is considered an ”almost outlier,” and is included in a separate table.
The default is almost = 0.5; values for this argument must always be greater than zero.
7.11. OUTLIER 127
tcrate Defines the rate of decay for the temporary change outlier regressor. This value must
be a number greater than zero and less than one. The default value is tcrate=0.7
** (12/period), where period is the number of observations in one year (for monthly
time series, 4 for quarterly time series). This formula for the default value of tcrate
ensures the same rate of decay over an entire year for series of different periodicity. If
the frequency of the time series is less than 4 (ie, period < 4), then there is no default
value, and the user will have to enter a value of tcrate if a temporary change outlier
was specified in the variables argument of the regression spec, or if temporary change
outliers are specified in the types argument of this spec. If this argument is specified in
the regression spec, it is not necessary to include it in this spec.
DETAILS
A level shift (LS) at the first data point cannot be estimated since the level of the series prior to the given data
is unknown. Therefore, no LS test statistics is calculated for the first data point. Also, an LS at the last data
point cannot be distinguished from an AO there, and an LS at the second data point cannot be distinguished
from an AO at the first data point. Thus, no LS statistics are calculated for the second and last data points if
AOs are also being detected. LS statistics that are not calculated are set to and printed out as 0.
Similarly, a temporary change (TC) outlier at the last data point cannot be distinguished from an AO there,
so no TC statistics is calculated for the last data point if AOs are also being detected. TC statistics that are
not calculated are set to and printed out as 0.
The addone method works in the following way. The program calculates t-statistics for each type of outlier
specified (AO, TC and/or LS) at all time points for which outlier detection is being performed. If the maximum
absolute outlier t-statistic exceeds the critical value, then an outlier has been detected and the appropriate
regression variable is added to the model. The program then estimates the new model (the old model with the
detected outlier added) and looks for an additional outlier. This process is repeated until no additional outliers
are found. At this point, a backward deletion process is used to delete “insignificant” outliers (those whose
absolute t-statistics no longer exceed the critical value) from the model. This is done one at a time beginning
with the least significant outlier, until all outliers remaining in the model are significant. During backward
deletion the usual (non-robust) residual variance estimate is used, which can yield somewhat different outlier
t-statistics than those obtained during outlier detection.
The addall method follows the same general steps as the addone method, except that on each outlier
detection pass the addall method adds to the model all outliers with absolute t-statistics exceeding the critical
value. Typically several of the outliers added this way will be found to be insignificant when the new model is
estimated. The addall method thus depends heavily on the backward deletion process (much more than does
the addone method) to remove unnecessary outliers added to the model in the detection phase.
The differences between the addone and addall schemes can produce different final sets of detected outliers.
Two practical differences between the methods are worth noting. First, the addone method generally takes more
computation time than does addall. Second, the addall method may add so many outliers on a detection pass
that it exceeds the maximum number of regression variables allowed in a model. In this case the program prints
an error message to this effect and stops. Suggested remedies are to raise the cutoff value so fewer outliers are
detected, or to switch to the addone method, for which this phenomenon is much less likely.
For either method, the outlier t-statistics for all possible time points on each detection pass can be printed
by specifying print = iterations. This option generates considerable output.
Choosing the critical value requires both judgement and experience. Based on a simulation study involving
series of length up to 200 generated from low order nonseasonal ARIMA models, Chang, Tiao, and Chen (1988)
recommended critical values of 3 for high sensitivity in detection of AO outliers, 3.5 for medium sensitivity, and
4 for low sensitivity.
Outlier detection begins with the model specified by the regression and arima specs and with estimated
parameters. If the estimate spec is absent, the outlier spec forces estimation of the model (with default
estimation options) prior to outlier detection.
If outliers are suspected at specific known time points, then they may be included in the model by adding
the appropriate AO, TC, or LS regression variables to the model in the regression spec.
Outlier detection results can vary depending on the regARIMA model specified: observations are classified as
outliers because the model fits them less well than most of the other observations. Therefore a very inadequate
regARIMA model can yield inappropriate outlier adjustments.
If one or more temporary level shift t-statistics indicate that a run of 2 (or more) successive level shifts
cancels, a user-defined regressor can be used to capture the temporary level shift effect. In this way, two or
more level shifts can be replaced by one user-defined regressor. The usertype argument should be set to ls for
this regressor, so the user defined regressor is treated as a level shift. For more information on how to specify
user-defined regressors, see the regression spec documentation in Section 7.13.
Another technique can be used if the span of observations affected by the temporary level shift is small.
Individual AO outliers can be specified starting at the point of the first level shift and stopping with the point
before the final level shift.
EXAMPLES
Example 1 Simultaneously search for both AO and LS outliers over the entire time series, using the
addone method and a critical value that depends on the number of observations in the
interval searched for outliers (default options). If the number of level shifts present in the
model following outlier detection is two or more, compute t-statistics to test whether each
run of 2, . . . , 5 successive level shifts cancels to form a temporary level shift. Though the
estimate spec is absent, the presence of the outlier spec forces model estimation with
default estimation options.

data = (138 128 ... 297) }
arima { model = (0 1 1)(0 1 1)12 }
outlier { lsrun = 5 types=(ao ls) }
Example 2 Search only for AO outliers using the addall method and a critical value of t = 4.0. Because
the span argument is present in the series spec, only the time frame given there (January
7.11. OUTLIER 129
1980 through December 1992) is used in model estimation and in outlier detection. The two
level shifts specified in the regression spec are not tested for cancelation into a temporary
level shift since lsrun takes on its default value of 0.

Data = (138 128 ... 297)
Span = (1980.Jan, 1992.Dec) }
Regression { Variables = (LS1981.Jun LS1990.Nov) }
Arima { Model = (0 1 1)(0 1 1)12 }
Estimate { }
Outlier { Types = AO Method = Addall Critical = 4.0 }
Example 3 Estimate the model using the same span as in Example 2, but search only for LS outliers
in 1987 and 1988. The default addone method is used, but with a critical value of t = 3.0.
Each pair of successive LSs is tested for possible cancelation into a temporary LS.

data = (138 128 ... 297)
span = (1980.jan, 1992.dec) }
arima { model = (0 1 1)(0 1 1)12 }
estimate { }
outlier { types = ls
critical = 3.0
lsrun = 2
span = (1987.jan, 1988.dec) }
Example 4 Estimate the model using the same span as in Examples 2 and 3, but search for AO, TC
and LS outliers. The default addone method is used, but with a critical value of tAO = 3.0
for AO outliers, tLS = 4.5 for LS outliers, and tT C = 4.0 for TC outliers.

data = (138 128 ... 297)
span = (1980.jan, 1992.dec) }
arima { model = (0 1 1)(0 1 1)12 }
estimate { }
outlier { critical = (3.0, 4.5, 4.0)
types = all
}
7.12 PICKMDL
DESCRIPTION
Specifies that the ARIMA part of the regARIMA model will be sought using an automatic model selection
procedure similar to the one used by X-11-ARIMA/88 (see Dagum 1988). The user can specify which types
of models are to be fitted to the time series in the procedure and can change the thresholds for the selection
criteria.
USAGE
pickmdl { mode = both
method = best
file = "my.mdl"
fcstlim = 25.0
bcstlim = 25.0
qlim = 15.0
overdiff = 0.99
identity = all
outofsample = yes
print = (none autochoice)
savelog = automodel
}
ARGUMENTS
bcstlim Sets the acceptance threshold for the within-sample backcast error test when backcasts
are specified by setting mode=both. The absolute average percentage error of the back-
casted values is then tested against the threshold. For example, bcstlim=25 sets this
threshold to 25 percent. The value entered for this argument must not be less than zero,
or greater than 100. The default for bcstlim is 20 percent.
fcstlim Sets the acceptance threshold for the within-sample forecast error test. The absolute
average percentage error of the extrapolated values within the last three years of data
must be less than this value if a model is to be accepted by the pickmdl automatic
modeling selection procedure. For example, fcstlim=20 sets this threshold to 20 percent.
The value entered for this argument must not be less than zero, or greater than 100. The
default for fcstlim is 15 percent.
file Valid path and filename of the file containing the models used in the pickmdl automatic
model selection procedure. The models are specified using the same notation as in the
model argument of the arima spec; see DETAILS below. This argument must be
specified; there is no default.
identify Determines how automatic identification of outliers (via the outlier spec) and/or auto-
matic trading day regressor identification (via the aictest argument of the regression
7.12. PICKMDL 131
spec) are done within the pickmdl automatic model selection procedure. If identify
= all, automatic trading day regressor and/or automatic outlier identification (done in
that order if both are specified) are done for each model specified in the automatic model
file. If identify = first, automatic trading day regressor and/or automatic outlier
identification are done the first model specified in the automatic model file. The deci-
sions made for the first model specified are then used for the remaining models. The
identification procedures are redone for the selected model, if the model selected is not
the first. The default is identify = first.
method Specifies whether the pickmdl automatic model selection procedure will select the first
model which satisfies the model selection criteria (method = first) or the estimated
model with the lowest within-sample forecast error of all the model which satisfies the
model selection criteria (method = best). The default is method = first.
mode Specifies that the program will attempt to find a satisfactory model within the set of
candidate model types specified by the user, using the criteria developed by Statistics
Canada for the X-11-ARIMA program and documented in Dagum (1988); see DETAILS.
The fitted model chosen will be used to produce a year of forecasts if mode = fcst, or
will produce a year of forecasts and backcasts if mode = both. The default is mode =
fcst. The forecast spec can be used to override the number of forecasts and backcasts
used to extend the series. The model will be chosen from the types read in from a file
named in the file argument (specified above). Do not use both arima and pickmdl in
the same specification file.
outofsample Determines which kind of forecast error is used for pickmdl automatic model evaluation
and selection. If outofsample=yes, out-of-sample forecasts errors are used; these are
obtained by removing the data in the forecast period from the data set used to estimate
the model and to produce one year of forecasts (for each of the last three years of data). If
outofsample=no, within-sample forecasts errors are used. That is, the model parameter
estimates for the full series are used to generate forecasts for each of the last three years
of data. For conformity with X-11-ARIMA, outlier adjustments are made to the forecasted
data that have been identified as outliers. The default is outofsample=no.
overdiff Sets the threshold for the sum of the MA parameter estimates in the overdifferencing test.
The program computes the sum of the seasonal (for models with at least one seasonal
difference) or non-seasonal (for models with at least one non-seasonal difference) MA
parameter estimates. If the sum of the non-seasonal MA parameter estimates is greater
than the limit set here, the pickmdl automatic model selection procedure will reject the
model because of overdifferencing. If the sum of the seasonal MA parameter estimates
is greater than the limit set here, the pickmdl automatic model selection procedure will
print out a warning message suggesting the use of fixed seasonal effects in the regression
spec, but will not reject the model. The default for this argument is 0.9; values entered
for this argument should not be any lower than 0.9, and must not be greater than 1.
print The save option is not available for this spec. The tables available for output are listed
in Table 7.24; all tables are included in the default printout. For a complete listing of
the brief and default print levels for this spec, see Appendix B.
Table 7.24: Available Output Tables for Pickmdl

pickmdlchoice pmc model choice of pickmdl automatic model selection pro-
cedure
header hdr header for the pickmdl output
usermodels umd output for each model used in the pickmdl automatic
model selection procedure
qlim Sets the acceptance threshold for the p-value of the Ljung-Box Q-statistic for model
adequacy. The p-value associated with the fitted model’s Q must be greater than this
value for a model to be accepted by the pickmdl automatic model selection procedure.
For example, qlim = 10 sets this threshold to 10 percent. The value entered for this
argument must not be less than zero, or greater than 100. The default for qlim is 5
percent.
savelog Setting savelog=automodel or savelog=amd causes the result of the model selection
procedure to be output to the session log file (see section 2.6 for more information on the
log file).
DETAILS
The pickmdl spec cannot be used in the same spec file as the automdl or arima specs, or when the file
argument is specified in the estimate spec.
The default settings for the pickmdl automatic model selection procedure classify a model as acceptable
if (1) the absolute average percentage error of the extrapolated values within the last three years of data is
less than 15 percent, (2) the p-value associated with the fitted model’s Ljung-Box Q-statistic test of the lack of
correlation in the model’s residuals must be greater than 5 percent, and (3) there are no signs of overdifferencing.
There is an indication of overdifferencing if the sum of the non-seasonal MA parameter estimates (for models
with at least one non-seasonal difference) is greater than 0.9. No model is selected when none of the models
of the types in the model file is acceptable. Any of these criteria can be changed using the fcstlim, qlim, or
overdiff arguments.
Note that if there is a regression spec in the spec file, the regression terms specified there will be used with
all the ARIMA models evaluated by the automatic model selection procedure. The original series is transformed
as specified in the transform spec.
The X-11-ARIMA program developed by Statistics Canada uses the following model types in its automatic
modeling procedure:
7.12. PICKMDL 133
(0, 1, 1)(0, 1, 1)s

(0, 1, 2)(0, 1, 1)s
(2, 1, 0)(0, 1, 1)s
(0, 2, 2)(0, 1, 1)s
(2, 1, 2)(0, 1, 1)s
where s denotes the seasonal period (see Dagum 1988). These model types cannot be used if a fixed seasonal
effect is specified in the regression spec.
Each model type in the file designated by the file argument is listed on a separate line, with "X" at the end
of each line except the last.
Users can select one of the models to be a ”default” model by marking the end of the line with an asterisk
("*") rather than an "X". This will allow the program to use the default regARIMA model to generate
preadjustment factors based on the regressors specified by the user in the regression spec if a model is not
selected by the automatic modeling procedure. No forecasts (or backcasts) are generated if none of the models
are selected by the procedure.
An example using the X-11-ARIMA default models is given below:
(0 1 1)(0 1 1) *
(0 1 2)(0 1 1) X
(2 1 0)(0 1 1) X
(0 2 2)(0 1 1) X
(2 1 2)(0 1 1)
EXAMPLES
Example 1 Use the automatic ARIMA modeling procedure to select a model and use it to extend the
series with one year of forecasts. Trading day and stable seasonal regression effects are to
be included in the models. A default seasonal adjustment is to be performed.

data = (138 128 ... 297) }
regression { variables = (td seasonal) }
pickmdl { mode = fcst file = "nosdiff.mdl" }
estimate { }
x11 { }
The contents of nosdiff.mdl are given below:
(1 1 0) X
(2 1 0) X
(0 1 1) *
(0 1 2) X
(2 1 2)
Example 2 Similar to Example 1, except that the forecast acceptance threshold is changed to 20 percent,
the chi-square acceptance threshold is set to 10 percent, and the overdifferencing acceptance
threshold is changed to 0.99. Also, the first acceptable model will be selected, and automatic
outlier identification will be done for all the models listed in nosdiff.mdl.

data = (138 128 ... 297) }
pickmdl { mode = fcst file = "nosdiff.mdl"
method = first fcstlim = 20 qlim = 10
overdiff = 0.99 identify = all }
outlier { }
estimate { }
x11 { }
Example 3 The same as Example 1, except that out-of-sample forecast errors are used in the model
identification and selection process.

data = (138 128 ... 297) }
pickmdl { mode = fcst file = "nosdiff.mdl"
outofsample=yes }
estimate { }
x11 { }
7.13. REGRESSION 135
7.13 REGRESSION
DESCRIPTION
Specification for including regression variables in a regARIMA model, or for specifying regression variables
whose effects are to be removed by the identify spec to aid ARIMA model identification. Predefined regression
variables are selected with the variables argument. The available predefined variables provide regressors
modeling a constant effect, fixed seasonality, trading-day and holiday variation, additive outliers, level shifts,
and temporary changes or ramps. change-of-regime regression variables can be specified for seasonal and trading-
day regressors. User-defined regression variables can be added to the model with the user argument. Data
for any user-defined variables must be supplied, either in the data argument, or in a file specified by the file
argument (not both). The regression spec can contain both predefined and user-defined regression variables.
USAGE
regression { variables = (
const
seasonal or sincos[1, 2, 3]
td or tdnolpyear or tdstock[31] or
td1coef or td1nolpyear or tdstock1coef[31]
lom or loq
lpyear
easter[8] or sceaster[8] or easterstock[8]
labor[8] thank[1]
ao1967.apr ls1972.sep tc1979.sep
so1974.mar
rp1965.nov-1968.may tl1969.mar-1969.may
)
print = (none)
save = (rmx)
savelog = aictest
user = (cnybefore cnyafter IdulFitr strike)
usertype = (holiday holiday holiday2 ao)
start = 1995.jan
data = (25 0.1 · · · ) or file = "weather.dat"
format = "(2f5.1)"
aictest = ( easter user
td or tdnolpyear or tdstock or
td1coef or td1nolpyear or tdstock1coef
lom or loq or lpyear )
aicdiff = (2.0,,3.0,)
chi2test = yes
chi2testcv = 0.005
}
ARGUMENTS
aicdiff Defines the amount by which the AIC value (corrected for the length of the series, or
AICC) of the model with the regressor(s) specified in the aictest argument must fall
below the AICC of the model without these regressor(s) in order for the model with the
regressors to be chosen. The default value is aicdiff=0.0.
If only one value is given for this argument (aicdiff = 3.5), then this critical value is
used for all types of regressors. If a list of up to four values is given (aicdiff = (3.5,
4.0, 4.0, 5.5)), then the AIC difference for trading day regressors is set to the first
list entry (3.5 in this case), the AIC difference for length of month regressors is set to
the second list entry (4.0), the AIC difference for Easter regressors is set to the third
list entry (4.0), and the AIC difference for user-defined regressors is set to the fourth
list entry (5.5). A missing value, as in aicdiff = (3.25,,3.25,), is set to the default
critical value.
For more information on how this option is used in conjunction with the aictest argu-
ment, see DETAILS.
aictest Specifies that an AIC-based selection will be used to determine if a given set of regression
variables will be included with the regARIMA model specified. The only entries allowed
for this variable are td, tdnolpyear, tdstock, td1coef, td1nolpyear, tdstock1coef,
lom, loq, lpyear, easter, easterstock, and user. If a trading day model selection is
specified, for example, then AIC values (with a correction for the length of the series,
henceforth referred to as AICC) are derived for models with and without the specified
trading day variable. By default, the model with smaller AICC is used to generate
forecasts, identify outliers, etc. If more than one type of regressor is specified, the AIC-
tests are performed sequentially in this order: (a) trading day regressors, (b) length of
month / length of quarter / leap year regressors, (c) Easter regressors, (d) user-defined
regressors. If there are several variables of the same type (for example, several trading
day regressors), then the aictest procedure is applied to them as a group. That is, either
all variables of this type will be included in the final model or none. See DETAILS for
more information on the testing procedure. If this option is not specified, no automatic
AIC-based selection is performed.
chi2test Specifies that Chi-squared statistics will be be used to determine if groups of user-defined
holiday regressors will be kept in the regARIMA model. When chi2test = yes, Chi-
squared statistics will be generated for all user-defined holiday regression groups, and
those who which are not significant (at the level of the argument chi2testcv) are removed
from the regARIMA model. The default is chi2test = no, where no testing is done.
chi2testcv Sets the critical value used for the selection procedure in chi2test. The default is 0.01.
data Assigns data values to the user-defined regression variables. The time frame of the data
values must cover the time frame of the series (or of the span specified by the span
argument of the series spec, if present). It must also cover the time frame of forecasts
and backcasts requested in the forecast spec. The data values are read in free format.
The numerical values given in this argument are assigned in the order in which the user-
defined variables are named in the user argument. This assignment proceeds through all
the user-defined variables for the first time point, then through all the variables for the
second time point, etc. If the data argument is used, the file argument cannot be used.
file Name of the file containing data values for all user-defined regression variables. The
filename must be enclosed in quotes. If the file is not in the current directory, the path
must also be given. As with the data argument, the time frame of the data values must
cover both the series and any forecasts and backcasts requested. If the file argument is
used, the data argument cannot be used.
format Denotes the format used when reading the data for the regression variables from the file
named in the file argument. Four types of input are accepted:
a. free format, in which all numbers on a line will be read before continuing to the next
line, and the numbers must be separated by one or more spaces (not by commas or
tabs) (example: format="free");
b. a valid Fortran format, which must be enclosed in quotes and must include the
initial and terminal parentheses (example: format="(6f12.0)");
c. “datevalue” format, in which the year, month or quarter, and the associated value
for each of the user-defined regression variables for a given observation are given in
this order in free format on individual lines in the data file. Thus, a line of the data
file with three regressors having the values 0, 0, and 1 respectively for July of 1991
would have the form 1991 7 0 0 1. All the user-defined regressors must be on the
same record, and in the order of their appearance in the user argument (example:
format= "datevalue");
d. the “x12save” format X-13ARIMA-SEATS uses to save a table. This allows the user
to read in a file saved from a previous X-13ARIMA-SEATS run (example: format =
"x12save");
e. a variant of “free” format where the numbers must be separated by one or more
spaces (not by commas or tabs), and decimal points are expressed as commas (a
convention in some European countries). (example: format="freecomma");
f. a variant of “datevalue” format, where the year, month or quarter, and value of
each observation are found in this order in free format on individual lines, where
decimal points are expressed as commas. Thus, a line of the data file containing the
value 1355.34 for July of 1991 would have the form 1991 7 1355,34. The number
of preceding blanks can vary (example: format="datevaluecomma").
If no format argument is given the data will be read in free format. Format cannot be
used with the data argument, only with file.
print and save Table 7.25 gives the available output tables for this spec. All these tables are included
in the default printout, except regressionmatrix and dailyweights. Also, if there is
only one type of outlier in the regARIMA model, then only the combined outlier table
will print out, and the specific tables for the individual outlier effect (ao, ls, tc, so) will
be suppressed. For a complete listing of the brief and default print levels for this spec,
see Appendix B.
7.26.
Table 7.25: Available Output Tables for Regression
itname short save? description of table

regressionmatrix rmx + values of regression variables with associated
dates
aictest ats · output from AIC-based test(s) for trading day,
Easter, and user-defined regression variables
outlier otl + combined regARIMA outlier factors (table A8)
aoutlier ao + regARIMA additive (or point) outlier factors
(table A8.AO)
levelshift ls + regARIMA level change, temporary level
change and ramp outlier factors (table A8.LS)
seasonaloutlier so + regARIMA seasonal outlier factors (table
A8.SO)
transitory a13 + regARIMA transitory component factors from
user-defined regressors (table A13)
temporarychange tc + regARIMA temporary change outlier factors
(table A8.TC)
tradingday td + regARIMA trading day factors (table A6)
holiday hol + regARIMA holiday factors (table A7)
regseasonal a10 + regARIMA user-defined seasonal factors (table
A10)
userdef usr + factors from user-defined regression variables
(table A9)
chi2test cts · output from chi-squared based test for groups
of user-defined regression variables
dailyweights tdw · Daily weights from trading day regressors, nor-
malized to sum to seven
Table 7.26: Available Log File Diagnostics for Regression

aictest ats Test results from the AICC-based regressor selection procedure
chi2test cts Test results from the Chi-squared based regressor selection pro-
cedure
start The start date for the data values for the user-defined regression variables. The default
is the start date of the series. Valid values are any date up to the start date of the series
(or up to the start date of the span specified by the span argument of the series spec,
if present).
user Specifies names for any user-defined regression variables. Names are required for all user-
defined variables to be included in the model. The names given are used to label estimated
coefficients in the program’s output. Data values for the user-defined variables must be
supplied, using either the data or file argument (not both). The maximum number of
user-defined regression variables is 52. (This limit can be changed—see Section 2.8.)
usertype Assigns a type of model-estimated regression effect to each user-defined regression vari-
able. It causes the variable and its estimated effects to be used and be output in the
same way as a predefined regressor of the same type. This option is useful when trying
out alternatives to the regression effects provided by the program.
The type of the user-defined regression effects can be defined as a constant (constant),
seasonal (seasonal), trading day (td), stock trading day (tdstock), length-of-month
(lom), length-of-quarter (loq), leap year (lpyear), moving holiday (easter, the US hol-
idays thanks and labor), outlier (ao, ls, rp, so or tc), a user-defined transitory compo-
nent for SEATS (transitory) or other user-defined (user) regression effects. In addition
to these types, users can specify up to 5 different user-defined holidays (holiday, holi-
day2, holiday3, holiday4, and holiday5). This gives the user flexibility in specifying
more than one holiday, and the chi-squared statistic is generated separately for these
user-defined holidays.
One effect type can be specified for all the user-defined regression variables defined in the
regression spec (usertype=td), or each user-defined regression variable can be given its
own type (usertype=(td td td td td td holiday user)). Once a type other than
user has been assigned to a user-defined variable, further specifications for the variable
in other arguments, such as aictest or noapply, must use this type designation, not
user. If this option is not specified, all user-defined variables have the type user. See
DETAILS for more information on assigning types to user-defined regressors.
variables List of predefined regression variables to be included in the model. Data values for
these variables are calculated by the program, mostly as functions of the calendar. See
DETAILS for a discussion and a table of the available predefined variables. Also see
Section 4.3 for additional information and a table defining the actual regression variables
used.

b Specifies initial values for regression parameters in the order that they appear in the
variables and user arguments. If present, the b argument must assign initial values
to all regression coefficients in the regARIMA model, and must appear in the spec file
after the variables and user arguments. Initial values are assigned to parameters
either by specifying the value in the argument list or by explicitly indicating that it is
missing as in the example below. Missing values take on their default value of 0.1. For
example, for a model with two regressors, b=(0.7, ) is equivalent to b=(0.7,0.1), but
b=(0.7) is not allowed. For a model with three regressors, b=(0.8,,-0.4) is equivalent
to b=(0.8,0.1,-0.4). To hold a parameter fixed at a specified value, immediately follow
the value in the b list with an ‘f’, e.g., b=(0.7f, 0.1).
centeruser Specifies the removal of the (sample) mean or the seasonal means from the user-defined
regression variables. If centeruser=mean, the mean of each user-defined regressor is
subtracted from the regressor. If centeruser=seasonal, means for each calendar month
(or quarter) are subtracted from each of the user-defined regressors. If this option is
not specified, the user-defined regressors are assumed to already be in an appropriately
centered form and are not modified.
eastermeans Specifies whether long term (400 year) monthly means are used to remove seasonality from
the Easter regressor associated with the variable easter[w], as described in footnote 5
of Table 4.1 (eastermeans=yes), or, instead, monthly means calculated from the span of
data used for the calculation of the coefficients of the Easter regressors (eastermeans=no).
The default is eastermeans=yes. This argument is ignored if no built-in Easter regressor
is included in the regression model, or if the only Easter regressor is sceaster[w] (see
DETAILS).
noapply List of the types of regression effects defined in the regression spec whose model-
estimated values are not to be removed from the original series before the seasonal
adjustment calculations specified by the x11 spec are performed.
Applicable types are all modelled trading day effects (td), Easter, Labor Day, and
Thanksgiving-Christmas holiday effects (holiday), point outliers (ao), level changes and
ramps (ls), temporary changes (tc), seasonal outliers (so), user-defined seasonal regres-
sion effects (userseasonal), and the set of user-defined regression effects (user).
tcrate Defines the rate of decay for the temporary change outlier regressor. This value must be
a number greater than zero and less than one. The default value is tcrate=0.7 ** (12
/ period), where period is the number of observations in one year (for monthly time
series, 4 for quarterly time series. This formula for the default value of tcrate ensures the
same rate of decay over an entire year for series of different periodicity. If the frequency
of the time series is less than 4 (ie, period < 4), then there is no default value, and the
user will have to enter a value of tcrate if a temporary change outlier was specified in
the variables argument.
DETAILS
If forecasting is performed, X-13ARIMA-SEATS creates data values for the selected predefined regression variables
for the entire forecast period. If there are any user-defined regression variables, then data values must also be
supplied for them for the entire forecast period (similarly for the backcasts). In addition to the limit of 52 user-
defined regression variables, there is an overall limit of 80 regression variables in the model. (These limits can
be changed—see Section 2.8.) The latter limit is on the total number of predefined and user-defined regression
variables plus the number of regression variables added automatically by the outlier spec. The maximum length
of the series of user-defined regression variables, not including the forecast period, is 780. (This limit can also
be changed—see Section 2.8.)
Table 7.27: Predefined regression variables
Variable Description
const Trend constant regression variable to allow for a nonzero overall mean for the differenced data.
seasonal Fixed seasonal effects parameterized via s−1 seasonal contrast variables (s = seasonal period).
The resulting variables allow for month-to-month (or quarter-to-quarter, etc.) differences in
level, but have no net effect on overall level. Seasonal cannot be used with sincos and also
not in models with seasonal differencing except as a partial change of regime variable (see
DETAILS where additional change of regime options are described, as in Table 7.28).
sincos[ ] Fixed seasonal effects (for s = seasonal period) parameterized via trigonometric regression
variables of the form sin(ωj t) and cos(ωj t) at seasonal frequencies ωj = (2πj/s) for 1 ≤ j ≤ s/2
(dropping sin(ωj t) ≡ 0 for j = s/2 for s even). Each frequency to be included must be
specified, i.e., for monthly series sincos[1, 2, 3, 4, 5, 6] includes all seasonal frequencies
while sincos[1, 2, 3] includes only the first three. Sincos[ ] cannot be used with seasonal
or in models with seasonal differencing.
td Estimate monthly (or quarterly) flow trading-day effects by including the tdnolpyear vari-
ables (see below) in the model, and by handling leap-year effects either by re-scaling (for
transformed series) or by including the lpyear regression variable (for untransformed series).
Td can only be used for monthly or quarterly series, and cannot be used with tdnolpyear,
td1coef, td1nolpyear, lpyear, lom, loq, tdstock, or tdstock1coef. If td is specified, do not
specify adjust = lpyear or adjust = lom (adjust = loq) in the transform spec. Several
change of regime options are described in DETAILS, as in Table 7.28.
tdnolpyear Include the six day-of-week contrast variables (monthly and quarterly flow series only): (no. of
Mondays) − (no. of Sundays), . . . , (no. of Saturdays) − (no. of Sundays). Tdnolpyear cannot
be used with td, td1coef, td1nolpyear, tdstock, or tdstock1coef. Several change of regime
options are described in DETAILS, as in Table 7.28.
td1coef Estimate monthly (or quarterly) flow trading-day effects by including the td1nolpyear variable
(see below) in the model, and by handling leap-year effects either by re-scaling (for transformed
series) or by including the lpyear regression variable (for untransformed series). Td1coef
can only be used for monthly or quarterly series, and cannot be used with td, tdnolpyear,
td1nolpyear, lpyear, lom, loq, tdstock, or tdstock1coef. If td1coef is specified, do not
specify adjust = lpyear or adjust = lom (adjust = loq) in the transform spec. Several
td1nolpyear Include the weekday-weekend contrast variable (monthly and quarterly flow series only):
(no. of weekdays) − 52 (no. of Saturdays and Sundays). Td1nolpyear cannot be used with
td, tdnolpyear, td1coef, tdstock, or tdstock1coef. Several change of regime options are
described in DETAILS, as in Table 7.28.
lpyear Include a contrast variable for leap-year (monthly and quarterly flow series only): 0.75 for
leap-year Februaries (first quarters), -0.25 for non-leap year Februaries (first quarters), 0.0
otherwise. Lpyear cannot be used with td, td1coef, tdstock, or tdstock1coef. Several
lom Include length-of-month as a regression variable. If lom is requested for a quarterly series,
X-13ARIMA-SEATS uses loq instead. Requesting lom when s 6= 12 or 4 results in an error. Lom
cannot be used with td, td1coef, tdstock, or tdstock1coef. Several change of regime options
are described in DETAILS, as in Table 7.28.
loq Include length-of-quarter as a regression variable. If loq is requested for a monthly series,
X-13ARIMA-SEATS uses lom instead. The same restrictions that apply to lom apply to loq.
Several change of regime options are described in DETAILS, as in Table 7.28.
Table 7.27: Predefined regression variables (continued)
tdstock[w] Estimate day-of-week effects for inventories or other stocks reported for the wth day of each
month. The value w must be supplied and can range from 1 to 31. For any month of length
less than the specified w, the tdstock variables are measured as of the end of the month. Use
tdstock[31] for end-of-month stock series. Tdstock can be used only with monthly series and
cannot be used with tdstock1coef, td, tdnolpyear, td1coef, td1nolpyear, lom, or loq.
tdstock1coef[w] Estimate a constrained stock trading day effect for inventories or other stocks reported for the
wth day of each month. The value w must be supplied and can range from 1 to 31. For any
month of length less than the specified w, the tdstock1coef variables are measured as of the
end of the month. Use tdstock1coef[31] for end-of-month stock series. Tdstock1coef can
be used only with monthly series and cannot be used with tdstock, td, tdnolpyear, td1coef,
td1nolpyear, lom, or loq.
easter[w] Easter holiday regression variable for monthly or quarterly flow data which assumes the level
of daily activity changes on the w−th day before Easter and remains at the new level until the
day before Easter. The value w must be supplied and can range from 1 to 25. To estimate
complex effects, several of these variables, differing in their choices of w, can be specified.
labor[w] Labor Day holiday regression variable (monthly flow data only) that assumes the level of daily
activity changes on the w−th day before Labor Day and remains at the new level until the day
before Labor Day. The value w must be supplied and can range from 1 to 25.
thank[w] Thanksgiving holiday regression variable (monthly flow data only) that assumes the level of
daily activity changes on the w−th day before or after Thanksgiving and remains at the new
level until December 24. The value w must be supplied and can range from −8 to 17. Values
of w < 0 indicate a number of days after Thanksgiving; values of w > 0 indicate a number of
days before Thanksgiving.
sceaster[w] Statistics Canada Easter holiday regression variable (monthly or quarterly flow data only)
assumes that the level of daily activity changes on the (w − 1)−th day and remains at the
new level through Easter day. The value w must be supplied and can range from 1 to 24.
To estimate complex effects, several of these variables, differing in their choices of w, can be
specified.
easterstock[w] End of month stock Easter holiday regression variable for monthly or quarterly stock data.
This regressor is generated from the easter[w] regressors. The value w must be supplied and
can range from 1 to 25. To estimate complex effects, several of these variables, differing in
their choices of w, can be specified.
aodate Additive (point) outlier variable, AO, for the given date or observation number. For series
with associated dates, AOs are specified as aodate. For monthly series this is aoyear.month
(e.g., ao1985.jul or ao1985.7), for quarterly series this is aoyear.quarter (e.g., ao1985.1 for
an AO in the first quarter of 1985), and for annual series this is aoyear (e.g., ao1922). For
series without associated dates, AOs are specified as aoobservation number, e.g., ao50 for an
AO at observation 50. More than one AO may be specified. All specified outlier dates must
occur within the series. (AOs with dates within the series but outside the span specified by
the span argument of the series spec are ignored. )
lsdate Regression variable for a constant level shift (in the transformed series) beginning on the given
date, e.g., ls1990.oct for a level shift beginning in October 1990. More than one level shift
may be specified. Dates are specified as for AOs and the same restrictions apply with one
addition: level shifts cannot be specified to occur on the start date of the series (or of the span
specified by the span argument of the series spec).
Table 7.27: Predefined regression variables (continued)
tcdate Regression variable for a temporary level change (in the transformed series) beginning on the
given date, e.g., tc1990.oct for a temporary change beginning in October 1990. More than one
temporary change may be specified. Dates are specified as for AOs, and the same restrictions
apply.
sodate Regression variable for a seasonal outlier (in the transformed series) beginning on the given
date, e.g., so1990.oct for a seasonal outlier beginning in October 1990. More than one tempo-
rary change may be specified. Dates are specified as for AOs, and the same restrictions apply
with one addition: seasonal level shifts cannot be specified to occur on the start date of the
series (or of the span specified by the span argument of the series spec).
rpdate-date Ramp effect which begins and ends on the given dates, e.g., rp1988.apr-1990.oct. More than
one ramp effect may be specified. All dates of the ramps must occur within the series. (Ramps
specified within the series but with both start and end dates outside the span specified by the
span argument of the series spec are ignored.) Ramps can overlap other ramps, TLs, AOs,
and level shifts.
tldate-date Temporary level change effect which begins and ends on the given dates, e.g.,
tl1988.apr-1990.oct. More than one temporary level shift effect may be specified. All dates
of the temporary level shift regressor must occur within the series. (Temporary level shifts
specified within the series but with start or end dates outside the span specified by the span
argument of the series spec are ignored.) Temporary level shifts can overlap other TLs, ramps,
AOs, and level shifts.
If const is specified in the variables argument, then the resulting regression variable allows for an a constant
term in the series resulting from any differencing operations in the ARIMA model. If the ARIMA model involves
no differencing, this is simply the usual regression constant term for a nonzero overall mean; if the ARIMA
model does involve differencing, this regressor is called a trend constant. In the latter case the actual regression
variable created is defined such that, after differencing, it yields a column of ones. See Section 4.3 for discussion.
We generally recommend specifying td in the variables argument when trading-day effects are thought
to be present in a monthly flow time series, that is, a series whose values are monthly accumulations of daily
values. In this case, how the program handles leap-year effects depends on information from the transform
spec. If the series is transformed (Box-Cox or logistic transformation) then leap-year effects are removed by
prior adjustment: the series is divided before transformation by a set of factors lpt where lpt = 28.25/29 if t is
a leap year February, lpt = 28.25/28 if t is a non-leap year February, and lpt = 1.00 otherwise.
If the series is not transformed, then the leap-year regression variable lpyear is included in the model. Its
values, denoted by LPt are given by LPt = 29 − 28.25 if t is a leap year February, LPt = 28 − 28.25 if t is a
non-leap year February, and LPt = 0.00 otherwise. In both cases, the tdnolpyear regression variables, (no. of
Mondays) − (no. of Sundays), . . . , (no. of Saturdays) − (no. of Sundays), are also included in the model.Leap
year effects are the nonseasonal component of length-of-month effects. When type = trend is used in the x11
spec, with the result that there is no seasonal effect estimation and adjustment, then td handles length-of-month
effects instead of leap-year effects. That is, with a transformation, there is prior adjustment by the length-of-
month factors described in Table 7.27, and with no transformation, the lom regressor, whose value is the number
of days in the month, is added to the regression with the tdnolpyear regressors.
In any situation in which the user prefers to model length-of-month effects in a transformed series, the leap
year regressor is the nonseasonal component for the length-of-month (quarter) regressor. If the user prefers
to model length-of-month effects in a transformed series through the lom regression variable, this can be done
by specifying both lom and tdnolpyear, i.e., variables = (lom tdnolpyear ...). If the user prefers to
prior adjust an untransformed series for length-of-month effects, this can be done by specifying variables =
(tdnolpyear ...) in the regression spec and adjust = lom in the transform spec.
If adjust=lom is specified in the transform spec, then including either td or lom in the variables list
leads to a conflict. The conflict occurs either because two requests have been made to re-scale the series by
dividing by length of month, or because both a length-of-month rescaling and the lom regression variable have
been requested (which will generally lead to a singular system of equations for the regression coefficients). In
this case, the user should either (i) remove adjust=lom from the transform spec, or (ii) in the variables list,
replace td by tdnolpyear, or drop lom.
For quarterly flow time series the same trading-day options are available, and the above comments apply
with lom replaced by loq.
The values lom and loq are equivalent —if either is specified, the seasonal period specified in the series
spec determines which is used. Thus, period = 12 implies lom and period = 4 implies loq. Also, note that
lom or loq can be specified without tdnolpyear. This could be done to account for fixed seasonality due
to length-of-month (or length-of-quarter) effects for a series with no day-of-week specific effects. Predefined
length-of-period variables are available only for monthly or quarterly flow series.
For stock series, such as inventories, the program can estimate trading-day effects only for monthly series.
Tdstock[w], where w can range from 1 to 31, creates six regression variables contrasting six days of the week
with the seventh - see Section 4.3. The value w must be specified; it denotes the day of the month for which
the stock is reported or the last day of the month, whichever is smaller. Therefore, tdstock[31] is used for
end-of-month stocks.
The holiday effect regression variables (for Easter, Labor Day, and Thanksgiving) are for flow series. The
Easter variable can be specified for either monthly or quarterly series. The Labor Day and Thanksgiving
variables are only for monthly series.
If a series is designated as a stock or a flow series by using the type argument of the series or composite
spec, then trading day and Easter regressors specified in variables argument need to agree with this type - one
cannot specify stock trading day regressors for a flow series. If a series type is not specified, then any trading
day or holiday regressor may be used with the series.
Change-of-regime regression variables can be specified for seasonal (seasonal), trigonometric seasonal
(sincos), trading day (td, tdnolpyear, tdstock, td1coef, td1nolpyer, or tdstock1coef), leap year
(lpyear), length-of-month (lom), and length-of-quarter (loq) regression variables. Two types of change-of-
regime regressors are available: full and partial.
As Table 7.28 shows, change of regime regressors are specified by appending the change date, surrounded
by one or two slashes, to the name of a regression variable in the variables argument of the regression
spec. The date specified for the change of regime divides the series being modelled into two spans, an early
span containing the data for times prior to the change date and a late span containing the data from on and
after this date. Partial change of regime variables are restricted to one of these two spans, being zero in the
complementary span. The full change of regime variables estimate both the basic regression of interest and
the partial change of regime regression for the early span. For example, the full change of regime specification
variables = (td/1990.jan/)is equivalent to the specification variables = (td td/1990.jan//). It causes
Table 7.28: Change of Regime Regressor Types and Syntax
Type Syntax Example

Full change of regime regressor reg/date/ td/1990.jan/
Partial change of regime regressor, zero before change date reg//date/ td//1990.jan/
Partial change of regime regressor, zero on and after change date reg/date// td/1990.jan//
the program to output the coefficients estimated for td and for td/1990.jan// along with trading day factors
for their combined effects.
The coefficients resulting from use of a full change of regime regression have convenient interpretations: Let
the basic regressors be denoted by Xjt , and let t0 be the change point. Then the partial change of regime
regressors for the early regime are
(
E Xjt for t < t0
Xjt =
0 for t ≥ t0
L E
and those for the late regime can be calculated as Xjt = Xjt − Xjt . For the data transformed as indicated in
the transform spec, the effect estimated by the full change of regime regression has the form
X X X X
E L E
aj Xjt + bj Xjt = aj Xjt + (aj + bj )Xjt .
j j j j
From the right-hand-side formula, we observe that the coefficients aj of the basic regressors Xjt can be
L E
interpreted at the coefficients of the late-span regressors Xjt , and the coefficients bj of the Xjt can be interpreted
as measuring the change in the coefficients of the late-span regressors required to obtain coefficients for the early-
span effects. Therefore, statistically significant bj indicate the nature of the change of regime.
We illustrate two other natural uses for partial change of regime variables. First, the specification variables
= (td//1990.jan/) can be used to estimate the trading day component of a series that has no statistically
significant trading day effects prior to 1990, but possibly significant effects beginning in this year. Second, when
an ARIMA model with seasonal differencing is specified in the arima spec, or in the models estimated by the
automdl spec, then the specification variables = (seasonal//1990.jan/) can be used to estimate a fixed
change in a somewhat variable seasonal pattern that takes place in January of 1990 and to test for the statistical
significance of the estimated change.
The effect of the argument aictest can be to delete a regressor set named in the variables list from this
list, or to add a regressor set to the variables list. The effect of a nonzero (positive) value of aicdiff is to make
it more difficult for the aictest procedure to include in the model the variable being tested. Let ∆AICC denote
the value associated with the aicdiff argument, which by default is zero. Let AICC with (AICC without ) denote
the AICC value of the model with (without) a set of regressors specified in the aictest argument. If this set is
not named in the variables list, it will be added to the regression model if
AICC with + ∆AIC < AICC without .

If this set is named in the variables list, it will be retained in the regARIMA model only if this inequality
holds.
In the second case, if aictest = (tdstock), it is the end-of-month stock variables, specified by tdstock[31],
are the variables added, because 31 is the default value for w in tdstock[w].
There are more possibilities if aictest = (easter) and no Easter effect regressors appear in the variables
list. Then three additional models are considered, the three models obtained by augmenting the specified reg-
ARIMA model with the regressor easter[w] for w = 1, 8, 15 respectively. The Easter regressor whose model has
the smallest AICC is retained if its AICC is smaller than the model with no Easter regressors by at least the
amount ∆AIC = 0; otherwise, the model without Easter regressors is selected.
Simulation experiments we have conducted suggest that AICC does not distinguish with high reliability
between easter[w] regressors whose w values differ by less than seven. The out-of-sample forecast diagnostics
produced by the history spec can sometimes distinguish between such regressors by showing that one provides
persistently more accurate forecasts, and therefore presumably better describes the Easter effect in the data.
Similar to the case for Easter, when aictest = (td) and no trading day regressors appear in the variables
list, then additional models are considered. These are the models obtained by augementing the specified reg-
ARIMA model with full and one coefficient trading regressors, depending on the type of series (td and td1coef
for flow series, tdstock and tdstock1coef for stock series).
The trading day regressor whose model has the smallest AICC is retained if its AICC is smaller than the
model with no trading day regressors by at least the amount ∆AIC = 0; otherwise, the model without tradind
day regressors is selected.
When regressors appear in both the aictest and variables arguments, the type of regressors specified should
be identical. An exception for this is for trading day regressors. The entry aictest = td serves as a correct
entry for any type of flow or stock trading day regressor. The sample day for stock trading day variables and
the date specified for change-of-regime regressors should not be included in the aictest argument; its value will
be taken from the entry in the variables argument. For example, if variables=(tdstock[15] ao1995.jan),
then the entry for aictest can be tdstock or td.
Another exception is for Easter regressors. The entry aictest = easter serves as a correct entry for any
type of flow or stock Easter regressor. The window length of the Easter regressor should not be included
in the aictest argument; they will be assumed from the entry in the variables argument. For example, if
variables=(easterstock[15] ao1995.jan), then the entry for aictest can be easterstock or easter.
Note that this is not affected by setting type = stock or type = flow in the series or composite specs;
the entries aictest = td and aictest = easter can still be used for both stock and flow series. However, you
cannot set type = flow in the series spec and aictest = tdstock.
Regressors specified by the aictest argument must also be able to be included with other regressors specified
either in the variables and the aictest arguments. For example, the following regression spec is incorrectly
specified, as the td and lom arguments cannot be specified together in the variables argument:
regression{
variables = td
aictest = lom
}
Using tdnolpyear instead would allow a model with the 6 trading day regressors and the length-of-month
regressors.
In addition, users should not specify aictest = lom for series that are not monthly series, and aictest =
loq for series that are not quarterly.
As mentioned above, trading day regressors are always tested before length-of-month (-quarter) or leap year
regressors. If options specified in the regression spec lead to trading day and leap year regressors in the same
regARIMA model, then the program will test the trading day and leap year regressors together if aictest =
td is specified, or will test the sets of regressors separately if aictest = (td lpyear) is specified.
User-defined variables should be input to the program in deseasonalized form (unless they are seasonal re-
gressors). The deseasonalization method described in Section C.1.3 is likely to be the appropriate one, because
regressors are additive components of the regARIMA model. If deseasonalization is not done, then the seasonal
factors will not include all estimated seasonal effects. Another problem is that regressors with seasonal compo-
nents are likely to have estimated coefficients, and estimated effects, that are more correlated with one another
and therefore more difficult to interpret.
If a type is assigned in to a user-defined variable with the usertype argument, the factor derived from
the user-defined regression variables of that type will be combined with the regression factor from variables
of the same type specified in the regression spec. The resulting factor will be adjusted out of the series for
the seasonal adjustment factor calculations determined by the x11 spec unless the type name appears in the
noapply argument.
Setting usertype=seasonal will cause seasonal factors to be created from the user-defined regressors that will
be adjusted out of the original series before the seasonal adjustment specified by the x11 spec is calculated. Com-
bined seasonal factors are created from the X-11 and regression factors. In addition, if noapply=userseasonal
is specified, the user-defined seasonal regressors are treated exactly like seasonal regressors specified in the
variables argument: the seasonal effect estimated from these regressors will not adjusted out of the series
prior to seasonal adjustment. The effects estimated by Table 7.27 seasonal regressors specified in the variables
argument are not available as output. If it is desired to remove these effects from the series prior to seasonal
adjustment, this can be done in by setting save=rmx to save the regressors in an output file. From this file,
the regressors can be input to the program as user-defined regressors with usertype=seasonal to achieve the
desired removal.
Note that if format = "datevalue" or format = "x12save", the starting date of the user defined regres-
sor(s) is automatically read from the data file. Therefore, the starting date need not be specified with the start
argument of the regression spec.
Trading day and/or holiday regressors may not be specified simultaneously in the regression and x11regression
specs unless the noapply option is used to specify that the effects estimated by either the regression or
x11regression spec not be used to adjust the series.
The two choices for the argument eastermeans yield noticeably different holiday factors. But the choice has
no effect on forecasts (provided the regARIMA model used includes seasonal differencing or the fixed seasonal
regressors) and usually has only negligible effects on the combined seasonal and holiday factors, because the
seasonal factors change to compensate for the differences between the choices.
Table 7.29 give the monthly means for February, March, and April that are used to obtain deseasonalized
Easter regressors under eastermeans = yes; the means for other months are zero. These calendar means
Easter effect February March April

length (w) Mean Mean Mean
1 0.0000 0.2660 0.7340
2 0.0000 0.2810 0.7190
3 0.0000 0.2967 0.7033
4 0.0000 0.3125 0.6875
5 0.0000 0.3304 0.6696
6 0.0000 0.3483 0.6517
7 0.0000 0.3654 0.6346
8 0.0000 0.3820 0.6180
9 0.0000 0.3976 0.6024
10 0.0000 0.4136 0.5864
11 0.0000 0.4305 0.5695
12 0.0000 0.4477 0.5523
13 0.0000 0.4643 0.5357
14 0.0000 0.4807 0.5193
15 0.0000 0.4973 0.5027
16 0.0000 0.5146 0.4854
17 0.0000 0.5319 0.4681
18 0.0000 0.5490 0.4510
19 0.0000 0.5661 0.4339
20 0.0000 0.5830 0.4170
21 0.0000 0.5999 0.4001
22 0.0003 0.6163 0.3835
23 0.0011 0.6311 0.3677
24 0.0021 0.6451 0.3528
25 0.0037 0.6576 0.3387
Table 7.29: 500 Year (1600-2099) means for Easter regressors of different window length w.
were generated from frequencies of the date of Easter for a 500 year period (1600-2099). These frequencies
were computed from dates given in Bednarek (2007) which were checked using information from Montes (2001,
1997b, 1997a); the algorithm used by Montes to compute the date of Easter for the Gregorian calendar is given
in Duffet-Smith (1981).
For quarterly series, the mean of the first quarter is equivalent to the sum of the February and March means
from Table 7.29, the mean for the second quarter is equivalent to the April mean, and the means for other
quarters are zero.
For a nonseasonal time series, an adjustment for trading day and holiday effects estimated by means of this
spec can be obtained by setting type=trend in the x11 spec.
When the b=() argument is used to fix coefficients, AIC and the other model selection statistics may become
invalid, see the DETAILS section of estimate.
For more information concerning the modeling of holiday effects and the detection and modeling of trading
day effects, see Findley and Soukup (2000) and Lin and Liu (2002).
EXAMPLES
Example 1 Estimate a model with ARIMA (0 1 1) errors, fixed seasonal effects, and a trend constant.
SERIES { TITLE = "Monthly sales" START = 1976.JAN

DATA = (138 128 ... 297) }
REGRESSION { VARIABLES = (CONST SEASONAL) }
ARIMA { MODEL = (0 1 1) }
ESTIMATE { }
Example 2 Specify a model to fit sine and cosine variables with the 4th and 5th seasonal frequency
by ordinary least squares to the final irregular component of a series to test if ”visually
significant” spectrum peaks at these frequencies are statistically significant.
series { title = "Irregular Component of Monthly Sales"

start = 1976.jan
file = "sales.d13"
format = "x12save"
}
regression { variables = (const sincos[4,5]) }
estimate { }
spectrum { savelog=peaks }
Example 3 Specify regression variables for trading-day, Easter, Labor Day, and Thanksgiving effects in a
monthly time series. The duration in number of days is specified for each holiday effect. Since
td is specified and the series is log transformed, the original series (before transformation)
is divided by the leap-year factors, and the tdnolpyear regression variables are fit to the
transformed series. The regression coefficients are estimated by the identify spec through
a regression of the maximally differenced series (after transformation and length-of-month
adjustment) on the correspondingly differenced regression variables. The identify spec
then produces various sample ACFs and PACFs (of the regression residuals) to be used for
identifying an ARIMA model for the regression errors.

Data = (138 128 ... 297) }
Regression { Variables = (TD Easter[8] Labor[10] Thank[3]) }
Identify { Diff = (0 1) SDiff = (0 1) }
Example 4 Estimate a model including the same regressors as in Example 3, and also the lom regression
variable in place of the division of the series by standard leap-year effects that the argument
value td invokes. (Replacing the value of td with tdnolpyear prevents the division by
the standard leap year effects.) Perform a test (using AICC) of the significance of the
trading-day and Easter regressors. Note that the program will test the significance of the
6 trading-day regressors first, then the significance of the length-of-month regressor, and
finally the significance of the Easter regressor. An ARIMA (0 1 1)(0 1 1)12 model is used
for the regression error series.

data = (138 128 ... 297) }
regression { variables = (tdnolpyear lom easter[8] labor[10] thank[3])
aictest = (lom td easter) }
arima { model = (0 1 1)(0 1 1) }
estimate { }
Example 5 Specified regression variables are a one coefficient stock trading day regressor and a end of
month stock Easter regressor. Since the sample day specified in the trading day regressor is
31, it is an end of month stock regressor as well. Decide (using AICC) if the stock trading-day
and Easter regressors should be kept in the model.
series { title = "Retail inventory of food products"

start = 1990.jan data = "foodri.dat"
type = stock
}
regression { variables = (
tdstock1coef[31]
easterstock[8]
)
aictest = ( td easter )
}
arima { model = (0 1 1)(0 1 1) }
x11 { }
Example 6 Estimate a model with trading-day effects, two AOs, and two LSs for a quarterly seasonal
series. Accounting for these effects, the transformed series follows an ARIMA (0 1 1)(0 1 1)4
model.
Series { Title = "Quarterly Sales" Start = 1963.1 Period = 4

Data = (1039 1241 ... 2210) }
Regression { Variables = (AO1967.1 LS1985.3 LS1987.2 AO1978.1 TD) }
Arima { Model = (0 1 1)(0 1 1) }
Estimate { }
Example 7 Estimate a user-defined regression variable for a temporary level-shift from the third quarter
of 1985 through the first quarter of 1987. The effect of the temporary level shift is removed
through the regression performed by the identify spec, prior to the computation of ACFs
and PACFs for identification of the ARIMA part of the model.
series {title = "Quarterly sales" start = 1981.1

data = (301 294 ... 391) period = 4 }
regression {user = tls
data = (0 0 0 0 0 0 0 0 0 0 0 0 ...
0 0 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 ... 0) }
identify { diff = (0 1) sdiff = (0 1) }
Example 8 Same as Example 7, except that the temporary level shift regressor is used.
series {title = "Quarterly sales" start = 1981.1

data = (301 294 ... 391) period = 4 }
regression { variables = tl1985.03-1987.01 }
identify { diff = (0 1) sdiff = (0 1) }
Example 9 Estimate a model that involves a constant, fixed seasonal effects, and two user-defined
regression variables. The data for the latter two variables is stored in the file weather.dat
in the current directory. This file includes data on several other variables not being used
in the model. The data for the two user-defined regression variables is extracted from this
larger file using a Fortran format that skips the first 16 columns in the file. The start date
is specified since the data set of user-defined regression variables begins before the data for
the time series being modelled.
series {title = "Monthly Riverflow" start = 1970.1

data = (8.234 8.209 ... 8.104) period = 12 }
regression { variables = (seasonal const)
user = (temp precip)
file = "weather.dat"
format = "(t17,2f8.2)"
start = 1960.1 }
arima { model = (3 0 0)(0 0 0) }
estimate { }
Example 10 Estimate a model for a monthly retail inventory series with end-of-month stock trading-day
effects and one AO. The transformed series, minus the regression effects, follows an ARIMA
(0 1 0)(0 1 1)12 model. Decide (using AICC) if the stock trading-day regressors should be
kept in the model.
series {title = "Retail Inventory - Family Apparel"

start = 1967.1 period = 12
data = (1893 1932 ... 3201 )
type = stock }
regression { variables = (tdstock[31] ao1980.jul)
aictest=tdstock }
arima { model = (0 1 0)(0 1 1) }
estimate { }
Example 11 Estimate a model for a monthly retail sales series with stable seasonal and trading day
regressors. Include regressors for a change-of-regime in both sets of regressors in December
of 1985. The transformed series, minus the regression effects, follows an ARIMA (0 1 1)
model.
series { title = "Retail Sales - Televisions"

start = 1976.1 period = 12 type = flow
file = ’tvsales.ori’ }
regression { variables = (td/1985.dec/ seasonal/1985.dec/) }
estimate { }
Example 12 Similar to example 11, only partial change-of-regime regressors are used in conjunction with
the seasonal and trading day regressors so that the extra regressors are set to zero before
December of 1985.
series {title = "Retail Sales - Televisions"

start = 1976.1 period = 12 type = flow
file = ’tvsales.ori’ }
regression { variables = (td td//1985.dec/
seasonal seasonal//1985.dec/) }
estimate { }
Example 13 Estimate a model with two AOs, and two LSs for a quarterly seasonal series. Accounting
for these effects, the transformed series follows an ARIMA (0 1 1)(0 1 1)4 model. The two
LSs are very close together, and may offset each other.

Data = (1039 1241 ... 2210) }
Regression { Variables = (AO2001.3 LS2007.1 LS2007.3 AO2008.4) }
Arima { Model = (0 1 1)(0 1 1) }
Estimate { }
Example 14 The same as Example 13, except that we have replaced the two level shift outliers with a
temporary level shift regressor.

Data = (1039 1241 ... 2210) }
Regression { Variables = (AO2001.3 TL2007.1-2007.2 AO2008.4) }
Arima { Model = (0 1 1)(0 1 1) }
Estimate { }
Example 15 Specified regression variables are a trend constant and trading day effects. Use the automatic
modeling procedure to select an ARIMA model. Additively seasonally adjust the series after
preadjusting for the trading day regression effects.
series { title = "Exports of pasta products"

start = 1980.jan data = "pasta.dat" }
regression { variables = (const td) }
automdl { }
x11 { mode = add }
Example 16 The regression effects selected are seasonal means, a constant, several outliers, trading day,
and an Easter effect. There are user-defined regression variables for special sales promotions
in 1988, 1989 and 1990, which are located in the file promo.dat in 3f12.0 format. The
ARIMA part of the model is (2,1,0). Seasonally adjust the series after pre-adjusting for
all the regression effects. Remove the Easter effects and trading day effects from the final
seasonally adjusted series. Generate 24 forecasts.
series{ title = "Retail sales of children’s apparel"

file = "capprl.dat" start = 1975.1 }
regression{
variables = (const td ao1976.oct ls1991.dec easter[8]
seasonal)
user = (sale88 sale89 sale90)
start = 1975.1 file = "promo.dat" format = "(3f12.0)"
}
arima{ model = (2 1 0) }
forecast{ maxlead = 24 }
x11{ save=seasonal appendfcst=yes }
Example 17 The same as Example 16, except that the user-defined regression effect will be handled the
same way as additive outliers with regard to prior adjustments, final adjustments, print files,
and save files.
series{ title = "Retail sales of children’s apparel"

file = "capprl.dat" start = 1975.1 }
regression{
variables = (const td ao1976.oct ls1991.dec easter[8]
seasonal)
user = (sale88 sale89 sale90)
start = 1975.1 file = "promo.dat" format = "(3f12.0)"
usertype = ao
}
x11{ save=seasonal appendfcst=yes }
Example 18 Specify a regARIMA model with trading day and outlier terms. Specify starting values for
the regression coefficients, and hold the coefficients of the outlier regressors fixed at these
values. Use this model to generate 12 forecasts (by default, since an x11 spec is present).
Perform a default multiplicative seasonal adjustment, after prior adjustment for trading day
and outlier factors.
series{
format = "2L"
title = "Midwest Total Starts"
file = "mwtoths.dat"
name = "MWTOT "
}
transform{ function=log }
arima{ model=(0 1 2 )(0 1 1 ) }
estimate{ save=mdl }
regression{
variables = (ao1977.jan ls1979.jan ls1979.mar ls1980.jan td)
b = ( -0.7946F -0.8739F 0.6773F -0.6850F 0.0209
~0.0107 -0.0022 0.0018 ~0.0088 -0.0074 )
}
x11{ }
Example 19 Read in the data from a file using a predefined X-11 data format. Note that the starting date
is taken from the information provided in the data file, so it does not have to be specified.
Specify a regARIMA model with trading day and holiday terms. Perform automatic outlier
identification, and print out model diagnostics. Use this model to generate 12 forecasts.
Perform a multiplicative seasonal adjustment, using a 3x3 seasonal moving average, after
prior adjustment for trading day, outlier and holiday factors. Remove the holiday and trading
day factors from the final seasonally adjusted series. Save the trading day and holiday factors
in individual output files.
Series {
Format="1L" File="bdptrs.dat" Name="BDPTRS"
Title="Department Store Sales" }
Transform { Function=Log }
Regression { Variables=( Td Easter[8] )
Save = ( Td Holiday ) }
Arima { Model=(0 1 1)(0 1 1) }
Outlier { }
Estimate { }
Check { }
Forecast { }
X11 {
Mode = Mult Seasonalma = S3X3
Title=( "Department Store Retail Sales Adjusted For"
"Outlier, Trading Day, And Holiday Effects" )
}
Example 20 This spec file reads in a set of seasonal regressors saved from a previous X-13ARIMA-SEATS
run. The series adjusted for regression effects (including the user-defined seasonal effect) is
saved.
series{ title = "US Total Housing Starts"

file = "ustoths.dat" start = 1990.1
period = 4 save = b1}
regression{
user = (s1 s2 s3)
usertype = seasonal
start = 1985.1 file = "seasreg.rmx"
format = "x12save"
}
outlier{ }
Example 21 This example shows how to specify a groups of user-defined holiday regressors for payments
made to child care workers in Taiwan. Holiday regressors are specified for Chinese New Year,
the Moon Festival, and the Mid Fall Festival. The chi2test option is used to determine
which of the user-defined holiday regressors are significant.
series{
file="serv.dat"
start=1991.jan
span=(1993.jan,)
title = "Payment to family nanny, taiwan"
}
transform{ function=log }
regression{
variables = ( AO1995.Sep AO1997.Jan AO1997.Feb )
user=(
Beforecny Betweencny Aftercny
Beforemoon Betweenmoon Aftermoon
Beforemidfall Betweenmidfall Aftermidfall
)
file="u1u2u3.dat"
format="datevalue"
start=1991.1
usertype=(
holiday holiday holiday
holiday2 holiday2 holiday2
holiday3 holiday3 holiday3
)
chi2test = yes
savelog = chi2test
}
arima{ model=(0 1 1)(0 1 0) }
check{ }
forecast{
maxlead=12
}
estimate{ savelog=(aic aicc bic) }
7.14. SEATS 157
7.14 SEATS
DESCRIPTION
An optional spec invoking the production of model based signal extraction using SEATS, a seasonal adjustment
program developed by Victor Gómez and Agustin Maravall at the Bank of Spain.
The user can set options which control ARIMA model estimation if done within the SEATS module (epsiv
and maxit), perform checks on the model submitted to the SEATS modules (qmax, rmod and xl). The user
can also choose options to decompose the trend-cycle into a long-term trend and a cycle component using the
modified Hodrick-Prescott filter (hpcycle and hplan).
USAGE
seats{ appendfcst = yes
hpcycle = yes
qmax = 20
rmod = 0.85
signifsc = 2
statseas = yes
out = 2
print = ( s10 s11 s12 s1s s2s)
save = ( s10 s11 )
savelog = ( normalitytest seatsmodel )
}
ARGUMENTS
appendfcst Determines if forecasts will be included in certain SEATS tables selected for storage with
the save option. If appendfcst=yes, then forecasted values will be stored with table s10.
If appendfcst=no, no forecasts will be stored. The default is to not include forecasts.
hpcycle If hpcycle = yes, then the program will decompose the trend-cycle into a long-term
trend and a cycle component using the modified Hodrick-Prescott filter. The default
is not to perform this decomposition (hpcycle = no). For more information on the
Hodrick-Prescott filter, see Kaiser and Maravall (2001).
out Sets level of seasonal decomposition diagnostic output. The default (out = 1) produces
the standard SEATS output in which filter and diagnostic output are obtained from
infinite (Wiener-Kolmogorov) filters and signal extraction error and revisions statistics
are associated with semi-infinite or bi-infinite data. With out=0, the most complete
output, or out=2, a more abbreviated output, all of the filter output and most of the
signal extraction error and revisions statistics are finite-sample quantities for the available
data.
If tables are specified in the print argument, out is set to 2; else, the default is out = 1.
print and save Table 7.30 gives the available tables that can be both printed out and saved for this spec.
The specsa and specirr tables are printed out by default. Choices here override the
selection made with the out argument - if one or more of these tables is selected, no other
SEATS output will be produced.
Table 7.30: Output Tables for Seats spec that can be Printed and Saved
Name Small Description of table

trend s12 final SEATS trend component
seasonal s10 final SEATS seasonal component
irregular s13 final SEATS irregular component
seasonaladj s11 final SEATS seasonal adjustment component
transitory s14 final SEATS transitory component
adjustfac s16 final SEATS combined adjustment factors
adjustmentratio s18 final SEATS adjustment ratio
trendfcstdecomp tfd forecast of the trend component
seasonalfcstdecomp sfd forecast of the seasonal component
seriesfcstdecomp ofd forecast of the series component
seasonaladjfcstdecomp afd forecast of the final SEATS seasonal adjustment
transitoryfcstdecomp yfd forecast of the transitory component
seasadjconst sec final SEATS seasonal adjustment with constant term included
trendconst stc final SEATS trend component with constant term included
totaladjustment sta total adjustment factors for SEATS seasonal adjustment
diffseasonaladj dsa fully differenced transformed SEATS seasonal adjustment
difftrend dtr fully differenced transformed SEATS trend
seasonalsum ssm seasonal-period-length sums of final SEATS seasonal compo-
nent
Note that in the descriptions for the tables named diffseasonaladj and difftrend
given in Table 7.30, the term “fully differenced” means differenced to the order of the
sum d + D of the nonseasonal and seasonal differencing orders of the ARIMA model.
Table 7.31 gives a listing of tables that can only be saved by the program. Specifying
one of the these tables in the print argument will have no effect on the printout - they
should only be used with the save argument.
Note that many of the series specified in Table 7.31 are only produced in the finite filter
calculations are used for the SEATS decomposition - you cannot save these tables if out
= 1 in the seats spec. Also, the component models cannot be saved when out = 2.
to save certain tables as percentages rather than ratios. Specifying these table names in
the print argument will not change the output of the program, and the percentages are
only produced when a log transformation is specified in the transform spec.
7.14. SEATS 159
Table 7.31: Output Tables for Seats spec that can Only be Saved

componentmodels mdc models for the components
filtersaconc fac concurrent finite seasonal adjustment filter
filtersasym faf symmetric finite seasonal adjustment filter
filtertrendconc ftc concurrent finite trend filter
filtertrendsym ftf symmetric finite trend filter
squaredgainsaconc gac squared gain for finite concurrent seasonal adjustment filter
squaredgainsasym gaf squared gain for finite symmetric seasonal adjustment filter
squaredgaintrendconc gtc squared gain for finite concurrent trend filter
squaredgaintrendsym gtf squared gain for finite symmetric trend filter
timeshiftsaconc tac time shift for finite concurrent seasonal adjustment filter
timeshifttrendconc ttc time shift for finite concurrent trend filter
wkendfilter wkf end filters of the semi-infinite Wiener-Kolmogorov filter

seasonalpct pss final seasonal factors, expressed as percentages if appropriate
irregularpct psi final irregular component, expressed as percentages if appropriate
transitorypct psc final transitory component, expressed as percentages if appropriate
adjustfacpct psa combined adjustment factors, expressed as percentages if appropriate
qmax Sets a limit for the Ljung-Box Q statistic, which is used to determine if the model provided
to the SEATS module is of acceptable quality. Default is qmax=50.
When model coefficients are fixed in the arima or regression specs, it is often necessary
to choose a larger value of qmax to keep SEATS from changing the model.
7.33.
signifsc A parameter used to adjust the frequencies examined in a test for significant seasonality.
The default is signifsc = 0.
statseas If statseas=no, the program will not accept a stationary seasonal model, and will change
the seasonal part of the model to (0 1 1). If statseas=yes, the program will accept a
stationary seasonal model. The default is statsea = yes.
Table 7.33: Available Log File Diagnostics for Seats

seatsmodel smd Model used by the SEATS module for signal extraction
x12model xmd Model submitted to the SEATS module
normalitytest nrm Normality test
overunderestimation oue Over-under estimation diagnostics
totalsquarederror tse Total mean squared error
componentvariance cvr Component variances
concurrentesterror cee Concurrent estimation error
percentreductionse prs Percent reduction standard error
averageabsdiffannual aad Annual Average absolute difference
seasonalsignif ssg Test for seasonal significance
bias Corrects for the bias that may occur in multiplicative decomposition when the period-to-
period changes are relatively large when compared to the overall mean. This argument
should only be set when a log transformation is used.
If Bias = 1, a correction is made for the overall bias for the full length of the series and
for the forecasting period. This is the default value.
If Bias = -1, a correction is made so that, for every year (including the forecasting pe-
riod), the annual average of the original series equals the annual average of the seasonally
adjusted series, and also (very approximately) equals the annual average of the trend.
If Bias = 0, no bias correction is done. No other values are allowed.
centerir If centerir=yes, all regression preadjustments submitted to the SEATS module will be
centered by removing the mean effect. If centerir=no, the regression factors will not be
changed. The default is not to center the regression preadjustments.
epsiv Convergence criteria for ARIMA estimation within the SEATS module; this is used when
the SEATS module determines that a model should be changed or re-estimated. This
should be a small positive number; the default is 0.001.
epsphi When Phi(B) contains a complex root, it is allocated to the seasonal if its frequency
differes from the seasonal fequencies by less than epsphi degrees. Otherwise, it goes to
the cycle. The default is 2.
hplan A parameter that is used to determine the modified Hodrick-Prescott filter. By default,
the program will set this parameter automatically according to the seasonal period of
the series. For more information on the Hodrick-Prescott filter, see Kaiser and Maravall
(2001).
7.14. SEATS 161
maxbias A positive number which sets a limit that checks if a multiplicative adjustment should
be used on a specific series. When the average value of the differences (in absolute value)
between the annual means of the original and seasonally adjusted series is larger than
maxbias, a warning message is printed out. The number is expressed in percentage
points of the level of the series. Default is 0.5.
maxit Number of iterations allowed for ARIMA estimation within the SEATS module; should
be a positive integer. Default is 20.
noadmiss When noadmiss=yes, if the model submitted to SEATS does not lead to an admis-
sible decomposition, it will be replaced with a decomposable model. Otherwise when
noadmiss=no, no approximation is done in this case. The default is noadmiss=no.
rmod Limit for the modulus of an AR root. If the modulus of an AR root is larger than rmod,
the root is assigned to the trend; if the modulus of an AR root is smaller than rmod,
the root is assigned to the cycle. The default value of rmod is 0.80.
xl When the modulus of an estimated root falls in the range (XL, 1), it is set to 1.00 if the
root is in the AR polynomial. If the root is in the MA polynomial, it is set to xl. The
default is 0.99.
DETAILS
The most recent versions of TRAMO and SEATS, as well as more documentation, are available at the Bank
of Spain web site (http://www.bde.es/webbde/es/secciones/servicio/software/econom.html). Possible
delays in the updating of versions may cause slight differences between the X-13A-S version of SEATS and the
one at the Bank of Spain web site.
For more details on the SEATS seasonal adjustment method and diagnostics, see Maravall (1995), Gómez and
Maravall (1996) and Gómez and Maravall (2001b); for comparisons of SEATS and X-11 seasonal adjustments
and filters, see Hood, Ashley, and Findley (2000), Hood (2002b), Findley and Martin (2003), and Findley, Wills,
Aston, Feldpausch, and Hood (2003).
Note that there are other output files that were saved by the SEATS program that are available when
running the X-13ARIMA-SEATS program. These output files can contain forecasts, components or diagnostics
generated from the SEATS model-based adjustment performed. Table 7.34 shows the file extensions that are
used to save the corresponding special output file from SEATS in the same way the short table names are used
as file extensions in storing individual tables to separate files. These extensions do not have to be specified
in the save argument - these files will be produced for every X-13ARIMA-SEATS run with a SEATS seasonal
adjustment. Section 3.2 gives details on the naming conventions used for X-13ARIMA-SEATS saved output. Note
that many of these tables will not be produced when out=2.
The matrix formulas of McElroy (2008a) for (nonstationary) ARIMA signal extraction substantially simplify
those of Bell and Hillmer (1988). We will motivate them from re-expressions of the standard regression formulas
for stationary (or more general) linear unobserved component decompositions of mean zero data wt , t = 1, . . . , n
into uncorrelated mean zero components
wt = ut + vt .
Table 7.34: X-13ARIMA-SEATS File Extensions for Special SEATS Saved Output
SEATS X-13ARIMA-SEATS Produced when Contents

file name extension out=2? of file
f.txt tbf no Forecasts of the series, seasonally
adjusted series, and components of
the model-based seasonal adjustment,
saved in columns separated by white
space
o.txt tbo no Series, seasonally adjusted series, and
components of the model-based sea-
sonal adjustment, saved in columns
separated by white space
rogtable.out rog yes Selected statistics from the growth
rate output
summarys.txt sum yes Summary information and diagnostics
from SEATS adjustment
table-s.out tbs yes Annotated listing of the series, the sea-
sonally adjusted series, and compo-
nents of the model-based seasonal ad-
justment, saved in columns separated
by white space
z.txt tbz no Log of the series, seasonally adjusted
series, and components of the model-
based seasonal adjustment, saved in
columns separated by white space
SEATS file name gives the file name saved by the SEATS program.
X-13ARIMA-SEATS extension gives the file extension used to save the output from the corre-
sponding SEATS output file.
Thus, with u = (u1 , . . . , un )0 , v = (v1 , . . . , vn )0 , and w = (w1 , . . . , wn )0 , we require Σuv ≡ Euv 0 = 0n×n , which
yields the variance matrix decomposition
Σww = Σuu + Σvv
and the covariance matrix formula
0
Σuw ≡ Eu (u + v) = Σuu .
Hence, given Σuu and Σvv , the basic formula
β = Σuw Σ−1
ww
0
for the coefficient matrix minimizing E (u − βw) (u − βw), which provides the mean square optimal linear
estimate û of u from w,
û = βw,
7.14. SEATS 163
can be evaluated as
−1
β = Σuu (Σuu + Σvv )
−1 −1 −1
Σ−1

= uu + Σvv Σvv . (7.11)
Similarly, the variance matrix of the error e = u − û reduces to
Σee = Σuu − Σuu Σ−1

ww Σuu
−1 −1 −1

= Σuu + Σvv . (7.12)
The less familiar formulas (7.11) and (7.12) generalize to the ARIMA case.
Now consider ARIMA data Y1 . . . , Yn with differencing polynomial δY (B) = 1 + δ1 B 1 + · · · + δd B d (d ≥ 1)
resulting in mean zero stationary wt = δY (B) Yt , d + 1 ≤ t ≤ n. We assume there is a signal plus noise
decomposition of Yt into difference-stationary components
Yt = St + Nt
with differencing operators δS (B) and δN (B) of degree dS , respectively dN , with no common zeroes and
with δY (B) = δS (B) δN (B), resulting in mean zero processes
ut = δS (B) St , vt = δN (B) Nt .

that are uncorrelated, Eut vt+h = 0, for all t, h. For example, if δY (B) = (1 − B) 1 − B 12 and St is the
seasonal component for monthly data, then
2
δS (B) = 1 + B + · · · + B 11 , δN (B) = (1 − B) .
Stationary case formulas do not apply to estimate nonstationary St : variance matrices of ARIMA data
cannot be estimated consistently.
Consider the simplest difference stationary model, the random walk :
zt = zt−1 + at or (1 − B)zt = zt − zt−1 = at , (7.13)

0
with
P uncorrelated, zero-mean at with Ea2t = σa2 for t = 2, . . . , n. The variance matrix of a = (a2 , . . . , an ) ,
2
being a = σa In−1 , can be consistently estimated from the available zt −zt−1 . The variance matrix of z1 , . . . , zn
cannot: For t ≥ 2, we have
t
X
zt = zt−1 + at = zt−2 + at−1 + at = · · · = z1 + aj .
j=2
Assuming z1 is uncorrelated with all at (Assumption A of Bell, 1984), for any k ≥ 0,
 
t+k
X
Ezt zt+k = Ezt2 + Ezt  aj 
j=t+1
= Ezt2 + 0 = Ez12 + (t − 1) σa2 .
Clearly Ez12 cannot be estimated consistently from one datum z1 .

In the ARIMA signal extraction case, assuming that Y1 , . . . , Yd are uncorrelated with all wt , McElroy (2008a)
shows that the mean square optimal linear estimate of St , 1 ≤ t ≤ n is
Ŝ = βY (7.14)
with
−1
β = ∆0S Σ−1 0 −1
uu ∆S + ∆N Σvv ∆N ∆0N Σ−1
vv ∆N . (7.15)
Here ∆S implements the calculation of ut = δS (B) Yt , dS+1 ≤ t ≤ n and Σuu is the variance matrix of
udS +1 , . . . , un . Similarly, ∆N implements vt = δN (B) Yt , dN +1 ≤ t ≤ n and Σvv is the variance matrix of
vdN +1 , . . . , vn . The variance matrix Σee of the signal extraction error e = S − Ŝ is given by
−1
Σee = ∆0S Σ−1 0 −1
uu ∆S + ∆N Σvv ∆N . (7.16)
SEATS, and its implementation in X-13ARIMA-SEATS , use the procedure of Hillmer and Tiao (1979) to derive
ARIMA models for St and Nt from the ARIMA model for Yt (assuming this ARIMA model has an ”admissible”
decomposition). Here St can denote any of the seasonal decomposition components. (Seasonal, trend, irregular,
seasonally adjusted series, etc.) From the ARIMA models for St and Nt , the matrices Σuu and Σvv can be
obtained, and therefore also the matrix of filters β for producing the component estimates Ŝt , 1 ≤ t ≤ n, as well
as Σee . From Σee , standard errors and confidence intervals for Ŝt can be obtained (which do not account for
modeling error). When the log-transformation is used for modeling, Ŝt and associated confidence intervals are
exponentiated to obtain the estimates and confidence intervals for the observed data’s seasonal decomposition
components.
The program does not use the matrix formulas (7.14)–(7.15) to calculate the component estimates Ŝt , 1 ≤
t ≤ n. Instead, the original method of SEATS is used, which does not involve the time-consuming inversion of
large matrices for long series. This ”Wiener-Kolmogorov” method produces identical component estimates, but
only bi-infinite-sample approximations to Σee of (7.16) and to the associated standard errors and confidence
intervals. Similarly for the filter diagnostics. Setting out=2 in the seats spec of X-13ARIMA-SEATS causes the
matrix-based (finite-sample) versions of almost all diagnostics to be produced. These will also be available in
the Bank of Spain’s next release of SEATS.
The finite-sample filter diagnostics (squared gain and time-shift functions) are illustrated and compared with
infinite-filter diagnostics in Findley and Martin (2006). A derivation, analysis and comparison of one of the
7.14. SEATS 165
finite-sample over-/underestimation tests is given in Findley, McElroy, and Wills (2005). The general derivation
of the finite-sample versions of these tests and their asymptotic distributions is given in McElroy (2006, 2008b).
Finite-sample versions of other diagnostics have also been implemented, see McElroy and Gagnon (2006, 2008).
The tests are goodness-of-fit tests for the time series model chosen for the series, each focussing on the
statistical properties of the models it yields for the seasonal factors, the seasonally adjusted series, the trend, and
the irregular and on the properties these models predict for the variances and certain covariances of the estimates
of these components. When the differencing operator for the ARIMA model for the (usually log-transformed)
D D
series is (1 − B)d (1 − B s ) = (1 − B)d+D 1 + B + · · · + B s−1 for s = 4 or 12, the basic component model
s−1 D

assumptions are that application of 1 + B + · · · + B to the (usually log-transformed) seasonal component
produces a stationary series whose ARMA model is known, and that application of (1 − B)d+D to the seasonally
adjusted series and trend, which yields what we call the fully differenced seasonally adjusted series and trend,
does likewise for these components. Often, statistically significant values of the test statistics arise because
D
application of 1 + B + · · · + B s−1 or (1 − B)d+D to these components yields a series which is not stationary
over the whole time interval of the observed series. This nonstationarity can often be detected in graphs of the
outputs of these differencing operators applied to the estimated components. These outputs are available as the
seasonalsum, diffseasonaladj and difftrend tables listed in Table 7.30 and the graphs can be obtained from
X-13-Graph (see Lytras 2012a, 2012b). Examination of the graphs when there is nonstationarity will frequently
reveal shorter data intervals over which acceptable goodness-of-fit results can be obtained.
We recommend that series for which a stationary seasonal model is chosen (i.e., a model with seasonal AR
or MA coefficients but no seasonal differencing) should not be seasonally adjusted. Adjustments of such series
are susceptible to large revisions and are conceptually problematic, because seasonal factors of a given calendar
month quickly change from indicating an increase to indicating a decrease, or vice versa. Thus, the repetitive
quality inherent in the concept of seasonality is lacking. Changing the seasonal part of the model to (011), as
statseas=no does, rarely produces more stable results and often imparts seasonality to the seasonally adjusted
series.
EXAMPLES
Example 1 A SEATS seasonal adjustment will be generated from the model determined by the automatic
modeling procedure. The transformation will be selected by the automatic transformation
selection procedure. Outlier identification will be performed for point, level change, and
temporary change outliers.
SERIES { TITLE="EXPORTS OF TRUCK PARTS"

START =1987.1
FILE = "X21109.ORI"
PERIOD = 12
}
TRANSFORM { FUNCTION = AUTO }
REGRESSION { AICTEST = TD }
AUTOMDL { }
OUTLIER { TYPES = (AO LS TC) }
FORECAST { MAXLEAD = 24 }
SEATS { SAVE = S11 }
Example 2 A SEATS seasonal adjustment will be generated from the model specified by the user. Setting
out=2 in the seats spec will cause the finite sample output to be used, and allow the user to
save finite sample filter diagnostics. A revision history of the seasonally adjusted series and
the trend component will be performed, and the percent revisions of the seasonally adjusted
series and the trend component will be saved in separate files.
Series { Title="Quarterly Exports Of Mangos"

Start =1990.1 File = "Xmango.Ori" Period = 4
}
Regression { Aictest = Td }
Arima { Model = (0 1 1)(0 1 1) }
Forecast { Maxlead = 8 }
Seats { Out = 2
Save = ( Squaredgainsaconc Timeshiftsaconc )
Savelog = Overunderestimation
}
History { Estimates = (Sadj Trend)
Save = ( Sarevisions Trendrevisions )
}
7.15. SERIES 167
7.15 SERIES
DESCRIPTION
Required spec that provides X-13ARIMA-SEATS with the time series data, a descriptive title for the series, the
starting date of the series, the seasonal period (12 for monthly data, 4 for quarterly data,) and an optional
restricted span (subset) within the time series to be used for the analysis. The data can either be included in
the series spec by using the data argument, or they can be obtained from a file by using the file argument.
Note that if X-13ARIMA-SEATS is run using a data metafile, the series should not be specified in this spec, since
data files are specified in the data metafile (for more details, see Section 2.5).
USAGE
series{ title = "Example Series"
start = 1967.1
span = (1970.1,)
modelspan = (1985.Jan, 0.Dec)
name = "tstsrs"
data = (480 · · · 1386) or file = "example.dat"
format = "2r"
decimals = 2
precision = 1
comptype = add
compwt = 1.0
print = (none + header)
save = (spn)
appendfcst = yes
appendbcst = no
type = stock
}
ARGUMENTS
appendbcst Determines if backcasts will be included in certain tables selected for storage with the
save option. If appendbcst=yes, then backcasted values will be stored with tables a16,
b1, d10, d16, and h1 of the x11 spec, table s10 of the seats spec, tables a6, a7, a8, a8.tc,
a9, and a10 of the regression spec, and tables c16 and c18 of the x11regression spec.
If appendbcst=no, no backcasts will be stored. The default is to not include backcasts.
appendfcst Determines if forecasts will be included in certain tables selected for storage with the save
option. If appendfcst=yes, then forecasted values will be stored with tables a16, b1,
d10, d16, and h1 of the x11 spec, tables a6, a7, a8, a8.tc, a9, and a10 of the regression
spec, and tables c16 and c18 of the x11regression spec. If appendfcst=no, no forecasts
will be stored. The default is to not include forecasts.
comptype Indicates how a component series of a composite (also called aggregate) series is incorpo-
rated into the composite. These component series can be added into the (partially formed)
composite series (comptype=add), subtracted from the composite series (comptype=sub),
multiplied by the composite series (comptype=mult), or divided into the composite series
(comptype=div). The default is no aggregation (comptype=none).
compwt Specifies that the series is to be multiplied by a constant before aggregation. This con-
stant must be greater than zero (for example, compwt=0.5). This argument can only be
used in conjunction with comptype. The default composite weight is one.
data Vector containing the time series data. The data are read row-wise in the following
format: there must be at least one blank space, comma, or carriage return separating
each of the data values. The number of observations is automatically determined as the
length of the data vector supplied. If the data argument is used, the file argument
cannot be used.
decimals Specifies the number of decimals that will appear in the seasonal adjustment tables of the
main output file. This value must be an integer between 0 and 5, inclusive (for example,
decimals=5). The default number of decimals is zero.
file Name of the file containing the time series data. The filename must be enclosed in quotes.
If the file is not in the current directory, the complete filename including the path must
be given. Valid path and filenames depend on the computer operating system. If the file
argument is used, the data argument cannot be used.
format Denotes the format to be used in reading the time series data from the named file, when
the data are not in free format. Several types of input can be used:
b. a valid Fortran format, which should be enclosed in quotes and must include the
initial and terminal parentheses (example: format="(6f12.0)").
c. a two character code which corresponds to a set of data formats used in previous
versions of X-11 and X-11-ARIMA (example: format="1r");
d. “datevalue” format, where the year, month or quarter, and value of each observation
are found in this order in free format on individual lines. Thus, a line of the data
file containing the value 32531 for July of 1991 would have the form 1991 7 32531.
The number of preceding blanks can vary (example: format="datevalue");
e. the format X-13ARIMA-SEATS uses to save a table. This allows the user to read in a
file saved from a previous X-13ARIMA-SEATS run (example: format="x12save");
f. the format that the TRAMO and SEATS programs use to read in a series and its
descriptors. This enables X-13ARIMA-SEATS to read in a data file formatted for the
TRAMO modeling program or the SEATS seasonal adjustment program. (example:
format="tramo");
g. a variant of “free” format where the numbers must be separated by one or more
7.15. SERIES 169
h. a variant of “datevalue” format, where the year, month or quarter, and value of
In the predefined X-11 data formats, the data is stored in 6 or 12 character fields, along
with a year and series label associated with each year of data. For a complete list of
these formats and how they are used, see DETAILS.
If no format argument is given, the data will be read in free format. Format cannot be
used with the data argument, only with file.
modelspan Specifies the span (data interval) of the data to be used to determine all regARIMA
model coefficients. This argument can be utilized when, for example, the user does not
want data early in the series to affect the forecasts, or, alternatively, data late in the
series to affect regression estimates used for preadjustment before seasonal adjustment.
As with the span spec detailed above, the modelspan argument has two values, the
start and end date of the desired span. A missing value defaults to the corresponding
start or end date of the span of the series being analyzed. For example, for monthly data,
the statement modelspan=(1968.1, ) causes whatever regARIMA model is specified in
other specs to be estimated from the time series data starting in January, 1968 and
ending at the end date of the analysis span. A comma is necessary if either the start or
end date is missing. The start and end dates of the model span must both lie within
the time span of data specified for analysis in the series spec, and the start date must
precede the end date.
of modelspan to always be the most recent occurrence of a specific calendar month
(quarter for quarterly data) in the span of data analyzed, where per denotes the calendar
month (quarter). Thus, if the span of data considered ends in a month other than
December, modelspan=(,0.dec) will cause the model parameters to stay fixed at the
values obtained from data ending in the next-to-final calendar year of the span.
name The name of the time series. The name must be enclosed in quotes and may contain up
to 64 characters. Up to the first 16 characters will be printed as a label on every page.
When specified with the predefined formats of the format argument, the first six (or
eight, if format="cs") characters of this name are also used to check if the program is
reading the correct series, or to find a particular series in a file where many series are
stored.
period Seasonal period of the series. If seasonal adjustments are generated, the only values
currently accepted by the program are 12 for monthly series and 4 for quarterly series.
Otherwise, any seasonal period up to 12 can be specified. (This limit can be changed—see
Section 2.8.) The default value for period is 12.
precision The number of decimal digits to be read from the time series. This option can only be
used with the predefined formats of the format argument. This value must be an integer
between 0 and 5, inclusive (for example, precision=5). The default is zero. If precision
is used in a series spec that does not use one of the predefined formats, the argument is
ignored.
the default printout, except seriesplot and adjoriginalplot. For a complete listing
of the brief and default print levels for this spec, see Appendix B.
span Limits the data utilized for the calculations and analysis to a span (data interval) of
the available time series. The span argument has two input values, the start and end
date of the span. A missing value defaults to the corresponding start or end date of the
input time series. For example, assuming monthly data, the statement span=(1968.1,
) specifies a span starting in January, 1968 and ending at the end date of the series input
through the data or file argument. A comma is necessary if either the start or end date
is missing. The start and end dates of the span must both lie within the series, and the
start date must precede the end date.
Table 7.35: Available Output Tables for Series
name short save description of table

header hdr · summary of options selected for this run of X-13-
ARIMA-SEATS
span a1 + time series data, with associated dates (if the
span argument is present, data are printed
and/or saved only for the specified span)
seriesplot a1p · plot of the original series
specfile spc · contents of input specification file used for this
run
savefile sav · list of files to be produced by the X-13-
ARIMA-SEATS run
specorig sp0 + spectral plot of the first-differenced original series
missingvaladj mva + original series with missing values replaced by reg-
ARIMA estimates
calendaradjorig a18 + original series adjusted for regARIMA calendar
effects
outlieradjorig a19 + original series adjusted for regARIMA outliers
adjoriginal b1 + original series, adjusted for prior effects and fore-
cast extended
adjorigplot b1p · plot of the prior adjusted original series aug-
mented by prior-adjusted forecasts (if specified);
if no prior factors or forecasts are used, the origi-
nal series is plotted
start The start date of the time series in the format start=year.seasonal period . (See Section
7.15. SERIES 171
3.3 and the examples below.) The default value of start is 1.1. (See DETAILS.)
title A title describing the time series. The title must be enclosed in quotes and may contain
up to 79 characters. It will be printed on each page of the output (unless the -p option
is evoked; see Section 2.7).
type Indicates the type of series being input. If type = flow, the series is assumed to be a
flow series; if type = stock, the series is assumed to be a stock series. The default is to
not assign a type to the series.

divpower An integer value used to re-scale the input time series prior to analysis. The program
divides the series by ten raised to the specified value. For example, setting divpower =
2 will divide the original time series by 102 , while divpower = -4 will divide the series
by 10−4 . Integers from -9 to 9 are acceptable values for divpower. If this option is not
specified, the time series will not be re-scaled.
missingcode A numeric value in the input time series that the program will interpret as a missing
value. This option can only be used in input specification files requiring a regARIMA
model to be estimated or identified automatically. The default value is -99999. Example:
missingcode=0.0.
missingval The initial replacement value for observations that have the value of missingcode. The
subsequent replacement procedure is described in DETAILS. The default value of miss-
ingval is 1000000000. Example: missingval=10D10.
saveprecision The number of decimals stored when saving a table to a separate file with the save
argument. The default value of saveprecision is 15. Example: saveprecision=10.
trimzero If trimzero=no, zeroes at the beginning or end of a time series entered via the file
argument are treated as series values. IF trimzero=span, causes leading and trailing
zeros to be ignored outside the span of data being analyzed (the span argument must
be specified with both a starting date and an ending date). The default (trimzero=yes)
causes leading and trailing zeros to be ignored. Note that when the format argument is
set to either free, datevalue, x12save, or tramo, all values input are treated as series
values, regardless of the value of trimzero.
DETAILS
The number of observations and the series end date are determined by the program after reading in the data.
X-13ARIMA-SEATS accepts a maximum of 780 observations. (This limit can be changed—see Section 2.8.)
If spec files are copied from one directory to another or from one computer system to another, verify that
the path and filenames in their file arguments remain valid.
The series spec cannot appear in a spec file with the composite spec. The latter signifies that a seasonal
adjustment of a composite series is to be calculated.
Table 7.36 gives a description of the default formats for each of the valid two-character X-11 format codes
for the format argument, as well as the corresponding Fortran format.
Table 7.36: Default Formats for Each X-11 Format Code
Fortran Format Fortran Format

Code for Monthly Data for Quarterly Data Description
1r (12f6.0,I2,A6) (4(12x,f6.0),I2,A6) Year and identifier on the right, data in 6-
digit fields
2r (6f12.0,/,6f12.0,I2,A6) (4f12.0,24x,I2,A6) Year and identifier on the right of the second
line, data in 12-digit fields
1l (A6,I2,12f6.0) (A6,I2,4(12x,f6.0)) Year and identifier on the left, data in 6-digit
fields
2l (A6,I2,6f12.0,/,8x,6f12.0) (A6,I2,4f12.0) Year and identifier on the left of the first line,
data in 12-digit fields
2l2 (A8,I4,6f11.0,2x,/,12x,6f11.0,2x) (A8,I4,4f11.0,2x) Four digit year and identifier on the left of
the first line, data in 11-digit fields
cs (A8,I2,10X,12E16.10,18X) (A8,I2,10X,12E16.10,18X) Data in CANSIM data base utility format,
data in 16-digit fields
cs2 (A8,I4,12X,12E16.10,14X) (A8,I4,12X,12E16.10,14X) Data in the new CANSIM data base utility
format (called CANSIM2), data in 16-digit
fields
These formats can be modified by using the precision argument. If precision is used in a series spec that
doesn’t use an X-11 format code, this argument is ignored.
Note that if one of the X-11 format codes is specified (or if format = "datevalue", format = "datevaluecomma",
format = "tramo", or format = "x12save"), the start of the series is automatically read from the data file.
Therefore, the starting date need not be specified with the start argument of the series spec.
If a data metafile is used to process a group of input files using a single input spec file, the X-11 formats
should be avoided. These formats require the name of the series (specified name) to verify that the data is in
the file. This implies that all the data files in the data metafile would be required to use the same series name.
This is often not desirable.
When doing a formatted read of a data file, X-13ARIMA-SEATS discards sequences of zeroes at the ends of the
series (unless trimzero=no). This convention is used to allow input of series stored in certain formats—Example
3 below gives an illustration. If the zeros at the ends of the series are true data values, trimzero=no will cause
them to be treated as such. However, if the zeroes at the beginning of a given series are real and the zeroes
implied at the end of the series are not (due to blanks at the end of the line), then the file must be modified so
that it can be read in free format. Example 4 below demonstrates this conversion.
The span and modelspan arguments can be used with the forecast spec to generate out-of-sample forecast
comparisons by excluding data at the end of the series. When either of these arguments are present, model
estimation will use data only for the specified span. Forecasting then (by default) proceeds from the end of the
span, producing comparisons of the withheld data with the forecasts. (See Example 4 of the forecast spec.)
Note that if the beginning date specified in the modelspan argument is not the same as the starting date
in the span argument, backcasts cannot be generated by the program, regardless of the value of the maxback
argument of the forecast spec.
7.15. SERIES 173
When the program encounters a value equal to the value of missingcode in the original series, it inserts an
additive outlier for that observation time into the set of regression variables of the model the series and then
replaces the missing value code with a value large enough to be considered an outlier during model estimation.
After the regARIMA model is estimated, the program adjusts the original series using factors generated from
these missing value outlier regressors. The adjusted values are estimates of the missing values.
If a series is designated as a stock or a flow series by using the type argument, then trading day and Easter
regressors specified in regression spec need to agree with this type - one cannot specify stock trading day
regressors for a flow series. If a series type is not specified, then any trading day or holiday regressor may be
used with the series.
EXAMPLES
Note: The following examples, do not show “complete” spec files in the sense that useful output is not produced
unless additional specs (e.g., x11 or arima and estimate) are also included.
Example 1 Specify a time series with the data argument.
series{
title = "A Simple Example"
start = 1967.jan # period defaults to 12
data = ( 480 467 514 505 534 546 539 541 551 537 584 854
522 506 558 538 605 583 607 624 570 609 675 861
.
.
.
1684 1582 1512 1508 1574 2303 1425 1386) }
Example 2 Drop observations from both the beginning and end of a quarterly series that starts in 1940
and ends in 1993. The first six years of data are dropped to restrict the analysis to post-
WWII data. The data held out for 1991–93 could be used to examine out-of-sample forecast
performance.
series { data = (879 899 985 ...) # There are 216 data values
start = 1940.1 # ending in 1993.4
period = 4 # Quarterly series
span = (1946.1, 1990.4) }
Example 3 This example shows how the X-13ARIMA-SEATS program can read data from files stored in a
format adopted from the X-11-ARIMA seasonal adjustment program. Here the data are avail-
able from July, 1970 through February, 1993, and are stored in the file c:\data\sales1.dat
as follows:
146.4 109.2 132.1 144.8 116.1 100.370SALES1

142.9 158.8 196.2 244.0 251.6 245.5 244.2 213.8 188.9 197.2 181.2 161.371SALES1
.
.
.
148.8 177.2 0 0 0 0 0 0 0 0 0 093SALES1
The data are stored in (12f6.1,i2,a6) format, with the last eight columns in each line
providing the year and series ID.
SERIES{ TITLE = "Monthly data in an X-11 format"

PERIOD = 12
FILE = "C:\DATA\SALES1.DAT" # a DOS path and file
PRECISION = 1
FORMAT = "1r" }
Since Fortran formatted reads treat blanks as zeroes, the input of the series obtains six
zeroes at the beginning. The input series also contains the ten zeroes at the end. As noted
in DETAILS, X-13ARIMA-SEATS discards the zeroes read in from both the beginning and
end of the series by default so that only the actual data are retained and assigned to the
correct months (146.4 to July, 1970, etc.). Also note that since the year is given on each
line, the user does not have to enter a start argument.
Example 4 This example illustrates the rare case of a data file that must be modified for correct in-
put to X-13ARIMA-SEATS . The original data file contains data for February, 1980 through
November, 1990 stored in (6f4.0,1x,i4) format as follows.
0 342 -256 491 0 0001

-234 922 -111 2 0 199 0002
.
.
.
581 -987 -423 10 0 0022
This file cannot be read in free format because several of the data entries run together and
because the file contains record counters (0001, 0002, . . . ) in columns 26–29. A free format
read would treat the record counters as data. The file cannot be read with (6f4.0) format
with a start date of February, 1980 because X-13ARIMA-SEATS with the default trimzero =
yes would incorrectly drop the zeroes at the first and last observations, and then erroneously
assign the value 342 to February 1980. Using trimzero = no would add extra zeroes to the
series, as the blank spaces at the beginning and end of the data set would be read as zero.
7.15. SERIES 175
The solution is to reformat the data file so it can be read in free format. This requires removal
of the record counters and separation of the data entries. The modified file, example4.new,
is as follows:
0 342 -256 491 0

-234 922 -111 2 0 199
.
.
.
581 -987 -423 10 0
Then the following series spec will correctly read the data from the file example4.new.
series {title = "Data read correctly in with trimzero = no"

start = 1980.2 period = 12
file = "example4.new" } # file is in current directory
Example 5 This example shows how the X-13ARIMA-SEATS program can read data in “date-value”
format. The data are available from July, 1970 through February, 1993, and are stored in
the file c:\data\sales1.edt as follows:
1970 7 14624
1970 8 10952
1970 9 13251
1970 10 14408
.
.
.
1993 1 14838
1993 2 17762
Each data record contains the year, month and value of a given observation of the time series.
SERIES{ TITLE = "Monthly data in a datevalue format"

PERIOD = 12
FILE = "C:\DATA\SALES1.EDT" # a DOS path and file
FORMAT = "DATEVALUE" TYPE = FLOW }
Note that as in the X-11-ARIMA format shown in Example 3 above, the starting date can
be read directly from the input file, so the user does not have to include a start argument.
Also, the type argument is used to specify that this is a flow series.
Example 6 The same as example 5, but this series will be used as a component in a composite adjust-
ment. The number of decimals displayed in the output is set to be 2, and the span of data
to be modelled will be set to be the start of the series through December, 1992
SERIES{ TITLE = "Monthly data in a datevalue format"

PERIOD = 12
FORMAT = "DATEVALUE"
COMPTYPE = ADD
DECIMALS = 2
MODELSPAN = (,1992.DEC)
}
Example 7 This example shows how the X-13ARIMA-SEATS program handles missing data. The same
data format is used as in the previous two examples, except a missing value code is inserted
for January of 1990:
1970 7 14624
1970 8 10952
1970 9 13251
1970 10 14408
.
.
.
1990 1 -99999.
.
.
.
1993 1 14838
1993 2 17762
The series spec below will replace the missing value code for January 1990 with a number
large enough to be considered an outlier, assuming a regARIMA model is estimated later in
the input specification file.
SERIES{ TITLE = "Monthly data in a date-value format"

PERIOD = 12
FORMAT = "DATEVALUE"
}
7.15. SERIES 177
Example 8 This example shows how the X-13ARIMA-SEATS program can read data from a file previously
saved by X-13ARIMA-SEATS (in a previous run, the outlier adjusted original series was stored
in the file c:\data\sales1.a11).
SERIES{ TITLE = "Monthly data in a file saved by X-12-ARIMA"

PERIOD = 12
FILE = "C:\DATA\SALES1.A11" # a DOS path and file
FORMAT = "X12SAVE" }
Note that as in the X-11-ARIMA format shown in Example 3 and the “datevalue” format
shown in Example 5 above, the starting date can be read directly from the input file, so a
start argument is not included.
Example 9 This example shows how the X-13ARIMA-SEATS program can read data in the special “date-
value” format that uses the convention of commas as decimal points. As in Example 5,
the data are available from July, 1970 through February, 1993, and are stored in the file
c:\data\sales1c.edt as follows:
1970 7 146,24
1970 8 109,52
1970 9 132,51
1970 10 144,08
.
.
.
1993 1 148,38
1993 2 177,62
Each data record contains the year, month and value of a given observation of the time series.
SERIES{ TITLE = "Monthly data in the comma variant of datevalue format"

PERIOD = 12
FILE = "C:\DATA\SALES1C.EDT" # a DOS path and file
FORMAT = "DATEVALUECOMMA" }
7.16 SLIDINGSPANS
DESCRIPTION
Optional spec providing sliding spans stability analysis. These compare different features of seasonal adjustment
output from overlapping subspans of the time series data. The user can specify options to control the starting
date for sliding spans comparisons (start), the length of the sliding spans (length), the threshold values deter-
mining sliding spans statistics (cutseas, cuttd, cutchng), how the values of the regARIMA model parameter
estimates will be obtained during the sliding spans seasonal adjustment runs (fixmdl), and whether regARIMA
automatic outlier identification is performed (outlier).
USAGE
slidingspans{ start = 1975.jan
length = 132
numspans = 3
cutchng = 3.0
cutseas = 3.0
cuttd = 2.0
outlier = yes
fixmdl = no
fixreg = outlier
print = (long -ssheader)
save = (sfspans)
savelog = (percent)
}
ARGUMENTS
cutchng Threshold value for the month-to-month, quarter-to-quarter, or year-to-year percent

changes in seasonally adjusted series. For a month (quarter) common to more than
one span, if the maximum absolute difference of its period-to-period percent changes
from the different spans exceeds the threshold value, then the month (quarter) is flagged
as having an unreliable estimate for this period-to-period change. This value must be
greater than 0; the default value is 3.0. Example: cutchng=5.0
cutseas Threshold value for the seasonal factors and seasonally adjusted series. For a month
(quarter) common to more than one span, if the maximum absolute percent change of its
estimated seasonal factors or adjustments from the different spans exceeds the threshold
value, then this month’s (quarter’s) seasonal factor or adjustment is flagged as unreliable.
This value must be greater than 0; the default value is 3.0. Example: cutseas=5.0
cuttd Threshold value for the trading day factors. For a month (quarter) common to more than
one span, if the maximum absolute percent change of its estimated trading day factors
from the different spans exceeds the threshold value, then this month’s (quarter’s) trading
7.16. SLIDINGSPANS 179
day factor is flagged as unreliable. This value must be greater than 0; the default value
is 2.0. Example: cuttd=1.0
fixmdl Specifies how the initial values for parameters estimated in regARIMA models are to be
reset before seasonally adjusting a sliding span. This argument is ignored if a regARIMA
model is not fit to the series.
If fixmdl=yes, the values for the regARIMA model parameters for each span will be set
to the parameter estimates taken from the original regARIMA model estimation. These
parameters will be taken as fixed and not re-estimated. This is the default for fixmdl.
If fixmdl=no, the program will restore the initial values to what they were when the
regARIMA model estimation was done for the complete series. If they were fixed in the
estimate spec, they remain fixed at the same values.
If fixmdl=clear, initial values for each span will be set to be the defaults, namely 0.1
for all coefficients, and all model parameters will be re-estimated.
fixreg Specifies the fixing of the coefficients of a regressor group in either a regARIMA model or
an irregular component regression. These coefficients will be fixed at the values obtained
from the model span (implicit or explicitly) indicated in the series or composite spec.
All other regression coefficients will be re-estimated for each sliding span. Trading day
(td), holiday (holiday, outlier (outlier), or other user-defined (user) regression effects
can be fixed. This argument is ignored if neither a regARIMA model nor an irregular
component regression is fit to the series, or if fixmdl=yes.
length The length of each span, in months or quarters (in accordance with the sampling interval)
of time series data used to generate output for comparisons. A length selected by the
user must yield a span greater than 3 years long and less or equal to 19 years long. If
the length of the span is not specified by the user, the program will choose a span length
based on the length of the seasonal filter selected by the user (or by the program if a
seasonal filter was not specified by the user) when the seasonal adjustment is performed
by the x11 spec, or by the level of the seasonal MA parameter coefficient (T heta), when
the seasonal adjustment is performed by the seats spec. For more information, see
DETAILS. Monthly data example: length=96
numspans Number of sliding spans used to generate output for comparisons. The number of spans
selected by the user must be between 2 and 4, inclusive. If this argument is not specified
by the user, the program will choose the maximum number of spans (up to 4) that can
be formed based on the length of the sliding spans given by the user (or selected by the
program if the length argument is not used). Example: numspans=4
outlier Specifies whether automatic outlier detection is to be performed whenever the regARIMA
model is re-estimated during the processing of each span. This argument has no effect if
the outlier spec is not used.
If outlier=keep, the program carries over any outliers automatically identified in the
original estimation of the regARIMA model for the complete time series, and does not
perform automatic outlier identification when a regARIMA model is estimated for one
of the sliding spans. If the date of an outlier detected for the complete span of data does
not occur in one of the sliding spans, the outlier will be dropped from the model for that
span. This is the default setting.
If outlier=remove, those outlier regressors that were added to the regression part of the
regARIMA model when automatic outlier identification was performed on the full series
are removed from the regARIMA model during the sliding spans analysis. Consequently,
their effects are not estimated and removed from the series. If outlier terms are included
in the regression spec, these will included in the model estimated for the spans. This
option gives the user a way to investigate the consequences of not doing automatic outlier
identification.
If outlier=yes, the program performs automatic outlier identification whenever a reg-
ARIMA model is estimated for a span of data.
see Appendix B.
savelog The only diagnostic available for output to the log file (see Section 2.6) is the percentage
of observations flagged as unstable for each of the estimates from the seasonal adjustment
estimates tested by the sliding spans analysis.
Specifying savelog=percents or savelog=pct will store this information into the log
file.
start The starting date for sliding spans comparisons. The default is the beginning month of
the second span. Example: start=1990.jan
additivesa Specifies whether the sliding spans analysis of an additive seasonal adjustment will be
calculated from the maximum differences of the seasonally adjusted series (additivesa
= difference) or from the maximum of an implied adjustment ratio of the original series
to the final seasonally adjusted series (additivesa = percent). This option will also
determine if differences (additivesa = difference) or percent changes (additivesa
= percent) are generated in the analysis of the month-to-month, quarter-to-quarter,
or year-to-year changes in seasonally adjusted series. The default is additivesa =
difference. If the seasonally adjusted series for any of the spans contains values that are
less than or equal to zero, the sliding spans analysis will be performed on the differences.
fixx11reg Specifies whether the irregular component regression model will be re-estimated during
the sliding spans analysis, if one is specified in the x11regression spec. If fixx11reg=yes,
the regression coefficients of the irregular component regression model are fixed through-
out the analysis at the values estimated from the entire series. If fixx11reg=no, the
irregular component regression model parameters will be re-estimated for each span.
The default is fixx11reg=yes.
x11outlier Specifies whether the AO outlier identification will be performed during the sliding spans
analysis for the irregular component regression specified in the x11regression spec. If
x11outlier=yes, AO outlier identification will be done for each span. Those AO outlier
regressors that were added to the irregular component regression model when automatic
Table 7.37: Default Output Tables for Slidingspans Spec

header hdr · header text for the sliding spans analysis
ssftest ssf · F-tests for stable and moving seasonality esti-
mated over each of the sliding spans
factormeans fmn · range analysis for each of the sliding spans
percent pct · table showing the percent of observations flagged
as unstable for the seasonal and/or trading day
factors, final seasonally adjusted series (if nec-
essary), and the month-to-month (or quarter-to-
quarter) changes
summary sum · tables, histograms and hinge values summarizing
the percentage of observations flagged for unstable
seasonal and/or trading day factors, final season-
ally adjusted series (if necessary), and month-to-
month (or quarter-to-quarter) changes
yysummary smy + additional tables, histograms and hinge val-
ues summarizing the percentage of observations
flagged for the year-to-year changes
indfactormeans fmi · range analysis for the implicit adjustment factors
of the indirectly seasonally adjusted series
indpercent pci · tables of the percent of observations flagged as
unstable for the seasonal factors and month-to-
month (or quarter-to-quarter) changes of the in-
direct seasonal adjustment
indsummary smi · tables, histograms and hinge values summarizing
the percentage of observations flagged for unstable
seasonal factors, month-to-month (or quarter-to-
quarter) and year-to-year changes for the indirect
adjustment
Table 7.38: Other Output Tables for Slidingspans Spec

yypercent pcy · additional entry for the percent of observations
flagged as unstable for the year-to-year changes
sfspans sfs + seasonal factors from all sliding spans
chngspans chs + month-to-month (or quarter-to-quarter) changes
from all sliding spans
saspans sas + seasonally adjusted series from all sliding spans
ychngspans ycs + year-to-year changes from all sliding spans
tdspans tds + trading day factors from all sliding spans
indyypercent pyi · additonal entry for the percent of observations
flagged as unstable for the year-to-year (or
quarter-to-quarter) changes of the indirect sea-
sonal adjustment
indyysummary syi · additional tables, histograms and hinge val-
ues summarizing the percentage of observations
flagged for the year-to-year changes of the indi-
rect seasonal adjustment
indsfspans sis + indirect seasonal factors from all sliding spans
indchngspans cis + indirect month-to-month (or quarter-to-quarter)
changes from all sliding spans
indsaspans ais + indirect seasonally adjusted series from all sliding
spans
indychngspans yis + indirect year-to-year changes from all sliding spans
AO outlier identification was done for the full series are removed from the irregular
component regression model prior to the sliding spans run. If x11outlier=no, then the
automatically identified AO outlier regressors for the full series are kept for each sliding
spans run. If the date of an AO outlier detected for the complete span of data does not
occur in one of the sliding spans, the outlier will be dropped from the model for that span.
The coefficients estimating the effects of these AO outliers are re-estimated whenever the
other irregular component regression model parameters are re-estimated. However, no
additional AO outliers are automatically identified and estimated. This option is ignored
if the x11regression spec is not used, if the selection of the aictest argument results
in the program not estimating an irregular component regression model, or if the sigma
argument is used in the x11regression spec. The default is x11outlier=yes.
DETAILS
This section provides some additional information about the arguments within the sliding spans spec. Section
6.2 contains a description of the sliding spans diagnostics and their interpretation. For more details on the
sliding spans procedure, see Findley, Monsell, Shulman, and Pugh (1990).
Different adjustment quantities are examined in a sliding spans analysis, depending on the mode of the
seasonal adjustment and whether trading day adjustment is done. For a multiplicative or log-additive seasonal
adjustment, the seasonal factors, and the month-to-month and year-to-year changes of the seasonally adjusted
series are analyzed. For a multiplicative or log-additive seasonal and trading day adjustment, the trading
day factors and seasonally adjusted series are analyzed as well. For an additive seasonal adjustment without
trading day adjustment, the seasonally adjusted series and the month-to-month and year-to-year changes of the
seasonally adjusted series are analyzed. If trading day adjustment is done, these analysis are performed for the
seasonal and trading day adjusted series.
WARNING: In the additive adjustment case, the presence of relatively small values or negative values in
the adjusted series can render unusable the percent change values which are the basis of almost all of the sliding
spans statistics. In this situation, usually only a subjective analysis of the spans of adjusted series obtained by
using saspan in the print or save arguments can be used to detect excessive instability. Further research is
needed to develop more useful sliding spans statistics for additive adjustments.
One important choice that needs to be made in a sliding spans analysis is the length of the overlapping
spans. When used with the x11 spec, the length of the span is based on the length of the seasonal filter since
here seasonal adjustment is performed with fixed length seasonal filters. Table 7.39 gives a listing of how long
the sliding span is by default for different seasonal filters of the x11 spec.
Seasonal Length of Span

filter (in years)
3-term 6
3x3 6
3x5 8
3x9 11
3 x 15 17
Table 7.39: Sliding span lengths for different seasonal filters chosen
Note that additional observations are added to the length of the span by default so that each of the sliding
spans begin in January (or the first quarter for quarterly series). If different seasonal filters are used for different
months or quarters, the longest span length of the seasonal filters chosen will be used.
If a stable seasonal filter is selected, the program will set the span to by dividing the length of the series by
the number of spans used (with the default number of spans set to 4). The length of the span can be no more
than 17 years long, and no shorter than 3 years long.
For model-based seasonal adjustments, the ARIMA model-based seasonal adjustment filters generated in
the seats are always as long as the data span being adjusted (when the ARIMA model specified has a moving
average component). Findley, Wills, Aston, Feldpausch, and Hood (2003) develops an approach for determining
span lengths that is based on an analysis of SEATS model-based adjustment filters associated with the airline
model, the model chosen for about half the series adjusted by SEATS, see Gómez and Maravall (1996). Since
values of θ and Θ are known for which the SEATS seasonal adjustment filters have gain and phase-shift properties
very close to those of the X-11 filters, as shown in Planas and Depoutot (2002) and Findley and Martin (2006),
the sliding span lengths used for SEATS adjustments within X-13A-S are calibrated to coincide with the span
lengths used for the X-11 filters when the two types of filters are close. In this way, the span length specifications
used for SEATS adjustments are anchored to those of the X-11 filters.
The table below gives the span length used by the program for a given value of Θ. Research of the type
described in Feldpausch, Hood, and Wills (2004) showed that for simulated series with known components, using
the sliding spans lengths based on the seasonal moving average parameter seemed to provide a more reliable
indication of inaccuracy in the seasonal adjustment than other diagnostics commonly used with SEATS seasonal
adjustments.
Seasonal Length of Span

MA (in years)
0.160 5
0.325 6
0.490 7
0.535 8
0.620 9
0.640 10
0.695 11
0.710 12
0.750 13
0.760 14
0.795 15
0.805 16
0.840 17
0.850 18
0.910 19
Table 7.40: Seasonal MA parameter greater than 0 at which the span length increases to the value indicated.
If automatic ARIMA modeling selected is selected by either the automdl or pickmdl spec, then the model
selected by the procedure is used for all the sliding spans. If no model is selected by the procedure, then no
model will be estimated during the sliding spans analysis.
While many of the tables in this spec cannot be saved as individual files, specifying the seasonal adjustment
diagnostic summary option with the -s flag flag at runtime allows the user to store information from the sliding
spans analysis into a diagnostic summary file (with the file extension .udg). In addition, the savelog argument
can write selected diagnostics into the log file for a given run (with the file extension .log). For more information,
see Section 2.6.
If a sliding spans analysis of the direct and indirect adjustments of a composite seasonal adjustment is
desired, the sliding spans analysis option must be specified for each of the component series. If the seasonal
filter length is not the same for each component, then the user will have to use the length argument defined
above in each of the input files of the component series to ensure that the spans analyzed for these series are of
the same length.
If the automatic seasonal filter selection option is used, the seasonal filters used to generate the original
seasonal adjustment will be used for the seasonal adjustment of each of the spans.
If an outlier specified by the user does not occur in a given span, that outlier will be dropped from the model
for that span, and will be re-introduced into the regARIMA model if it is defined in future spans. User-defined
regressors are checked to see if they are not constant in each span (i.e., all values of the regressor equal to zero).
EXAMPLES
Example 1 Multiplicative monthly seasonal adjustment, 3x9 seasonal factors for all calendar months.
Sliding spans analysis performed with default settings for all options.
SERIES { FILE = "TOURIST.DAT" START = 1976.1 }

SLIDINGSPANS { }
Example 2 Log-additive seasonal adjustment of quarterly data, 3x9 seasonal filters for first two quarters,
3x5 seasonal filters for last two quarters, 7-term Henderson trend filter. Sliding spans analysis
performed with threshold values for selected tests set to 5.0.
Series {
File = "qstocks.dat"
Start = 1967.1
Title = "Quarterly stock prices on NASDAC"
Freq = 4
}
X11 {
Seasonalma = ( S3x9 S3x9 S3x5 S3x5 )
Trendma = 7
Mode = Logadd
}
Slidingspans {
cutseas = 5.0
cutchng = 5.0
}
Example 3 Seasonal ARIMA model with regression variables used for trading day adjustment and for
automatic outlier identification and adjustment. Specified regression variables are a con-
stant, trading day effects, and ramp between May 1982 and Sept. 1982. The ARIMA part
of the model is (0,1,2)(0,1,1)12. Additively seasonally adjust the series after preadjusting
for the outliers and the trading day regression effects. Perform sliding spans analysis; incor-
porate any outliers found by the application of the automatic identification procedure to the
full series into the regARIMA model re-estimated for each of the sliding spans. The length
of the sliding spans is set so that the sliding spans statistics from this run can be compared
to a SEATS seasonal adjustment to be done in the next example.
series { title = "Number of employed machinists - X-11"

start = 1980.jan file = "machine.emp"
}
regression { variables = (const td rp82.may-82.oct) }
arima { model = (0 1 2)(0 1 1) }
outlier { }
estimate { }
check { }
forecast { }
x11 { mode = add save = d11}
slidingspans { outlier = keep
length = 144
}
Example 4 Same as example 3, except that a model-based seasonal adjustment is performed using the
seats spec
series { title = "Number of employed machinists - SEATS"

start = 1980.jan file = "machine.emp"
}
regression { variables = (const td rp82.may-82.oct) }
arima { model = (0 1 2)(0 1 1) }
outlier { }
estimate { }
check { }
forecast { }
seats { save = s11 }
slidingspans { outlier = keep
length = 144
}
Example 5 The predefined regression effects to be estimated are a constant, trading day and a fixed
seasonal. The ARIMA part of the model is (3, 1, 0). Generate 60 forecasts. Seasonally adjust
the series after pre-adjusting for the estimated trading day. Perform sliding spans analysis.
Re−estimate the values of the REGARIMA model parameters for each span.
series { title = "Cheese Sales in Wisconsin"

file = "cheez.fil" start = 1975.1 }
regression { variables = (const seasonal tdnolpyear) }
forecast { maxlead = 60 }
x11 { save = seasonal appendfcst = yes }
slidingspans { fixmdl = no }
Example 6 Sliding spans analysis will be performed on the multiplicative seasonal adjustment specified,
using 3 sliding spans of length 40 quarters as specified. This would allow the user to get some
indication of seasonal adjustment stability, even though the series is not long enough for a
complete sliding spans analysis with spans of the length most appropriate for 3x9 seasonal
filters (44 quarters).
Series {
File = "qstocks.dat"
Start = 1987.1
Title = "Quarterly stock prices on NASDAC"
Freq = 4
}
X11 {
Seasonalma = S3x9
}
Slidingspans {
Length = 40
Numspans = 3
}
7.17 SPECTRUM
DESCRIPTION
Optional spec that provides X-13ARIMA-SEATS with options related to spectral plots for monthly series. Users
can set the starting date of the span of data to be used to estimate the spectra (start) and the type of estimate
to be generated (type). For more information on the spectrum diagnostic, see Section 6.1.
USAGE
spectrum{ start = 1975.Jan
type = periodogram
print = (none + specsa + specirr)
save = (spn)
savelog = peaks
}
ARGUMENTS
the default printout, except seriesplot and adjoriginalplot. For a complete listing
of the brief and default print levels for this spec, see Appendix B.
7.42.
start The starting date of the span of data to be used to estimate the spectra the original,
seasonally adjusted, and modified irregular series. This date must be in the format
start=year.seasonal period . This can be used to determine if there are residual trading
day or seasonal effects in the adjusted data from, say, the last seven years. Residual effects
can occur when seasonal or trading day patterns are evolving. The default starting date
for the spectral plots is set to be 96 observations (8 years of monthly data) from the end
of the series. If the span of data to be analyzed is less than 96 observations long, it is set
to the starting date of this span of data. Example: start=1987.Jan.
type The type of spectral estimate used in the spectral plots output by the program. If type
= periodogram, the periodogram of the series is calculated and plotted. The default
(type=arspec) produces an autoregressive model spectrum of the series.
axis If axis=same, the spectra for the differenced original series and the differenced seasonally
adjusted series modified for extremes will be displayed using the same axis. If axis=diff,
the spectral plots will not be so constrained. The default is axis=diff.
7.17. SPECTRUM 189
Table 7.41: Available Output Tables for Spectrum

specorig sp0 + spectral plot of the first-differenced original series
specsa sp1 + spectral plot of differenced, X-11 seasonally ad-
justed series (or of the logged seasonally adjusted
series if mode = logadd or mode = mult)
specirr sp2 + spectral plot of outlier-modified X-11 irregular
series
specseatssa s1s + spectrum of the differenced final SEATS seasonal
adjustment
specseatsirr s2s + spectrum of the final SEATS irregular
specextresiduals ser + spectrum of the extended residuals
specresidual spr + spectral plot of the regARIMA model residuals
speccomposite is0 + spectral plot of first-differenced aggregate series
specindirr is2 + spectral plot of the first-differenced indirect sea-
sonally adjusted series
specindsa is1 + spectral plot of outlier-modified irregular series
from the indirect seasonal adjustment
Table 7.42: Available Log File Diagnostics for Spectrum

peaks spk Visually significant peaks in spectra
dirpeaks dpk Visually significant peaks in spectra for direct seasonal adjust-
ment
indpeaks ipk Visually significant peaks in spectra for indirect seasonal ad-
justment
decibel Determines whether spectral estimates will be expressed in terms of decibel units decibel
= yes, as shown in equation (6.1). The estimates are plotted on the untransformed scale
if decibel = no. The default is decibel = yes.
difference If difference=no, the spectrum of the (transformed) original series or seasonally ad-
justed series is calculated; if difference=first, the spectrum of the month-to-month
differences of these series is calculated. The default (difference=yes) will apply a
max(d + D − 1, 1) difference to the (transformed) original series and seasonally adjusted
series before computing the spectrum, where d is the order of regular differencing and D
is the order of seasonal differencing in the regARIMA model specified for the series. If
no regARIMA model is specified, the default order of differencing is 1.
maxar An integer value used to set the maximum order of the AR spectrum used as the default
type of spectrum plot. Integers from 1 to 30 are acceptable values for maxar. If this
option is not specified, the maximum order will be set to 30.
peakwidth Allows the user to set the width of the band used to determine spectral peaks. The
default value is peakwidth = 1.
series Allows the user to select the series used in the spectrum of the original (or composite)
series (Table G.0). Table 7.43 shows the series that can be specified with this argument
- the default is series = adjoriginal (or b1).
Table 7.43: Choices Available for the series Argument

original a1 original series
outlieradjoriginal a19 original series, adjusted for regARIMA outliers
adjoriginal b1 original series, adjusted for user specified and reg-
ARIMA prior effects
modoriginal e1 original series modified for extremes
Name gives the name of each series which can be specified for use with the series
arguments.
Short gives a short name for the tables of the series argument.
7.17. SPECTRUM 191
Note that if the x11 spec is not specified, the original series modified for extremes will
not be generated; the setting series = modoriginal will be ignored, and the default
setting will be used instead.
siglevel Sets the significance level for detecting a peak in the spectral plots. The default is
siglevel = 6.
DETAILS
A routine searches each of the spectra for peaks at the seasonal and trading day frequencies. A warning message
is printed out if visually significant peaks are found, and the plot in which a peak was found is printed out.
When the restricted output (the -n flag) option is used, the plot is not produced in the main output, but a
message is printed suggesting that the user rerun the program without the -n flag.
For more information on the spectrum diagnostic, see Section 2.1 of Findley, Monsell, Bell, Otto, and Chen
(1998) and Soukup and Findley (1999) as well as Section 6.1.
EXAMPLES
Example 1 This example shows how to obtain a spectrum plot of the first differences (month-to-month
differences) of the logarithms of the series to check if the series has seasonal or trading day
effects. This is a complete spec file.
series{ title = "Spectrum analysis of Building Permits Series"

start = 1967.Jan
file = "permits.dat"
format = "(12f6.0)"
print = none
}
transform{
function = log
print = none
}
spectrum{
start = 1987.Jan
print = (none +specorig)
savelog = peaks
}
Example 2 This spec file for a composite seasonal adjustment will generate the periodogram rather than
the AR-spectrum, and will store information on whether there are peaks in the indirect sea-
sonal adjustment. The periodogram for the indirect seasonal adjustment will be saved. This
is a complete spec file, but must be run with other spec files in a metafile; more information
on composite seasonal adjustment is available in Section 7.4.
composite { title="TOTAL ONE-FAMILY Housing Starts"

name="C1FTHS" save=(indseasonal) }
title="Composite adj. of 1-Family housing starts"
save=(D10) }
spectrum { savelog = indpeaks
type = periodogram
save = is1 }
7.18. TRANSFORM 193
7.18 TRANSFORM
DESCRIPTION
Specification used to transform or adjust the series prior to estimating a regARIMA model. With this spec
the series can be Box-Cox (power) or logistically transformed, length-of-month adjusted, and divided by user-
defined prior-adjustment factors. Data for any user-defined prior-adjustment factors must be supplied, either
in the data argument, or in a file specified by the file argument (not both). For seasonal adjustment, a set of
permanently removed factors can be specified and also a set of factors that are temporarily removed until the
seasonal factors are calculated.
USAGE
transform{ function = log or power = 0.0

adjust = lom
title = "prior adjustment factors"
start = 1975.jan
data = (1.25 · · · 1.90) or file = "prioradj.dat"
format = "(6f12.3)"
name = "Adjfac"
aicdiff = 0.0
mode = ratio
type = temporary
print = (none)
save = (prioradj)
savelog = atr
}
ARGUMENTS
adjust Perform length-of-month adjustment on monthly data (adjust = lom), length-of-quar-

ter adjustment on quarterly data (adjust = loq), or leap year adjustment of monthly
or quarterly data (adjust = lpyear). (See DETAILS.)
Do not use the adjust argument if td or td1coef is specified in the variables argument
of the regression or x11regression specs, or if additive or pseudo-additive seasonal
adjustment is specified in the mode argument of the x11 spec. Leap year adjustment
(adjust = lpyear) is only allowed when a log transformation is specified in either the
power or function arguments.
aicdiff Defines the difference in AICC needed to accept no transformation when the auto-
matic transformation selection option is invoked (function=auto). The default value
is aicdiff = -2.0 for monthly and quarterly series, aicdiff = 0.0 otherwise. For
more information on how this option is used to select a transformation see DETAILS.
data An array containing one or two series of preadjustment factors which, unless mode=diff
(see below), must have positive values intended for division into the corresponding values
of the input time series. The default value is a vector of ones (no prior adjustment). When
data (or file) is used, an adjustment factor must be supplied for every observation in
the series (or for the span specified by the span argument of the series spec, if present).
Generally, an adjustment factor must also be supplied for each forecast (and backcast)
desired. (See DETAILS.) The adjustment factors are read in free format. If a start date
is supplied for the adjustment factors, then they may start before the beginning of the
series. If the data argument is used, the file argument cannot be used. When mode =
diff, the values in data are subtracted from the series, and they need not be positive.
Two series can be input via the data argument when both permanent and temporary
prior-adjustment factors are specified in the type set - see DETAILS for more informa-
tion.
file Name of the file containing the user-defined prior-adjustment factors. The filename must
be enclosed in quotes. If the file is not in the current directory, the path must also
be given. If the file argument is used, the data argument cannot be used. The value
restrictions are the same as for data. If the data in the file are not in free format, the
format argument must be used.
If both permanent and temporary prior-adjustment factors are specified in the type set,
the factors can be input from a single file or from two files - see DETAILS for more
information.
format Denotes the format used to read the prior adjustment factors from a file. Five types of
input are accepted:
b. a valid Fortran format, which should be enclosed in quotes and must include the
initial and terminal parentheses (example: format="(6f12.0)").
c. a two character code which corresponds to a set of data formats used in previous
versions of X-11 and X-11-ARIMA (example: format="1r");
d. “datevalue” format, where the year, month or quarter, and value of each observation
are found in this order in free format on individual lines. Thus, a line of the data
file containing the value 32531 for July of 1991 would have the form 1991 7 32531.
The number of preceding blanks can vary (example: format="datevalue");
e. the format X-13ARIMA-SEATS uses to save a table. This allows the user to read in a
file saved from a previous X-13ARIMA-SEATS run (example: format="x12save");
f. the format that the TRAMO and SEATS programs use to read in a series and its
format="tramo");
g. a variant of “free” format where the numbers must be separated by one or more
7.18. TRANSFORM 195
h. a variant of “datevalue” format, where the year, month or quarter, and value of
of preceding blanks can vary (example: format="datevaluecomma");.
In the predefined X-11 data formats, the data is stored in 6 or 12 character fields, with
a year and series label associated with each year of data. For a complete list of these
formats, see Table 7.36 in the DETAILS section of the series spec.
If no format argument is given, the data will be read in free format. Format can only
be used with the file argument, not with data.
If permanent and temporary prior-adjustment factors are input from two different files
with distinct formats, then up to two formats can be specified - see DETAILS for more
information.
function Transform the series Yt input in the series spec using a log, square root, inverse, or
logistic transformation. Alternatively, perform an AIC-based selection to decide between
a log transformation and no transformation (function=auto) using either the regARIMA
model specified in the regression and arima specs or the airline model (0 1 1)(0 1 1) (see
DETAILS). The default is no transformation (function = none). Do not include both
the function and power arguments. Note: there are restrictions on the values used in
these arguments when preadjustment factors for seasonal adjustment are generated from
a regARIMA model; see DETAILS.
Table 7.44: Transformations Available Using the function Argument
value transformation range for Yt equivalent power argument

none Yt all values power = 1
log log(Y
√ t) Yt > 0 for all t power = 0
1
sqrt 4 + 2( Yt − 1) Yt ≥ 0 for all t power = 0.5
inverse 2 − Y1t Yt 6= 0 for all t power = -1
Yt
logistic log( 1−Y t
) 0 < Yt < 1 for all t none equivalent
mode Specifies the way in which the user-defined prior adjustment factors will be applied to
the time series. If prior adjustment factors to be divided into the series are not given
as percents (e.g., (100 100 50 · · · )), but rather as ratios (e.g., (1.0 1.0 .5 · · · )), set
mode=ratio. If the prior adjustments are to be subtracted from the original series, set
mode=diff. If mode=diff is used when the mode of the seasonal adjustment is set to be
multiplicative or log additive in the x11 spec, the factors are assumed to be on the log
scale. The factors will be exponentiated to put them on the same basis as the original
series. If this argument is not specified, then the prior adjustment factors are assumed
to be percents (mode=percent).
If both permanent and temporary prior-adjustment factors are specified in the type
argument, then up to two values can be specified for this argument, provided they are
compatible (e.g., diff cannot be specified along with ratio or percent). See DETAILS
for more information.
name The name of the prior adjustment factors. The name must be enclosed in quotes and
may contain up to 64 characters. Up to the first 16 characters will be printed as a label
for the prior adjustment factors. When specified with the X-11 formats of the format
argument, the first six (or eight, if format="cs") characters of this name are also used
with the predefined formats to check that the program is reading the correct series, or
to find a particular series in a file where many series of factors are stored.
then the user can either specify series names for both sets of prior adjustment factors, or
no name should be entered - see DETAILS for more information.
power Transform the input series Yt using a Box-Cox power transformation,
(
log(Yt ) λ = 0;
Yt −→ yt = 2 λ
λ + (Yt − 1)/λ λ 6= 0.
This formula for the Box-Cox power transformation is constructed so that its values will
be close to Yt when λ is near 1 and close to logYt when λ is near zero. It also has the
property that the transformed value is positive when Yt is greater than 1.
The power λ must be given (e.g., power = .33). The default is no transformation (λ =
1), i.e., power = 1. The log transformation (power = 0), square root transformation
(power = .5), and the inverse transformation (power = -1) can alternatively be given
using the function argument. Do not use both the power and the function arguments
in the same spec file. Note: there are restrictions on the values used in these arguments
when preadjustment factors for seasonal adjustment are generated from a regARIMA
model; see DETAILS.
precision The number of decimal digits to be read from the file of prior adjustment factors. This
option can only be used with the predefined formats of the format argument. This value
must be an integer between 0 and 5, inclusive (for example, precision=5). The default
is zero. If precision is used in a transform spec that does not use one of the predefined
formats, the argument is ignored.
If both permanent and temporary prior-adjustment factors are specified in the type
argument, then up to two values can be specified for this argument - see DETAILS for
more information.
print and save Table 7.45 gives the available output tables for this spec. The aictransform, prior,
and prioradjusted tables are printed out by default For a complete listing of the brief
and default print levels for this spec, see Appendix B.
savelog Setting savelog=autotransform or savelog=atr causes the result of the automatic
transformation selection procedure to be output to the log file (see section 2.6 for more
information on the log file).
start The start date of the user-defined prior-adjustment factors. The default is the start date
of the series. Valid values are any date up to the start date of the series (or up to the
start date of the span specified by the span argument of the series spec, if present).
then up to two starting dates can be specified to read in the two sets of prior adjustment
factors - see DETAILS for more information.
7.18. TRANSFORM 197
Table 7.45: Available Output Tables for Transform

aictransform tac · output from AIC-based test(s) for transfor-
mation
seriesconstant a1c + original series with value from the constant
argument added to the series
seriesconstantplot acp · plot of original series with value from the
constant argument added to the series
prior a2 + prior-adjustment factors, with associated
dates
permprior a2p + permanent prior-adjustment factors, with
associated dates
tempprior a2t + temporary prior-adjustment factors, with
associated dates
prioradjusted a3 + prior-adjusted series, with associated dates
permprioradjusted a3p + prior-adjusted series using only permanent
prior factors, with associated dates
prioradjustedptd a4d + prior-adjusted series (including prior trad-
ing day adjustments), with associated
dates
permprioradjustedptd a4p + prior-adjusted series using only permanent
prior factors and prior trading day adjust-
ments, with associated dates
transformed trn + prior-adjusted and transformed data, with
associated dates
title A title for the set of user-defined prior-adjustment factors. The title must be enclosed in
quotes and may contain up to 79 characters.
type Specifies whether the user-defined prior-adjustment factors are permanent factors (re-
moved from the final seasonally adjusted series as well as the original series) or temporary
factors (removed from the original series for the purposes of generating seasonal factors
but not from the final seasonally adjusted series). If only one value is given for this ar-
gument (type = temporary), then only one set of user-defined prior-adjustment factors
will be expected. If both types of user-defined prior-adjustment factors are given (type
= (temporary permanent)), then two sets of prior adjustment factors will be expected,
for more information see DETAILS. The default is type = permanent.
constant Positive constant value that is added to the original series before the program models
or seasonally adjusts the series. Once the program finishes modeling and/or seasonally
adjusting the series with the constant value added, this constant is removed from the
seasonally adjusted series as well as the trend component.
trimzero If trimzero=no, zeroes at the beginning or end of a time series entered via the file
argument are treated as series values. If trimzero=span, causes leading and trailing
zeros to be ignored outside the span of data being analyzed (the span argument of the
series spec must be specified with both a starting date and an ending date). The default
(trimzero=yes) causes leading and trailing zeros to be ignored. Note that when the
format argument is set to either datevalue, x12save, or tramo, all values input are
treated as series values, regardless of the value of trimzero.
DETAILS
If a Box-Cox or logistic transformation is specified in conjunction with a length-of-month (or leap year) adjust-
ment and/or user-defined prior-adjustment factors, the time series is first adjusted for length-of-month and/or
prior factors, and then Box-Cox or logistically transformed. If both length-of-month and prior-adjustment fac-
tors are specified, then combined adjustment factors (length-of-month × prior adjustment) are used. Length-
of-quarter and leap year adjustments are handled in the same way.
If either lom and loq of the adjust argument is specified, the correct adjustment factor is determined by the
period specified in the series spec. In the case of a monthly input series Yt , each observation is first divided
by the number of days in that month (mt ) and then multiplied by the average length-of-month (30.4375),
resulting in (30.4375 × Yt )/mt . Length-of-quarter adjustments are performed in a similar manner, resulting in
(91.3125 × Yt )/qt , where qt is the length in days of quarter t. Forecasts of the transformed and length-of-month
adjusted data are transformed back to the original scale for output (see the documentation of the forecast
spec).
If adjustment factors are supplied for the forecast period, then forecasts of the prior-adjusted series will
be inverse-transformed (multiplied or, if mode = diff, added to) with these factors. If adjustment factors are
7.18. TRANSFORM 199
not supplied for the forecast period, then inverse-transformation of forecasts will only account for a Box-Cox or
logistic transformation and for any length-of-month (or length-of-quarter) adjustment —this effectively assumes
values of 1 for the user-defined prior-adjustment factors throughout the forecast period (or 0 if mode = diff).
When seasonal adjustment is requested with the x11 spec, any value of power or function can be used
for the purpose of forecasting the series with a regARIMA model. However, this is not the case when factors
generated from the regression coefficients are used to adjust either the original series or the final seasonally
adjusted series. In this case, the only accepted transformations are the log transformation (for multiplicative
or log-additive seasonal adjustments) and no transformation, which can be specified as power = 1 (for additive
seasonal adjustments).
This restriction of the transformation done because factors derived from the regression coefficients must
be the same type as factors generated by the seasonal adjustment procedure, so that combined adjustment
factors can be derived and adjustment diagnostics can be generated. If the regARIMA model is applied to a
log-transformed series, the regression factors are expressed in the form of ratios, which is the same form as
the seasonal factors generated by the multiplicative (or log-additive) adjustment modes. Conversely, if the reg-
ARIMA model is fit to the original series, the regression factors are measured on the same scale as the original
series, which matches the scale of the seasonal factors generated by the additive adjustment mode.
If no seasonal adjustment is done, any power transformation can be used.
When function=auto and the series being processed has all positive values, the program will choose between
no transformation and a log transformation by fitting a regARIMA model to the untransformed and transformed
series and choosing the log transformation except when
AICCnolog − AICClog < ∆AICC or AICClog + ∆AICC > AICCnolog (7.17)
where AICClog is the value of AICC from fitting the regARIMA model to the transformed series, AICCnolog
is the value of AICC from fitting the regARIMA model to the untransformed series, and ∆AICC is the value
entered for the aicdiff argument, with a default of -2. Negative values of ∆AICC bias the selection in favor of
the log transformation. The default of -2 is used not for statistical reasons but for convenience. Multiplicative
adjustment is appropriate for the great majority of Census Bureau series and we do not wish to inconvenience
users accustomed to multiplicative adjustments unless the statistical support for additive adjustment is rather
strong. The AICC values for the log transformed series (or any transformed series) is obtained by applying
an appropriate (Jacobian) adjustment to the log likelihood to make it compatible with the log likelihood of the
estimated model of the untransformed series. (The adjustment is printed in the output if print=lkf is specified
in the estimate spec.) If the series has a zero or negative value, no transformation is used.
If a regARIMA model has been specified in the regression and/or arima specs, then the procedure will
use this model to generate the AICC statistics needed for the test. If no model is specified, or the automatic
model identification procedure is specified via the automdl or pickmdl spec, the program will use the airline
model ((0 1 1)(0 1 1) in Box-Jenkins notation) and use this model to generate the AICC statistics.3
If seasonal adjustment is specified via the x11 or x11regression spec, the program will set the seasonal ad-
justment mode to one that is appropriate for the transformation selected (multiplicative for a log transformation,
additive for no transformation).
3 Note that if only the regression spec is specified, the X-13ARIMA-SEATS default ARIMA model is the (0 0 0)(0 0 0) model (in
Box-Jenkins notation). In other words, if the regression model includes trading day, but no ARIMA model is specified, then the
program will use a (0 0 0)(0 0 0) ARIMA model and trading day regressors to generate the AICC statistics.
The program currently does not allow the use of user-defined prior-adjustment factors with the automatic
transformation selection option.
Users specifying both temporary and permanent user-defined prior-adjustment factors must take advantage
of some special features built into the transform spec. For the arguments related to data input, the user
can specify an entry for each type of prior-adjustment factor. The type variable tells the program which type
of prior factor is being referred to by a given entry. For example, in the input specified below, the series of
temporary prior-adjustment factors is read from temp.fil using a (6F12.5) format. These factors start in
January of 1980. The series of permanent prior-adjustment factors, which starts in July of 1975, is read from
perm.fil using a (F15.3) format.
transform{
type=(temporary permanent)
file=("temp.fil" "perm.fil")
format=("(6F12.5)" "(F15.3)")
start=(1980.jan 1975.jul)
mode=(ratio percent)
}
If two entries are given for the file arguments but only one entry for each of the format, start, mode
and precision arguments, then the values given are assumed to apply to both sets of factors. The number of
values given for the name argument must match the number of prior-adjustments factors implied by the type
argument.
When the data argument is used to input two sets of prior-adjustment factors, the data is assumed to be
a matrix of two columns. The type assignment for the data columns is implied by the type argument. In the
example below, the first column of data is interpreted to be a temporary prior-adjustment factor (with values
of 1.055, 0.990, and 1.025), and the second column of data is interpreted to be a permanent prior-adjustment
factor.
transform{
type=(temporary permanent)
data=( 1.055 1.000
0.990 1.000
1.025 1.000
. .
. .
. .
1.033 1.000 )
start=1980.jan
mode=ratio
}
The same assumption is made when only one data file is given for two adjustment types, as in the input
below.
7.18. TRANSFORM 201
transform{
type=(temporary permanent)}
file="both.fil"}
start=1980.jan}
mode=ratio}
}
X-13ARIMA-SEATS accepts a maximum of 780 user-defined prior-adjustment factors of each type (temporary
or permanent), not including the forecast period. (This limit can be changed—see Section 2.8.)
The constant argument is sometimes useful when a series has a number of values either close to zero where
neither multiplicative nor additive seasonal adjustment seems to be effective, or in the case where a series has
zero or negative values and the series seems to behave in a manner that usually calls for a multiplicative seasonal
adjustment. Strategies for selecting the value of constant, as well as an application to Canadian air travel
series, can be found in Chen and Durk (2005).
EXAMPLES
Note: The following examples do not show “complete” spec files. Useful output is not produced unless addi-
tional specs (e.g., x11, identify, or arima and estimate) are added.
Example 1 Specify a user-defined prior adjustment for a strike in March and April of 1967, as well as a
length-of-month adjustment.
series { data = (879 899 462 670 985 973 ...)

start = 1967.jan }
transform { data = (1 1 .5 .75 1 1 ...)
mode = ratio
adjust = lom }
Example 2 Specify a constant to add to the series before modeling and seasonal adjustment. Use the
automatic transformation selection procedure to determine if a log transformation should
be used to transform the resulting series.
series { data = (6 79 98 42 4 73 85 26 ...)

start = 1997.1 period=4 }
transform { constant=45 function = auto }
Example 3 Specify a logarithmic transformation and also a user-defined adjustment by a price deflator
that changes current dollars to constant (real) dollars. A start date is specified for the
deflator series since it begins before the time series being modelled.
series {title = "Total U.S. Retail Sales --- Current Dollars"

file = "retail.dat"
start = 1980.jan }
transform {function = log
title = "Consumer Price Index"
start = 1970.jan # adj. factors start January, 1970
file = "cpi.dat"
format = "(12f6.3)" }
Example 4 Same as Example 3, only a pre-defined format is used to read in the user-defined adjustment
factors, and the factors are applied as temporary prior-adjustment factors.
series {title = "Total U.S. Retail Sales --- Current Dollars"

file = "retail.dat"
start = 1980.jan }
transform {function = log
title = "Consumer Price Index"
file = "cpi.dat"
format = "1R"
precision = 3
name = "cpi"
type = temporary
}
Example 5 Specify a cube root transformation to stabilize the variation of a quarterly time series.
SERIES {TITLE="Annual Rainfall"

FILE="RAIN.DAT"
PERIOD=4
START=1901.1}
TRANSFORM {POWER=.3333}
Example 6 This example uses two sets of user-defined prior-adjustment factors: one for the Consumer
Price Index that will be removed from the final seasonally adjusted series to convert the
value of the series to current dollars (a permanent prior effect), and a set of strike effects (a
temporary prior effect). Each set of factors is read from its own file. Since the files have the
same format, single values are entered for format and precision.
7.18. TRANSFORM 203
series {title = "Retail Sales of computers --- Current Dollars"

file = "rscomp.dat" start = 1980.jan
}
transform { function = log
title = "Consumer Price Index & Strike Effect"
type = (permanent temporary)
file = ("cpi.dat" "strike.dat")
format = "1R" precision = 3
name = ("cpi" "strike")
}
Example 7 Use the automatic transformation selection procedure to determine if a log transformation
should be used to transform the series. Since a regARIMA model is not specified, the
program will use an airline model to generate the AICC values needed for the test. The AICC
difference for the test has been reset to zero, so the program will choose the transformation
based on which model estimation yields the smaller value of AICC.
series {title = "Total U.K. Retail Sales"

file = "ukretail.dat"
start = 1978.jan
}
transform {function = auto
aicdiff = 0.0
}
7.19 X11
DESCRIPTION
An optional spec for invoking seasonal adjustment by an enhanced version of the methodology of the Census
Bureau X-11 and X-11Q programs. The user can control the type of seasonal adjustment decomposition cal-
culated (mode), the seasonal and trend moving averages used (seasonalma and trendma), and the type of
extreme value adjustment performed during seasonal adjustment (sigmalim). The output options, specified
by print and save, include final tables and diagnostics for the X-11 seasonal adjustment method. In X-13-
ARIMA-SEATS, additional specs can be used to diagnose data and adjustment problems, to develop compensating
prior regression adjustments, and to extend the series by forecasts and backcasts. Such operations can result in
a modified series from which the X-11 procedures obtain better seasonal adjustment factors. For more details
on the X-11 seasonal adjustment diagnostics, see Shiskin, Young, and Musgrave (1967), Lothian and Morry
(1978), and Ladiray and Quenneville (2001). Trading day effect adjustments and other holiday adjustments can
be obtained from the x11regression spec.
USAGE
x11{ mode = pseudoadd
seasonalma = s3x9
trendma = 13
sigmalim = (1.25 2.75)
title = "3x9 moving average, mad"
appendfcst = yes
appendbcst = no
final = user
print = ( brief +b2)
save = ( d10 d11 )
savelog = ( m7 q )
}
ARGUMENTS
appendbcst Determines if backcasts will be included in certain X-11 tables selected for storage with
the save option. If appendbcst=yes, then backcasted values will be stored with tables
tables a16, b1, d10, d16, and h1. If appendbcst=no, no backcasts will be stored. The
default is to not include backcasts.
appendfcst Determines if forecasts will be included in certain X-11 tables selected for storage with
the save option. If appendfcst=yes, then forecasted values will be stored with tables
a16, b1, d10, d16, and h1. If appendfcst=no, no forecasts will be stored. The default is
to not include forecasts.
final List of the types of prior adjustment factors, obtained from the regression and outlier
specs, that are to be removed from the final seasonally adjusted series. Additive outliers
7.19. X11 205
(final=ao), level change and ramp outliers (final=ls), temporary change (final=tc),
and factors derived from user-defined regressors (final=user) can be removed. If this
option is not specified, the final seasonally adjusted series will contain these effects.
mode Determines the mode of the seasonal adjustment decomposition to be performed. There
are four choices: (a) multiplicative (mode=mult), (b) additive (mode=add), (c) pseudo-
additive (mode=pseudoadd), and (d) log-additive (mode=logadd) decomposition. The
default mode is [mult], unless the automatic transformation selection procedure is invoked
in the transform spec; in the latter case, the mode will match the transformation selected
for the series (mult for the log transformation and add for no transformation).
print and save Table 7.46 gives the output tables that are available by default; Table 7.47 gives other
tables that can be printed or saved using this argument, while Table 7.48 shows the line
printer plots that can be specified using the print argument.
to save certain tables as percentages rather than ratios. Specifying these table names in
the print argument will not change the output of the program, and the percentages are
only produce when multiplicative or log-additive seasonal adjustment is specified by the
user in the mode argument; these quantities will be expressed as differences if mode =
add.
7.50.
seasonalma Specifies which seasonal moving average (also called seasonal ”filter”) will be used to
estimate the seasonal factors. These seasonal moving averages are n×m moving av-
erages, meaning that an n-term simple average is taken of a sequence of consecutive
m-term simple averages.
The seasonal filters shown in Table 7.51 can be selected for the entire series, or for
a particular month or quarter. If the same moving average is used for all calendar
months or quarters, only a single value need be entered. If different seasonal moving
averages are desired for some calendar months or quarters, a list of these must be entered,
specifying the desired seasonal moving average for each month or quarter. An example
for a quarterly series is the following: seasonalma=(s3x3 s3x9 s3x9 s3x9).
If no seasonal moving average is specified, the program will choose the final seasonal filter
automatically; this option can also be invoked by setting seasonalma=msr. This is done
using the moving seasonality ratio procedure of X-11-ARIMA/88, see DETAILS. This is a
change from previous versions of X-11 and X-11-ARIMA where, when no seasonal moving
average was specified, a 3×3 moving average was used to calculate the initial seasonal
factors in each iteration, and a 3×5 moving average to calculate the final seasonal factors.
This seasonal filtering sequence can be specified by entering seasonalma=x11default.
sigmalim Specifies the lower and upper sigma limits used to downweigh extreme irregular values
in the internal seasonal adjustment iterations. The sigmalim argument has two in-
put values, the lower and upper sigma limits. Valid list values are any real numbers
greater than zero with the lower sigma limit less than the upper sigma limit (example:
sigmalim=(1.8 2.8)). A missing value defaults to 1.5 for the lower sigma limit and 2.5
for the upper sigma limit. For example, the statement sigmalim=(,3.0) specifies that
Table 7.46: Default Output Tables for X11 spec

adjustdiff fad + final adjustment difference (only for pseudo-additive sea-
sonal adjustment)
adjustfac d16 + combined seasonal and trading day factors
adjustmentratio e18 + final adjustment ratios (original series / seasonally ad-
justed series)
calendar d18 + combined holiday and trading day factors
calendaradjchanges e8 + percent changes (differences) in original series adjusted for
calendar effects
combholiday chl + combined holiday prior adjustment factors, A16 table
ftestd8 d8f · F-tests for stable and moving seasonality, D8
irregular d13 + final irregular component
irrwt c17 + final weights for the irregular component
movseasrat d9a · moving seasonality ratios for each period
origchanges e5 + percent changes (differences) in original series
qstat f3 · quality control statistics
replacsi d9 + final replacement values for extreme SI-ratios (differ-
ences), D iteration
residualseasf rsf · F-test for residual seasonality
sachanges e6 + percent changes (differences) in seasonally adjusted series
seasadj d11 + final seasonally adjusted series
seasonal d10 + final seasonal factors
seasonaldiff fsd + final seasonal difference (only for pseudo-additive seasonal
adjustment)
tdaytype tdy · trading day factors printed by type of month
totaladjustment tad + total adjustment factors (only printed out if the original
series contains values that are ≤ 0)
trend d12 + final trend-cycle
trendchanges e7 + percent changes (differences) in final trend component se-
ries
unmodsi d8 + final unmodified SI-ratios (differences)
unmodsiox d8b + final unmodified SI-ratios, with labels for outliers and ex-
treme values
x11diag f2 · summary of seasonal adjustment diagnostics
yrtotals e4 · ratio of yearly totals of original and seasonally adjusted
series
argument.
7.19. X11 207
Table 7.47: Other Output Tables for X11

adjoriginalc c1 + original series modified for outliers, trading day and prior factors, C
iteration
adjoriginald d1 + original series modified for outliers, trading day and prior factors, D
iteration
autosf asf · automatic seasonal factor selection
extreme c20 + extreme values, C iteration
extremeb b20 + extreme values, B iteration
ftestb1 b1f · F-test for stable seasonality, B1 table
irregularadjao iao + final irregular component adjusted for point outliers
irregularb b13 + irregular component, B iteration
irregularc c13 + irregular component, C iteration
irrwtb b17 + preliminary weights for the irregular component
mcdmovavg f1 + MCD moving average of the final seasonally adjusted series
modirregular e3 + irregular component modified for zero-weighted extreme values
modoriginal e1 + original series modified for zero-weighted extreme values
modseasadj e2 + seasonally adjusted series modified for zero-weighted extreme values
modsic4 c4 + modified SI-ratios (differences), C iteration
modsid4 d4 + modified SI-ratios (differences), D iteration
replacsib4 b4 · preliminary replacement values for extreme SI-ratios (differences), B
iteration
replacsib9 b9 · replacement values for extreme SI-ratios (differences), B iteration
replacsic9 c9 + modified SI-ratios (differences), C iteration
robustsa e11 + robust final seasonally adjusted series
seasadjb11 b11 + seasonally adjusted series, B iteration
seasadjb6 b6 + preliminary seasonally adjusted series, B iteration
seasadjc11 c11 + seasonally adjusted series, C iteration
seasadjc6 c6 + preliminary seasonally adjusted series, C iteration
seasadjconst sac + final seasonally adjusted series with constant from transform spec
included
seasadjd6 d6 + preliminary seasonally adjusted series, D iteration
seasonalb10 b10 + seasonal factors, B iteration
seasonalb5 b5 + preliminary seasonal factors, B iteration
seasonalc10 c10 + preliminary seasonal factors, C iteration
seasonalc5 c5 + preliminary seasonal factors, C iteration
seasonald5 d5 + preliminary seasonal factors, D iteration
sib3 b3 + preliminary unmodified SI-ratios (differences)
sib8 b8 + unmodified SI-ratios (differences)
tdadjorig c19 + original series adjusted for final trading day
tdadjorigb b19 + original series adjusted for preliminary trading day
trendadjls tal + final trend-cycle adjusted for level shift outliers
trendb2 b2 + preliminary trend-cycle, B iteration
trendb7 b7 + preliminary trend-cycle, B iteration
trendc2 c2 + preliminary trend-cycle, C iteration
trendc7 c7 + preliminary trend-cycle, C iteration
trendconst tac + final trend component with constant from transform spec included
trendd2 d2 + preliminary trend-cycle, D iteration
trendd7 d7 + preliminary trend-cycle, D iteration
argument.
Table 7.48: Plots Specified by the print Argument
name description of plot

irregularplot plot of the final irregular component
origwsaplot plot of the original series with the final seasonally adjusted series
ratioplotorig month-to-month (or quarter-to-quarter) ratio plots of the original series
ratioplotsa month-to-month (or quarter-to-quarter) ratio plots of the seasonally adjusted
series
seasadjplot plot of the final seasonally adjusted series
seasonalplot seasonal factor plots, grouped by month or quarter
trendplot plot of the final trend-cycle
Name gives the name of each plot for use with the print arguments.

adjustfacpct paf combined adjustment factors, expressed as percentages if ap-
propriate
calendaradjchangespct pe8 percent changes in original series adjusted for calendar factors
irregularpct pir final irregular component, expressed as percentages if appro-
priate
origchangespct pe5 percent changes in the original series
sachangespct pe6 percent changes in seasonally adjusted series
seasonalpct psf final seasonal factors, expressed as percentages if appropriate
trendchangespct pe7 percent changes in final trend cycle
7.19. X11 209
Table 7.50: Available Log File Diagnostics for X11

alldiagnostics all All seasonal adjustment diagnostics listed in this table
fstableb1 fb1 F-test for stable seasonality, performed on the original series
fstabled8 fd8 F-test for stable seasonality, performed on the final SI-ratios
icratio icr ¯ C̄ ratio
I/
idseasonal ids Identifiable seasonality test result
m1 m1 M1 Quality Control Statistic
movingseasf msf F-test for moving seasonality
movingseasratio msr Moving seasonality ratio
q q Overall index of the quality of the seasonal adjustment
q2 q2 Q statistic computed without the M2 Quality Control statistic
Table 7.51: X-13ARIMA-SEATS Seasonal Filter Options and Descriptions
name description of option

s3x1 A 3×1 moving average.
stable Stable seasonal filter. A single seasonal factor for each calendar month or quarter
is generated by calculating the simple average of all the values for each month or
quarter (taken after detrending and outlier adjustment).
x11default A 3×3 moving average is used to calculate the initial seasonal factors in each
iteration, and a 3×5 moving average to calculate the final seasonal factors.
the upper sigma limit will be set to 3.0, while the lower sigma limit will remain at the
1.5 default. A comma is necessary if either sigma limit is missing. For an explanation of
how X-13ARIMA-SEATS uses these sigma limits to derive adjustments for extreme values,
see DETAILS.
title Title of the seasonal adjustment, in quotes, for the convenience of the user. This can be
a single title or a list of up to 8 titles; an example with two titles is:
title= ("3x9, trading day adjustment"

"for sales of sporting goods")
If a list is provided, each title must be on a separate line of the spec file. This list will be
printed on the title page below the series title. There is no default seasonal adjustment
title.
trendma Specifies which Henderson moving average will be used to estimate the final trend-cycle.
Any odd number greater than one and less than or equal to 101 can be specified. Example:
trendma=23. If no selection is made, the program will select a trend moving average based
on statistical characteristics of the data. For monthly series, either a 9-, 13- or 23-term
Henderson moving average will be selected. For quarterly series, the program will choose
either a 5- or a 7-term Henderson moving average.
type When type=summary, the program develops estimates of the trend-cycle, irregular, and
related diagnostics, along with residual seasonal factors and, optionally, also residual
trading day and holiday factors from an input series which is assumed to be either already
seasonally adjusted or nonseasonal. These residual factors are not removed. The output
series in the final seasonally adjusted series (table D11) is the same as the original series
(table A1). When type=trend, the program develops estimates for the final trend-cycle
and irregular components without attempting to estimate a seasonal component. The
input series is assumed to be either already seasonally adjusted or nonseasonal. With this
option, estimated trading day and holiday effects as well as permanent prior adjustment
factors are removed to form the adjusted series (table D11) as well as for the calculation
of the trend (table D12). When a metafile with a composite spec is used to obtain
an indirect adjustment of an aggregate, these options are used for components of the
aggregate that are not seasonally adjusted. In the default setting, type=sa, the program
calculates a seasonal decomposition of the series. With all three values of type, the final
seasonally adjusted series (printed in the D11 table of the main output file) is used to
form the indirect seasonal adjustment of the composite.
7.19. X11 211
calendarsigma Specifies if the standard errors used for extreme value detection and adjustment are com-
puted separately for each calendar month (quarter), or separately for two complementary
sets of calendar months (quarters). If calendarsigma=all, the standard errors will be
computed separately for each month (quarter). If calendarsigma=signif, the standard
errors will be computed separately for each month only if Cochran’s hypothesis test de-
termines that the irregular component is heteroskedastic by calendar month (quarter).
If calendarsigma=select, the months (quarters) will be divided into two groups, and
the standard error of each group will be computed. For the select option, the argument
sigmavec must be used to define one of the two groups of months (quarters). If cal-
endarsigma is not specified, the standard errors will be computed from 5 year spans of
irregulars, in the manner described in Dagum (1988).
centerseasonal If centerseasonal = yes, the program will center the seasonal factors combined with
user-defined seasonal regression effects. The default is centerseasonal = no.
keepholiday Determines if holiday effects estimated by the program are to be kept in the final sea-
sonally adjusted series. In the default setting, keepholiday=no, holiday adjustment
factors derived from the program are removed from the final seasonally adjusted series.
If keepholiday=yes, holiday adjustment factors derived from the program are kept in
the final seasonally adjusted series. The default is used to produce a series adjusted for
both seasonal and holiday effects.
print1stpass If print1stpass=yes, output from the seasonal adjustment needed to generate the irreg-
ular components used for the X-11 Easter and irregular regression adjustment procedures
will be printed out. If print1stpass=no, this output will be suppressed, and only the
tables associated with the X-11 Easter and irregular regression procedures are printed
out. The default is print1stpass=no. When print1stpass=yes, the print argument
controls which tables are actually printed.
sfshort Controls what seasonal filters are used to obtain the seasonal factors if the series is at
most 5 years long. For the default case, sfshort=no, a stable seasonal filter will be
used to calculate the seasonal factors, regardless of what is entered for the seasonalma
argument. If sfshort=yes, X-13ARIMA-SEATS will use the central and one sided seasonal
filters associated with the choice given in the seasonalma argument wherever possible.
sigmavec Specifies one of the two groups of months (quarters) for whose irregulars a group stan-
dard error will be calculated under the calendarsigma=select option. The user enters
the month(s) (either the full name of the month or the abbreviations shown in Sec-
tion 3.3) or quarter(s) (q1 for the first quarter, q2 for the second quarter, etc.) that
comprise one group; all remaining months or quarters comprise the second group. Exam-
ple: sigmavec=(jan feb dec). Warning: This argument can only be specified when
calendarsigma=select.
trendic Specifies the irregular-to-trend variance ratio that will be used to generate the end weights
for the Henderson moving average. The procedure is taken from ?). If this variable is not
specified, the value of trendic will depend on the length of the Henderson trend filter.
These default values closely reproduce the end weights for the set of Henderson trend
filters which originally appeared in X-11and X-11-ARIMA.

true7term Specifies the end weights used for the seven term Henderson filter. If true7term =
yes, then the asymmetric ends weights for the 7 term Henderson filter are applied for
observations at the end of the series where a central Henderson filter cannot be applied.
If true7term = no, then central and asymmetric weights from a 5 term Henderson filter
are applied, as in previous versions of the X-11-ARIMA program released by Statistics
Canada. The default is true7term = no.
DETAILS
Modes of seasonal adjustment: In any X-13ARIMA-SEATS seasonal adjustment, the original time series (O)
is decomposed into three basic components:
Trend-Cycle (C): The long-term and medium-to-long term movements of the series, including consequential
turning points.
Seasonal (S): Within-year fluctuations about the trend that recur in a very similar way in the same
month or quarter from year to year.
Irregular (I): The residual component that remains after seasonal and trend are removed from the
series (and also trading day and holiday effects once these have been identified). It is
characterized by movements of very short duration. These can be quite large if there are
strikes or other unusual economic events of short duration.
Depending mainly on the nature of the seasonal movements of a given series, several different models are
used to describe the way in which the components C, S, and I combine to form the original series O. X-13-
ARIMA-SEATS provides modes of seasonal adjustment appropriate for four different decomposition models. Table
7.52 gives the four values of the mode arguments and also the corresponding models, both for the original (O)
and seasonally adjusted series (SA).
Table 7.52: Modes of Seasonal Adjustment and Their Models
Entry for Name for Model for Model for

mode argument mode O SA
mult Multiplicative O =C ×S×I SA = C × I
add Additive O =C +S+I SA = C + I
pseudoadd Pseudo-Additive O = C × [S + I − 1] SA = C × I
logadd Log-Additive Log(O) = C + S + I SA = exp(C + I)
The default seasonal adjustment mode is multiplicative. This is because, for most seasonal economic time
series, the magnitudes of the seasonal fluctuations appear to increase and decrease proportionally with increases
and decreases in the level of the series, in a way that is proportional to the level. A series with this type
of seasonality is said to have multiplicative seasonality. To estimate the multiplicative components, the
program uses a ratio-to-moving average method whose details are given in Shiskin, Young, and Musgrave
(1967), Dagum (1988), Baxter (1994), and Ladiray and Quenneville (2001), among others. The pseudo-additive
7.19. X11 213
model is considered when some months (or quarters) have extremely small values (due to vacations or climate,
for example), and the remaining months appear to have multiplicative seasonality. If the magnitude of the
seasonal does not appear to be affected by the level of the series, then the series has additive seasonality,
and the additive mode is appropriate.
The log-additive mode gives an alternative multiplicative decomposition which can be useful for certain
econometric analysis, usually related to time series model considerations. For log-additive seasonal adjustment,
the trend component is computed from an additive decomposition of the logged series (log(O)), so the additive
trend must be exponentiated in order to derive a trend with the same units as the original series. This results
in a downwardly biased estimate of trend; this bias is adjusted in X-13ARIMA-SEATS using a bias-correction
procedure described in Thomson and Ozaki (2002).
For multiplicative, pseudo-additive and log-additive seasonal adjustment, the seasonal and irregular com-
ponents are assumed to be ratios centered about 1. In the main output they are expressed as percentages so
that they center about 100. For additive seasonal adjustment, the seasonal and irregular components are in the
same units as the original time series and vary about 0.
When a regARIMA model is specified with the regression and arima specs, trading day, holiday, outlier,
and other regression effects defined in the regression spec can be derived from the regression coefficients of a
regARIMA model and used to adjust the original series prior to seasonal adjustment. In this case, these effects
must be the same type as factors generated by the seasonal adjustment procedure, so that combined adjustment
factors can be derived and adjustment diagnostics can be generated. If the regARIMA model is fit to a log-
transformed series, the regression factors are expressed in the form of ratios, which is the same form as factors
generated by the multiplicative or log-additive adjustment modes. Conversely, if the regARIMA model is fit to
the original series, the regression factors are measured on the same scale as the original series, which matches the
scale of the components generated by the additive adjustment mode. Therefore, users should be careful to ensure
that the transformation specified by the function or power arguments of the transform spec is compatible
with the seasonal adjustment mode specified by the mode argument of the x11 spec. Furthermore, be aware
that the default value for the mode argument is multiplicative seasonal adjustment, which conflicts with the
default for the function and power arguments of the transform spec, which assume no transformation is
done. Currently, you cannot use regression effects to pre-adjust the original series for a pseudo-additive seasonal
adjustment.
Multiplicative and pseudo-additive seasonal adjustment give very similar results for most series with multi-
plicative seasonality, unless the seasonal amplitude of the series is large. If the smallest seasonal factor is 0.7 or
less, there will be noticeable differences between the multiplicative and pseudo-additive seasonal adjustments.
If the smallest seasonal factor is 0.5 or less, this difference is likely to be important. If a multiplicative seasonal
adjustment produces many more extreme values (meaning values of less than 100.0 in Table C17, especially 0)
in months (or quarters) with small seasonal factors than months with large seasonal factors, then the pseudo-
additive seasonal adjustment is likely to be better. For more details on when to use pseudo-additive seasonal
adjustment, see Baxter (1994).
For simplicity, this discussion has ignored trading day and holiday effects. When these are estimated, they
add additional factors to the decomposition and, depending on how they are defined, adjustment for them can
lead to larger differences between the annual totals of the adjusted series and the annual totals of the original
time series.
Downweighting of extreme irregulars: Let µI be the assumed mean of the irregular component (1.0
for multiplicative seasonal adjustment, 0.0 for additive). Let σX11 denote an estimate of the standard deviation
of the irregular component for a month or quarter. If the absolute value of It − µI is less than the lower sigma
limit multiplied by σX11 , the irregular value It receives full weight. If the absolute value of It − µI is more than
the upper sigma limit multiplied by σX11 , the irregular value receives zero weight, meaning that It is replaced
by µt for seasonal factor calculations. Otherwise, It is partially downweighted.
Automatic seasonal filter selection: This procedure is taken from X-11-ARIMA/88, see Dagum (1988).
For the first two seasonal adjustment iterations, a 3×3 moving average is used to calculate the initial seasonal
factors and a 3×5 moving average is used to calculate the final seasonal factor. In the third and final iteration,
a 3×3 moving average is used to calculate the initial seasonal factors, but for the final iteration the program
¯ S̄, also called the global MSR). Then the program chooses whether
calculates the moving seasonality ratio ( I/
to use a 3×3, 3×5, or 3×9 moving average based on the size of the global MSR. For more information on the
moving seasonality ratio, see Lothian (1984).
Forecast extension: As mentioned in the introduction, an important use of regARIMA models is to
extend the series with forecasts (and backcasts) to improve the seasonal adjustment of the most recent (and
the earliest) observations. Therefore, X-13ARIMA-SEATS will extend the series with one year of forecasts prior
to seasonal adjustment whenever a regARIMA model is specified with no forecast spec. To specify a seasonal
adjustment without forecast extension, set maxlead = 0 in the forecast spec.
Level shifts and the final Henderson trend: When level shifts are estimate and removed from the series
prior to seasonal adjustment, they are put back into the final Henderson trend cycle (Table D12), so that this
component will have the level of the observed data. A table of the trend cycle of the level shift adjusted time
series can also be obtained by setting print = trendadjls.
Easter adjustment: The Easter adjustment options in this spec cannot be used when regARIMA model
based holiday are specified in the regression spec, or if an Easter adjustment is specified within the x11regression
spec.
Table of SI values with labels for extreme values: Table D8.B is designed to provide users with direct
information about which of the unmodified Seasonal-Irregular values (the detrended series, henceforth called SI
values) produced by the X-11 seasonal adjustment program will be modified by extreme value adjustment (as
shown by the irregular weights in Table C17) or are likely to have been affected by regARIMA outliers (either
those specified by the user or those identified though the outlier spec).
Each SI value that has been identified as an X-11 extreme value is printed with a “*” next to it. SI values at
times at which a single regARIMA outlier occurs in the model are printed with a “#” next to them. Extreme
SI values at times associated with at least one regARIMA outlier are printed with a “&” next to them; SIs
at times with more than one regARIMA outlier will have a “@” next to them. All observations between (and
including) the starting and ending points of a ramp outlier are marked as if they were outliers.
With multiplicative seasonal adjustments, SI values before and after level shift outliers that are most likely
to have been affected by the level shift are marked with a “-” character next to the value. The number of
observations flagged in this way depends on the magnitude of the level shift outlier (as determined by its
regression coefficient estimate) and on the length of the Henderson filter used for the trend that generates the
SI-ratios, in the manner described in Table 7.53.
Treatment of nonseasonal series: A nonseasonal series can be decomposed into trend-cycle and irregular
components using the type=trend option. This decomposition is obtained by a simplification of the X-11
seasonal adjustment decomposition procedure that retains only the steps related to the Henderson trends and
extreme value detection. Example 7 below shows how the type=trend option can also be applied to a seasonally
7.19. X11 215
Table 7.53: Number of Surrounding

SI-ratios in Table D 8.B Assumed
Affected by a Level Shift
Percent Change Length of
in Level (∆L ) Henderson Filter
23 13 9 7 5
∆L ≤ 1.1 0 0 0 0 0
1.1 < ∆L ≤ 1.2 1 1 0 0 0
1.2 < ∆L ≤ 1.3 1 1 1 0 0
1.3 < ∆L ≤ 1.5 2 1 1 0 0
1.5 < ∆L ≤ 1.8 2 1 1 1 0
1.8 < ∆L ≤ 1.9 2 2 1 1 0
1.9 < ∆L ≤ 2.0 3 2 1 1 0
2.0 < ∆L ≤ 2.6 3 2 1 1 1
2.6 < ∆L ≤ 2.9 3 2 2 1 1
2.9 < ∆L ≤ 3.6 4 2 2 1 1
3.6 < ∆L ≤ 5.5 4 3 2 1 1
5.5 < ∆L 5 3 2 1 1
adjusted series to obtain an alternative trend used by several national statistical offices in place of the final
Henderson trend (D12) for a seasonal time series. The alternative is a slight update of the trend proposed by
Dagum (1996).
EXAMPLES
Example 1 Multiplicative seasonal adjustment with all default options (so the program uses the moving
seasonality ratio to select the seasonal filter length). The monthly series starts in January,
1976 and is stored in free format in the file klaatu.dat in the current directory.
Series { File="klaatu.dat" Start = 1976.1 }

X11 { }
Example 2 Multiplicative monthly seasonal adjustment, 3×9 seasonal factors for all months, 23-term
Henderson moving average for the trend-cycle. Perform a test (using a version of AIC that
adjusts for the length of the series) of the significance of the trading-day regressors in a
regression of the irregular component.
Series { File="klaatu.dat" Start = 1976.1 }

X11 { SeasonalMA = s3x9 TrendMA = 23 }
X11regression { variables = td aictest=td }
Example 3 Quarterly seasonal adjustment, 3×3 seasonal factors for first two quarters, 3×5 seasonal
factors for remaining two quarters, 7-term Henderson trend moving average.
series {
file="qhstarts.dat"
start = 1967.1
period=4
}
x11 {
seasonalma = (s3x3 s3x3 s3x5 s3x5)
trendma = 7
}
Example 4 Seasonal ARIMA model with regression effects are used to obtain preadjustments of monthly
data. No forecast extension will be done in this example. Specified regression variables are
a constant, trading day effects, and two level shifts, one in May, 1972 and one in September,
1976. The ARIMA part of the model is (0,1,2)(1,1,0)12 . Additively seasonally adjust the
series after preadjusting for the outlier, level-shift and trading day effects estimated using
the regARIMA model. Use sigma limits set to 2.0 and 3.5 to search for extreme values in
the irregular component of the seasonal decomposition. Use the alltables print level when
printing out seasonal adjustment output.
SERIES { TITLE = "EXPORTS OF LEATHER GOODS" START = 1969.JUL

DATA = (815 866 926 ... 942) }
REGRESSION { VARIABLES = (CONST TD LS1972.MAY LS1976.OCT) }
ARIMA { MODEL=(0 1 2)(1 1 0) }
ESTIMATE { }
FORECAST { MAXLEAD=0 }
X11 { MODE = ADD PRINT = ALLTABLES SIGMALIM = (2.0 3.5) }
Example 5 The predefined regression effects used are trading day variables and a constant. User-defined
regression variables are included to capture the effect of special sales promotions in 1988 and
1990. These variables are read in from the file special.dat. The ARIMA part of the model
is (3,1,0)(0,1,1)12 . The seasonal period, 12, is not specified since this is the default. Perform
a multiplicative seasonal adjustment on the series after pre-adjusting for the regARIMA
trading day and user-defined regression effects and extending the series with 12 forecasts
and 12 backcasts. A two-line list of seasonal adjustment titles is specified.
series { title = "Unit Auto Sales" file = "autosal.dat"

start = 1985.1 }
regression { variables = (const td)~~~user = (sale88 sale90)
file = "special.dat" format = "(2f12.2)" }
arima { model = (3 1 0)(0 1 1)12 }
forecast { maxlead=12 maxback=12 }
x11 { title = ( "Unit Auto Sales"
"Adjusted for special sales in 1988, 1990" )
}
7.19. X11 217
Example 6 Read in the data from a file using a predefined X-11 data format. Note that the starting
date will be taken from the information provided in the data file and so does not have to
be specified. Specify a regARIMA model for the log transformed data with certain outlier
terms. Use this model to generate 5 years of forecasts. Perform a multiplicative seasonal
adjustment using a 3×9 seasonal moving average for all months. Save the E2 table ( Table
D11 with D12 trend values substituted when the C17 value is zero) for use as with the next
example.
series { title="NORTHEAST ONE FAMILY Housing Starts"

file="cne1hs.ori" name="CNE1HS" format="2R" }
transform { function=log }
regression {
variables=(ao1976.feb ao1978.feb ls1980.feb
ls1982.nov ao1984.feb)
}
arima { model=(0 1 2)(0 1 1) }
forecast { maxlead=60 }
title="Adjustment of 1 family housing starts"
save = e2
}
Example 7 A continuation of Example 6. Use the type=trend option to obtain an alternative to D12
trend for seasonal time series due to Dagum (1996) as further evaluated and updated in
Dagum and Luati (2009) and Darne and Dagum (2009). Read in the data from the seasonally
adjusted series modified for outliers (AO and TC) and extreme values from the E2 file saved
in Example 6. The starting date will be taken from the information provided in the E2 file
and so does not have to be specified. The nonseasonal (0 1 1) model with be used to provide
the six forecasts of the E2 series so that the symmetric 13 term Henderson filter can be
applied. The AO outliers are commented out in the regression spec as these outliers have
already been removed from the E2 table, and therefore do not have to be accounted for in
the model. The default setting of forecast provides 12 forecasts, so it can be used. Note the
very low values of the sigma limits used to smooth the E2 values. Note also that the log
transformation is not used.
series {
title="Trend for NORTHEAST ONE FAMILY Housing Starts"
file="cne1hs.e2"
format="x12save"
}
regression {
variables=(
# ao1976.feb ao1978.feb
ls1980.feb ls1982.nov
# ao1984.feb
)
}
arima { model=(0 1 1) }
forecast { }
x11{
type=trend
sigmalim=(0.7, 1.0)
trendma=13
title="Updated Dagum (1996) trend of 1 family housing starts"
}
Example 8 The predefined regression effect is a constant. The user-defined regression variables are for
strikes in 1980, 1985, and 1991 and are located in the file strikes.dat. The ARIMA part
of the model is (0,1,1)(0,1,1)12. Since a model is specified in the spec, generate a year of
forecasts by default. Seasonally adjust the series after pre-adjusting for the user-defined
regression effects. Adjust the series for X-11 trading day before estimating the regARIMA
model.
series{ title="Automobile Sales"

file = "carsales.dat"
start = 1975.1 }
regression{ variables = ( const )
user = (strike80 strike85 strike90)
file = "strike.dat" format = "(3f12.0)"
}
arima{ model = (0 1 1)(0 1 1)12 }
x11{ title = ("Car Sales in US - Adjust for strikes in 80, 85, 90")
save=seasonal appendfcst=yes
}
x11regression { variables = td }
Example 9 Use the automatic transformation selection procedure to determine if a log transformation
should be used to transform the series. Since a regARIMA model is not specified, the pro-
gram will use an airline model to generate the AICC values needed for the test. The AICC
difference for the test has been reset to zero, so the program will choose the transformation
7.19. X11 219
based on which model estimation yields the smaller value of AICC. The choice of trans-
formation will determine if the seasonal adjustment will be a multiplicative or an additive
seasonal adjustment.
series {title = "Total U.K. Retail Sales"

file = "ukretail.dat"
start = 1978.jan
}
transform {function = auto
aicdiff = 0.0
}
x11 {save=d11
}
7.20 X11REGRESSION
DESCRIPTION
An optional spec for use in conjunction with the x11 spec for series without missing observations. This spec
estimates calendar effects by regression modeling of the irregular component with predefined or user-defined
regressors. The user can select predefined regression variables with the variables argument. The predefined
variables are for calendar (trading-day and holiday) variation and additive outliers. A change-of-regime option
is available with trading-day regressors. User-defined calendar effect regression variables can be included in
the model via the user argument. Data for any user-defined variables must be supplied, either in the data
argument, or in a file named in the file argument (not both). The regression model specified can contain both
predefined and user-defined regression variables.
USAGE
x11regression { variables = (td or td1coef or
tdstock[31] or tdstock1coef[31]
easter[8] labor[8]
thank[1]
ao1967.apr )
user = (temperature precip)
start = 1955.jan
data = (25 0.1 ...)
or
file = "weather.dat"
format = "(2f5.1)"
tdprior = ( 1.4 1.4 1.4 1.4 1.4
0.0 0.0 )
aictest = ( easter user
td or td1coef or tdstock or tdstock1coef )
aicdiff = -2.0
span = (1980.jan, 1995.dec)
sigma = 2.75
or
critical = 3.5
outliermethod = addone
outlierspan = (1973.may, 1992.sep)
usertype = holiday
prior = yes
print = ( brief +b15 )
save = ( c16 c18 )
savelog = aictest
}
7.20. X11REGRESSION 221
ARGUMENTS
aicdiff Defines the difference in AICC needed to accept a regressor specified in the aictest
argument. The default value is aicdiff=0.0. For more information on how this option
is used in conjunction with the aictest argument, see DETAILS.
aictest Specifies that an AIC-based comparison will be used to determine if a specified regression
variable should be included in the user’s irregular component regression model. The only
entries allowed for this variable are td, tdstock, td1coef, tdstock1coef, easter, and
user. If a trading day model selection is specified, for example, then AIC values (with
a correction for the length of the series, henceforth referred to as AICC) are derived
for models with and without the specified trading day variable. By default, the model
with smaller AICC is used to generate forecasts, identify outliers, etc. If more than
one type of regressor is specified, the AIC-tests are performed sequentially in this order:
(a) trading day regressors, (b) easter regressors, (c) user-defined regressors. If there are
several variables of the same type (for example, several td regressors), then the aictest
procedure is applied to them as a group. That is, either all variables of this type will
be included in the final model or none. See DETAILS for more information on the
testing procedure. If this option is not specified, no automatic AIC-based selection will
be performed.
critical Sets the critical value (threshold) against which the absolute values of the outlier t-sta-
tistics are compared to detect additive outliers (meaning extreme irregular values). This
argument applies unless the sigma argument is used, or the only regressor(s) estimated
is flow trading day. The assigned value must be a real number greater than 0. Example:
critical=4.0. The default critical value is determined by the number of observations
in the interval searched for outliers (see the outlierspan argument below). Table 7.22
gives default critical values for a number of outlier span lengths. Larger (smaller) critical
values predispose x11regression to treat fewer (more) irregulars as outliers. A large
value of critical should be used if no protection is wanted against extreme irregular
values.
data Assigns values to the user-defined regression variables. The time frame of the values
must cover the time frame of the series (or of the span specified by the span argument of
the series spec, if present). It must also cover the time frame of forecasts and backcasts
requested in the forecast spec. The data values are read in free format. The numerical
values given in this argument should be in the order in which the user-defined variables
are named in the user argument. This assignment should proceed through all the values
of the user-defined variables for the first time point, then through all the values for the
second time point, etc. If the data argument is used, the file argument cannot be used.
file Name of the file containing data values for all user-defined regression variables. The
filename must be enclosed in quotes. If the file is not in the current directory, the path
must also be given. As with the data argument, the time frame of the data values must
cover both the series and any forecasts and backcasts. If the file argument is used, the
data argument cannot be used.
format Indicates the format used when reading the values for the regression variables in the file
named in the file argument. Three types of input are accepted:
b. a valid Fortran format, which must be enclosed in quotes and must include the
initial and terminal parentheses (example: format="(6f12.0)");
c. “datevalue” format, in which the year, month or quarter, and the associated values
for each of the user-defined regression variables for a given observation are given, in
this order, in free format on individual lines in the data file. Thus, a line of the data
file with three regressors having the values 0, 0, and 1 respectively for July of 1991
would have the form 1991 7 0 0 1. All the user-defined regressors must be on the
format="datevalue");
d. the “x12save” format X-13ARIMA-SEATS uses to save a table. This allows the user to
read in a file saved from a previous X-13ARIMA-SEATS run (example: format="x12save").
e. a variant of “free” format where the numbers must be separated by one or more
f. a variant of “datevalue” format, where the year, month or quarter, and value of
decimal points are expressed as commas. Thus, a line of the data file with three
regressors having the values 0.5, 0, and 1.25 respectively for July of 1991 would
have the form 1991 7 0,5 0 1,25. All the user-defined regressors must be on the
format="datevaluecomma").
If no format argument is given the data will be read in free format. In free format, all
numbers on a line will be read before continuing to the next line, and the numbers must
be separated by one or more spaces (not by commas or tabs). Format cannot be used
with the data argument, only with file.
outliermethod Determines how the program successively adds detected outliers to the model. The
choices are method = addone or method = addall. See the DETAILS section of the
outlier spec for a description of these two methods. The default is method = addone.
This argument cannot be used if the sigma argument is used.
outlierspan Specifies start and end dates of the span of the irregular component to be searched for
outliers. The start and end dates of the span must both lie within the series, and the
start date must precede the end date. A missing value, e.g., outlierspan = (1976.jan,
), defaults to the start date or end date of the series, as appropriate. (If there is a span
argument in the series spec, then, in the above remarks, replace the start and end dates
of the series by the start and end dates of the span given in the series spec.) This
argument cannot be used with the sigma argument.
see Appendix B.
Table 7.54: Default Output Tables for X11regression

priortd a4 + prior trading day weights and factors
extremeval c14 + irregulars excluded from the irregular regres-
sion, C iteration
x11reg c15 · final irregular regression coefficients and
diagnostics
tradingday c16 + final trading day factors and weights
combtradingday c18 + final trading day factors from combined daily
weights
holiday xhl + final holiday factors
calendar xca + final calendar factors (trading day and holiday)
combcalendar xcc + final calendar factors from combined daily
weights
outlierhdr xoh · options specified for outlier detection including
critical value and outlier span
xaictest xat · output from AIC-based tests for trading day
and holiday
prior Specifies whether calendar factors from the irregular component regression are computed
in a preliminary run and applied as prior factors (prior=yes), or as a part of the seasonal
adjustment process (prior=no). The default is prior=no. The prior argument has
no effect when a regARIMA model is specified; in this case, the irregular component
regression is always computed before seasonal adjustment.
savelog Setting savelog=aictest or savelog=ats causes the results of the AIC-based selection
procedure specified by the aictest argument to be output to the log file (see Section 2.6
for more information on the log file).
Table 7.55: Other Output Tables for X11regression

extremevalb b14 + irregulars excluded from the irregular regres-
sion, B iteration
x11regb b15 · preliminary irregular regression coefficients
and diagnostics
tradingdayb b16 + preliminary trading day factors and weights
combtradingdayb b18 + preliminary trading day factors from com-
bined daily weights
holidayb bxh + preliminary holiday factors
calendarb bxc + preliminary calendar factors
combcalendarb bcc + preliminary calendar factors from combined
daily weights
outlieriter xoi + detailed results for each iteration of outlier
detection including outliers detected, outliers
deleted, model parameter estimates, and ro-
bust and non-robust estimates of the residual
standard deviation
outliertests xot · t-statistics for every time point of each outlier
detection iteration
xregressionmatrix xrm + values of irregular regression variables with
associated dates
xregressioncmatrix xrc + correlation matrix of irregular regression pa-
rameter estimates if used with the print ar-
gument; covariance matrix of same if used
with the save argument
sigma The sigma limit for excluding extreme values of the irregular components before trading
day (only) regression is performed. Irregular values larger than this number of standard
deviations from the mean (1.0 for multiplicative adjustments, 0.0 for additive adjust-
ments) are excluded as extreme. Each irregular has a standard error determined by its
month (or quarter) type. The month types are determined by the month length, by
the day of the week on which the month starts. This argument cannot be used when
regressors other than flow trading day are present in the model, or when the critical
argument is used. The assigned value must be a real number greater than 0; the default
is 2.5 (which is invoked only when the flow trading day variable(s) are the only regressor
estimated). Example: sigma=3.0.
span Specifies the span (data interval) of irregular component values to be used to estimate
the regression model’s coefficients. This argument can be utilized when, for example,
the user does not want data early in the series to affect regression estimates used for
preadjustment before seasonal adjustment. As with the modelspan spec detailed in the
series spec, the span argument has two values, the start and end date of the desired
span. A missing value defaults to the corresponding start or end date of the span of
the input series. For example, for monthly data, the statement span=(1968.1, ) causes
whatever irregular regression model is specified to be estimated from the time series data
starting in January, 1968 and ending at the end date of the analysis span. A comma is
necessary if either the start or the end date is missing. The start and end dates of the
model span must both lie within the time span of data specified for analysis in the series
spec, and the start date must precede the end date.
of span to always be the most recent occurrence of a specific calendar month (quarter
for quarterly data) in the span of data analyzed, where per denotes the calendar month
(quarter). Thus, if the span of data considered ends in a month other than December,
span=(,0.dec) will cause the regression coefficients to stay fixed at the values obtained
from data ending in December of the next-to-final calendar year of the span.
start The start date for the values of the user-defined regression variables. The default is the
start date of the series. Valid values are any date up to the start date of the series (or
up to the start date of the span specified by the span argument of the series spec, if
present).
tdprior User-input list of seven daily weights, starting with Monday’s weight, which specify a
desired X-11 trading day adjustment prior to seasonal adjustment. These weights are
adjusted to sum to 7.0 by the program. This option can be used only with multiplicative
and log-additive seasonal adjustments. The values must be real numbers greater than or
equal to zero. Example: tdprior=(0.7 0.7 0.7 1.05 1.4 1.4 1.05).
user Specifies the list of names of user-defined regression variables. A name is required for each
user-defined variable whose coefficients are to be estimated. The names given are used to
label estimated coefficients in the program’s output. Values for the user-defined variables
must be supplied, using either the data or the file argument (not both). The maximum
number of user-defined regression variables is 52. (This limit can be changed—see Section
2.8.)
usertype Assigns a type to the user-defined regression variables. The user-defined regression effects
can be defined as a trading day (td), stock trading day (tdstock), holiday (holiday,
easter, and the US holidays thanks and labor), additive outlier (ao), or other user-
defined (user) regression effects. A single effect type can be specified for all the user-
defined regression variables defined in the x11regression spec (usertype=td), or each
user-defined regression variable can be given its own type (usertype=(td td td td td
td holiday user)). See DETAILS for more information on assigning types to user-
defined regressors.
variables List of predefined regression variables to be included in the model. The values of these
variables are calculated by the program, as functions of the calendar in most cases. See
DETAILS for a discussion and a table of the available predefined variables.

almost Differential used to determine the critical value used for a set of ”almost” outliers -
outliers with t-statistics near the outlier critical value that are not incorporated into the
regARIMA model. After outlier identification, any outlier with a t-statistic larger than
Critical − almost is considered an ”almost outlier,” and is included in a separate table.
The default is almost = 0.5; values for this argument must always be greater than zero.
b Specifies initial values or fixed values for irregular component regression parameters in
the order in which they appear in the variables and user arguments. If present, the b
argument must assign initial values to all regression coefficients in the irregular compo-
nent regression model, and must appear in the spec file after the variables and user
arguments. Initial values are assigned to parameters either by specifying the value in
the argument list or by explicitly indicating that it is missing as in the example below.
Missing values take on the default value of 0.1. For example, for a model with two
regressors, b=(0.7, ) is equivalent to b=(0.7,0.1), but b=(0.7) is not allowed. For
a model with three regressors, b=(0.8,,-0.4) is equivalent to b=(0.8,0.1,-0.4). To
hold a parameter fixed at the specified value, immediately follow the value in the b list
with an ‘f’, e.g., b=(0.7f, 0.1).
centeruser Specifies the removal of the (sample) mean or the seasonal means from the user-defined
regression variables. If centeruser=mean, the mean of each user-defined regressor is
subtracted from the regressor. If centeruser=seasonal, means for each calendar month
(or quarter) are subtracted from each of the user-defined regressors. If this option is
not specified, the user-defined regressors are assumed to already be in an appropriately
centered form and are not modified.
eastermeans Specifies whether long term (400 year) monthly means are used to deseasonalize the
Easter regressor associated with the variable easter[w], as described in footnote 5 of
Table 4.1 (eastermeans=yes), or, instead, monthly means calculated from the span of
data used for the calculation of the coefficients of the Easter regressors (eastermeans=no).
The default is eastermeans=yes. This argument is ignored if no built-in Easter regressor
is included in the regression model, or if the only Easter regressor is sceaster[w] (see
Details).
forcecal Specifies whether the calendar adjustment factors are to be constrained to have the same
value as the product (or sum, if additive seasonal adjustment is used) of the holiday and
trading day factors (forcecal=yes), or not (forcecal=no). The default is forcecal=no.
This argument is functional only when both holiday and trading day regressors are spec-
ified in the variables argument of this spec.
noapply List of the types of regression effects defined in the x11regression spec whose model-
estimated values are not to be adjusted out of the original series or final seasonally
adjusted series. Available effects include modelled trading day effects (td) and Easter,
Labor Day, and Thanksgiving-Christmas holiday effects (holiday).
reweight Specifies whether the daily trading day weights are to be re-weighted when at least one
of the daily weights in the C16 output table is less than zero (reweight=yes), or not
(reweight=no). The default is reweight=no. This argument is functional only when
trading day regressors are specified in the variables argument of this spec. Note: the
default for previous versions of X-11 and X-11-ARIMA corresponds to reweight=yes.
umdata An input array of mean-adjustment values, to be subtracted from the irregular series
It (or Log It ) before the coefficients of a model with a user-defined regressor are es-
timated. This argument, or umfile, is used when the mean function for predefined
regressors described in DETAILS is incorrect for the model with user-defined regressors.
The mean-adjustment function depends on the mode of adjustment. See DETAILS for
more information.
The time frame of these values must cover the time frame of the series (or of the span
specified by the span argument of the series spec, if present). It must also cover the
time frame of forecasts and backcasts requested in the forecast spec. The data values
are read in free format. If the umdata argument is used, the umfile argument cannot
be used.
umfile Name of the file containing a series of mean-adjustment values to be subtracted from
the irregular series It (or Log It ) before the coefficients of a model with a user-defined
regressor are estimated. This replaces the mean function that is subtracted from It when
only predefined regressors are used, as described in DETAILS. The filename must be
enclosed in quotes. If the file is not in the current directory, the path must also be given.
As with the umdata argument, the time frame of the data values must cover both the
series and any forecasts or backcasts. If the file argument is used, the umdata argument
cannot be used.
umformat Denotes the format used when reading the data for the regression variables from the file
named in the umfile argument. Five types of input are accepted:
a. a valid Fortran format, which must be enclosed in quotes and must include the
initial and terminal parentheses (example: umformat="(6f12.0)");
b. “datevalue” format, in which the year, month or quarter, and the associated value
for the mean-adjustment for a given observation are given, in this order, in free
format on individual lines in the data file. Thus, a line of the data file with a mean
adjustment of 1.01 for July of 1991 would have the form 1991 7 1.01 (example:
umformat="datevalue");
c. the “x12save” format X-13ARIMA-SEATS uses to save a table. This allows the user to
read in a file saved from a previous X-13ARIMA-SEATS run (example: umformat="x12save");
d. a two character code which corresponds to a set of data formats used in previous
versions of X-11 and X-11-ARIMA (example: umformat="1r");
e. the format that the TRAMO and SEATS programs use to read in a series and its
umformat="tramo");
f. a variant of “free” format where the numbers must be separated by one or more
convention in some European countries). (example:
format="freecomma");
g. a variant of “datevalue” format, where the year, month or quarter, and value of
In the predefined X-11 data formats mentioned in (d), the data is stored in 6 or 12
character fields, with a year and series label associated with each year of data. For
a complete list of these formats, see the DETAILS section of the series spec. If no
umformat argument is given the data will be read in free format. Umformat cannot
be used with the umdata argument, only with umfile.
umname The name of the series of values stored in the file named in umfile. The name must be
enclosed in quotes and may contain up to 64 characters. Up to the first 16 characters will
be printed as a label for the user-defined mean of the mean-adjustment values. When
specified with the predefined formats of the umformat argument, the first six (or eight,
if umformat="cs") characters of this name are also used with the predefined formats to
check that the program is reading the correct series, or to find a particular series in a file
where many series are stored.
umprecision The number of decimal digits to be read from the user-defined mean. This option can
only be used with the predefined formats of the umformat argument. This value must
be an integer between 0 and 5, inclusive (for example, umprecision=5). The default is
zero.
umstart The start date for the mean-adjustment values specified in umdata or umfile. The
default is the start date of the series. Valid values are any date up to the start date of
the series (or up to the start date of the span specified by the span argument of the
series spec, if present).
umtrimzero If umtrimzero=no, zeroes at the beginning or end of the user mean time series entered
via the umfile argument are treated as series values. If umtrimzero=span, causes leading
and trailing zeros to be ignored outside the span of data being analyzed (the span
argument must be specified with both a starting date and an ending date). The default
(umtrimzero=yes) causes leading and trailing zeros to be ignored. Note that when the
format argument is set to either datevalue, x12save, or tramo, all values input are
treated as series values, regardless of the value of umtrimzero.
DETAILS
This spec is used to estimate a calendar effect, or other effect, from the irregular component It of a preliminary
seasonal adjustment that did not adjust for the effect. The estimation is done by ordinary least squares (OLS)
applied to a regression model for the effect.
In the simplest cases detailed below, the model has the form
It − 1.0 = β 0 Xt + et ,
where Xt is a regression vector with variables that describe the basic effect of interest. In other cases, a
more complicated linear transformation of It appears on the left of the model. In all cases, t-statistics, chi-
square statistics, and AIC’s are calculated from the OLS estimates as though the regression errors et were
independent and had constant variance. Unfortunately, the filtering operations used to produce It guarantee that
both assumptions about et are somewhat incorrect, enough that decisions made for the statistical significance
of estimated effects from the statistics just mentioned are often less reliable than decisions made for effects
estimated from a regARIMA model using the regression spec. That is, x11regression’s statistics are more
likely than regression’s to suggest that a significant effect is present when it is not.. For effects that are truly
significant, the estimates from the regression and x11regression specs are usually quite close. When they
differ appreciably, those from regression are better more often than those from x11regression. (The forecast
diagnostics of the history spec can be used to compare estimated effects for series of sufficient length, see
Findley, Monsell, Bell, Otto, and Chen 1998 and Findley and Soukup 2000.) Thus use of x11regression should
normally be reserved for series for which the user is unable to find a regARIMA model with good fit over the
data span of interest.
Appendix C gives a detailed discussion of the irregular component regression models and their factors. Brief
descriptions of the predefined regression variables that can be specified in the x11regression spec is given in Table
7.56 below.
Table 7.56: Predefined regression variables for x11regression
td Estimates monthly (or quarterly) flow trading-day effects by adding the tdnolpyear
variables (see Table 7.27) to the model. The derivations of February from the average
length of 28.25 are handled either by rescaling (for multiplicative adjustments) or by
including the lpyear regression variable (for additive and log-additive adjustments). Td
cannot be used with tdstock[], tdstock1coef[] or td1coef.
td1coef Estimate monthly (or quarterly) flow trading-day effects by including the td1nolpyear
variable (see below) in the model, and by handling leap-year effects either by rescaling
(for transformed series) or by including the lpyear regression variable (for untrans-
formed series). Td1coef can only be used for monthly or quarterly series, and cannot
be used with td, tdstock1coef[] or tdstock[].
Table 7.56: Predefined regression variables for x11regression (continued)
tdstock[w] Adds 6 stock trading-day variables to model the effect of the day of the week on a stock
series estimated for the wth day of each month. The value w must be supplied and can
range from 1 to 31. For any month of length less than the specified w, the tdstock
variables are measured as of the end of the month. Use tdstock[31] for end-of-month
stock series. Tdstock can be used only with monthly series and cannot be used with
td, tdstock1coef[] or td1coef.
tdstock1coef[w] Adds a constrained stock trading-day variable to model the effect of the day of the
week on a stock series estimated for the wth day of each month. The value w must
be supplied and can range from 1 to 31. For any month of length less than the spec-
ified w, the tdstock1coef variables are measured as of the end of the month. Use
tdstock1coef[31] for end-of-month stock series. Tdstock1coef can be used only with
monthly series and cannot be used with td, tdstock[] or td1coef.
easter[w] Easter holiday regression variable (monthly or quarterly flow data only) which assumes
the level of daily activity changes on the w−th day before Easter and remains at the
new level until the day before Easter. The value w must be supplied and can range
from 1 to 25. To estimate complex effects, several of these variables, differing in their
choices of w, can be specified.
labor[w] Labor Day holiday regression variable (monthly flow data only) that assumes the level
of daily activity changes on the w−th day before Labor Day and remains at the new
level until the day before Labor Day. The value w must be supplied and can range from
1 to 25.
thank[w] Thanksgiving holiday regression variable (monthly flow data only) that assumes the
level of daily activity changes on the w−th day before or after Thanksgiving and remains
at the new level until December 24. The value w must be supplied and can range from
−8 to 17. Values of w < 0 indicate a number of days after Thanksgiving; values of
w > 0 indicate a number of days before Thanksgiving.
sceaster[w] Statistics Canada Easter holiday regression variable (monthly or quarterly flow data
only) assumes that the level of daily activity changes on the (w −1)−th day and remains
at the new level through Easter day. The value w must be supplied and can range from
1 to 24. To estimate complex effects, several of these variables, differing in their choices
of w, can be specified.
aodate Adds an additive (point) outlier variable, AO, for the given date or observation number.
For series with associated dates, AOs are specified as aodate. For monthly series the form
is aoyear.month (e.g., ao1985.jul or ao1985.7), for quarterly series it is aoyear.quarter
(e.g., ao1985.1 for an AO in the first quarter of 1985). More than one AO may be
specified. All specified outlier dates must occur within the series. (AOs with dates
within the series but outside the span specified by the span argument of the series
spec are ignored.)
The regression model specified by x11regression is estimated from the series of irregulars of the B and C
iterations of the calculations of the x11 spec. If the spec file also includes the arima, automdl, or regression
spec, then the effects estimated via x11regression are obtained first, and they are removed from the data
used for the estimations, or the forecast and backcast extensions, specified by these other specs. The series
resulting from the calculations of these other specs is then decomposed by a second execution of the x11 spec
calculations in order to obtain the seasonal, trend, calendar-effect, and irregular components output by the
program. Similarly, if the x11 spec requests the Bateman-Mayes Easter-effect adjustment, this adjustment is
calculated from a series that has been preadjusted for the effects estimated by x11regression.
If forecasting is performed, X-13ARIMA-SEATS creates data values for the selected predefined regression
variables for the entire forecast period. If there are any user-defined regression variables, then data values must
also be supplied for them for the entire forecast period. In addition to the limit of 52 user-defined regression
variables, there is an overall limit of 80 regression variables in the model. (These limits can be changed—see
Section 2.8.) The latter limit is on the sum of the number of predefined and user-defined regression variables,
plus the number of regression variables generated from automatic outlier detection. The maximum length of
the series of user-defined regression variables, not including the forecast period, is 780. (This limit can also be
changed—see Section 2.8.)
Trading day and/or holiday adjustments may be obtained either from regARIMA or from irregular regression
models, but not from both. If these effects are estimated in both the regression and x11regression spec,
then the noapply option must be used to ensure that only one set of factors is used in the adjustment.
The effect of the argument aictest can be to delete a regressor set named in the variables argument from
this list, or to add a regressor set to the model specified by the variables argument. The effect of a positive
value of aicdiff is to make it more difficult for the aictest procedure to include in the model the variable
being tested. Let ∆AIC denote the value associated with the aicdiff argument, which by default is zero. Let
AICC with (and AICC without ) denote the AICC value of the model with (or without) a set of regressors specified
in the aictest argument. If this set is not named in the variables list, it will be added to the regression model
if
AICC with + ∆AIC < AICC without .
If this set is named in the variables list, it will be retained in the irregular component regression model only
if this inequality holds.
In the second case, if aictest = (tdstock), then the end-of-month stock variables, specified by tdstock[31],
are the variables added, because 31 is the default value for w in tdstock[w].
There are more possibilities if aictest = (easter) and no Easter effect regressors appear in the variables
argument. Then three additional models are considered, namely the models obtained by augmenting the specified
irregular component regression model with the regressor easter[w] for w = 1, 8, 15 respectively. The Easter
regressor whose model has the smallest AICC is retained if its AICC is smaller than the model with no Easter
regressors by at least the amount ∆AIC = 0; otherwise, the model without Easter regressors is selected.
When trading day regressors appear in both the aictest and variables arguments, the type of regressors
specified must be identical. The sample day for stock trading day variables and the date specified for change-
of-regime regressors should not be included in the aictest argument; they will be assumed from the entry in
the variables argument. For example, if variables=(tdstock[15] ao1995.jan), then the entry for aictest
should be tdstock.
If a trading day (td or tdstock) or holiday (holiday, or the specific US holidays easter, thanks, and
labor) regressor type is assigned in to a user-defined variable with the usertype argument, the factor derived
from the user-defined regression variables of that type will be combined with the regression factor from variables
of the same type specified in the x11regression spec. The resulting factor will be adjusted out of the series for
the seasonal adjustment factor calculations determined by the x11 spec unless the type name appears in the
noapply argument.
If x11regression is used in a spec file without an x11 spec, then the irregular component used for the
regression is that obtained from the default specification x11{}.
The two choices for the argument eastermeans yield noticeably different holiday factors. But the choice
usually has negligible effects on the combined seasonal and holiday factors, because the seasonal factors change
to compensate for the differences between the choices.
The monthly means used to obtain deseasonalized Easter regressors under eastermeans=yes were generated
from frequencies of the date of Easter for a 500 year period (1600-2099). These frequencies can be computed by
dates given in Bednarek (2007) which were checked using information from Montes (2001, 1997b, 1997a); the
algorithm used to compute the date of Easter for the Gregorian calendar is given in Duffet-Smith (1981).
For a nonseasonal time series, an adjustment for trading day and holiday effects estimated by means of this
spec can be obtained by setting type=trend in the x11 spec.
When the b=() argument is used to fix coefficients, AIC and the other model selection statistics may become
invalid, see the Details section of estimate.
EXAMPLES
Example 1 Multiplicative seasonal adjustment with all default options (so the program uses the moving
seasonality ratio to select the seasonal filter length). The monthly series starts in January,
1976 and is stored in free format in the file klaatu.dat in the current directory. A trading
day adjustment is done using a regression on the irregular component.
Series { File = "westus.dat"
Start = 1976.1
}
X11 { }
X11Regression { Variables = td
}
Example 2 Same as Example 1, only an AIC-based test will be performed to see if trading day and
Easter regressors should be included in the regression on the irregular component.
Series { File = "westus.dat"
Start = 1976.1
}
X11 { }
X11Regression { Variables = td
Aictest = (td easter)
}
Example 3 User-defined holiday regressors for the period both before and after Easter are included in
the irregular regression along with trading day regressors. AO outlier identification will be
performed during the irregular regression procedure.
series {
file = "ukclothes.dat"
start = 1985.Jan
}
x11 { }
x11regression{
variables = td critical = 4.0
user = (easter1 easter2) file = "ukeaster.dat"
usertype = holiday start = 1980.Jan
}
Example 4 Prior trading day weights are provided with this spec file. The trading day weights calculated
during the irregular regression will be combined with these weights for a combined trading
day component.
series {
file = "nzstarts.dat" start = 1980.Jan
}
x11 { }
x11regression{
variables = td
tdprior = (1.4 1.4 1.4 1.4 1.4 0.0 0.0)
}
Example 5 Perform a default seasonal adjustment. The trading day regressors in the x11regression
spec will be fixed to their initial values; the Easter holiday regressor will be estimated.
series{
format = ’2R’
title = ’MIDWEST ONE FAMILY Housing Starts’
name = ’CMW1HS’
file = ’cmw1hs.ori’
span = (1964.01,1989.03)
}
x11{ }
x11regression{
variables = (td easter[8])
b = ( 0.4453f 0.8550f -0.3012f 0.2717f
-0.1705f 0.0983f -0.0082)
}
Example 6 Use an irregular component regression to estimate the trading day effect (with change-of-
regime in January of 1990) and holiday effects.
series{
title = ’Motor Home Sales’
start = 1967.1
span = (1972.1, )
name = ’SB0562’
file = ’C:\final.x12\T0B05601.TXT’
format = ’2L’
}
X11REGRESSION { variables = ( td/1990.1/
easter[8] labor[10] thank[10] ) }
x11{
seasonalma = x11default
sigmalim = (1.8 2.8)
appendfcst = YES
save = (D11 D16)
}
Example 7 The predefined regression effects are trading day variables and a constant. The user-defined
regression variables are for strikes in 1980, 1985, and 1991 and are located in the file
strikes.dat. The ARIMA part of the model is (0,1,1)(0,1,1)12. Since a model is spec-
ified in the spec, generate a year of forecasts by default. The seasonal period, 12, is not
indicated since this is the default. Seasonally adjust the series after pre-adjusting for the
user-defined regression effects. Before estimating the regARIMA model, do a prior pass to
estimate a preliminary irregular and estimate trading day and Easter effects, and remove
the calendar effects from the series. A two-line seasonal adjustment title is specified.
series{ title = "Automobile Sales"

file = "carsales.dat"
start = 1975.Jan }
regression{ variables = (const)
user = (strike80 strike85 strike90)
file = "strike.dat"
format = "(3f12.0)" }
arima{ model = (0 1 1)(0 1 1)12 }
x11{ title = ("Car Sales in US"
"Adjust for strikes in 80, 85, 90")
save = seasonal appendfcst = yes
}
x11regression{ variables = ( td easter[8] ) }
A Codes Associated With the
X-13ARIMA-SEATS Graphics Metafile
As noted in section 2.7, the -g flag specifies a complete path name for a directory into which output will be
stored that is intended as input for a separate graphics program. The program also stores a graphics metafile
into this directory, which contains a list of the files stored by the program, along with codes that denote what
table has been stored in the corresponding file.
Table A.1 below provides a list of all the tables that can be stored by X-13ARIMA-SEATS in graphics mode,
along with the codes used in the graphics metafile to denote these files (in alphabetical order).
For example, if a record in the graphics metafile reads
sa g:\users\jones001\g2\StartsUS.d11
then the final seasonally adjusted series is stored in the file g:\users\jones001\g2\StartsUS.d11.
Table A.1: Codes Associated With the X-13ARIMA-SEATS Graphics Metafile
Code Description
acf residual autocorrelations
acf2 squared residual autocorrelations
adjcori composite series (prior adjusted)
ador original series (prior adjusted)
ahst concurrent and revised seasonal adjustments and revisions
aichst revision history of the likelihood statistics
ao regARIMA AO outlier component
arat final adjustment ratios
bct point backcasts and prediction intervals on the original scale
btr point backcasts and standard errors for the transformed data
cad regARIMA calendar adjusted original data
caf combined adjustment factors
cal combined calendar adjustment factors
ccal final combined calendar factors from irregular component regression
cfchst concurrent out-of-sample forecasts
chol combined holiday component
chss sliding spans of the changes in the seasonally adjusted series
cmpcad regARIMA calendar adjusted composite data
cmpoad regARIMA outlier adjusted composite data
cmpori composite time series data (for the span analyzed)
cmppadj prior adjusted composite data
235
236 APPENDIX A. GRAPHICS CODES
Table A.1: Codes Associated With the X-13ARIMA-SEATS Graphics Metafile (continued)
Code Description
cmpspor spectrum of the original series
csahst history of the change of the adjustments
ctd final combined trading day factors from irregular component regression
ctrhst history of the change of the trend-cycle values
fct point forecasts and prediction intervals on the original scale
fcthst revision history of the out-of-sample forecasts
fintst final outlier test statistics
fltsac concurrent seasonal adjustment filter
fltsaf symmetric seasonal adjustment filter
flttrnc concurrent trend filter
flttrnf symmetric trend filter
frfc factors applied to get adjusted series with forced yearly totals
ftr point forecasts and standard errors for the transformed data
idacf residual autocorrelations for different orders of differencing
idpacf residual partial autocorrelations for different orders of differencing
indahst concurrent and revised indirect seasonal adjustments and revisions
indao indirect additive outlier adjustment factors
indarat indirect final adjustment ratios
indcaf indirect combined adjustment factors
indcal indirect calendar component
indchss sliding spans of the changes in the indirect seasonally adjusted series
indfrfc factors applied to get indirect adjusted series with forced yearly totals
indirr indirect irregular component
indls indirect level change adjustment factors
indmirr irregular component modified for extremes from indirect adjustment
indmori original data modified for extremes from indirect adjustment
indmsa seasonally adjusted data modified for extremes from indirect adjustment
indrsi final replacement values for SI component of indirect adjustment
indsa indirect seasonally adjusted data
indsar rounded indirect final seasonally adjusted series
indsass sliding spans of the indirect seasonally adjusted series
indsat final indirect seasonally adjusted series with forced yearly totals
indsf indirect seasonal component
indsfss sliding spans of the indirect seasonal factors
indsi indirect unmodified SI component
indspir spectrum of indirect modified irregular component
indspsa spectrum of differenced indirect seasonally adjusted series
indtadj indirect total adjustment factors
indtrn indirect trend cycle
indyyss sliding spans of the year-to-year changes in the indirect seasonally adjusted series
irr final irregular component
irrwt final weights for irregular component
ls regARIMA level change outlier component
237
Code Description
mdlest regression and ARMA parameter estimates
mirr modified irregular series
mori original data modified for extremes
msa modified seasonally adjusted series
mvadj original series adjusted for missing value regressors
oad regARIMA outlier adjusted original data
ori time series data (for the span analyzed)
oricnt time series data plus constant (for the span analyzed)
orifctd series forecast decomposition (SEATS)
otl regARIMA combined outlier component
pacf residual partial autocorrelations
padj prior-adjusted data
padjt prior-adjusted data (including prior trading day adjustments)
ppradj permanent prior-adjusted data
ppradjt permanent prior-adjusted data (including prior trading day adjustments)
pprior permanent prior-adjustment factors
prior prior-adjustment factors
ptd prior trading day factors
regrsd regression residuals
rgseas regARIMA user-defined seasonal component
rhol regARIMA holiday component
rsi final replacement values for SI ratios
rtd regARIMA trading day component
sa final seasonally adjusted data
sac final seasonally adjusted series with constant value added
safctd final seasonally adjusted series forecast decomposition (SEATS)
sar rounded final seasonally adjusted series
sass sliding spans of the seasonally adjusted series
sat final seasonally adjusted series with forced yearly totals
seataf final combined adjustment factors (SEATS)
seatdsa fully differenced transformed seasonal adjustment (SEATS)
seatdtr fully differenced transformed trend component (SEATS)
seatirr final irregular component (SEATS)
seatsa final seasonally adjusted series (SEATS)
seatsf final seasonal component (SEATS)
seatssm seasonal-period-length sums of final seasonal component (SEATS)
seattrn final trend component (SEATS)
setarat final adjustment ratios (SEATS)
setsac final seasonally adjusted series with constant value added (SEATS)
settadj total adjustment factors (SEATS)
settrc final trend cycle with constant value added (SEATS)
settrns final transitory component (SEATS)
sf final seasonal factors
238 APPENDIX A. GRAPHICS CODES
Code Description
sffctd final seasonal component forecast decomposition (SEATS)
sfhst concurrent and projected seasonal component and their percent revisions
sfr seasonal factors, adjusted for user-defined seasonal regARIMA component
sfss sliding spans of the seasonal factors
sgsac squared gain of the concurrent seasonal adjustment filter
sgsaf squared gain of the symmetric seasonal adjustment filter
sgtrnc squared gain of the concurrent trend filter
sgtrnf squared gain of the symmetric trend filter
si final unmodified SI ratios
siox final unmodified SI ratios, with labels for outliers and extreme values
so regARIMA seasonal outlier component
spcsir spectrum of the irregular component (SEATS)
spcssa spectrum of the seasonally adjusted series (SEATS)
spexrsd spectrum of the extended residuals (SEATS)
spir spectrum of modified irregular series
spor spectrum of the original series
sprsd spectrum of the regARIMA model residuals
spsa spectrum of differenced seasonally adjusted series
tadj total adjustment factors
tc regARIMA temporary change outlier component
tdss sliding spans of the trading day factors
tprior temporary prior-adjustment factors
trancmp regARIMA transitory component
tranfcd final transitory component forecast decomposition (SEATS)
trn final trend cycle
trnfctd final trend component forecast decomposition (SEATS)
trnhst concurrent and revised Henderson trend-cycle values and revisions
tssac time shift of the concurrent seasonal adjustment filter
tstrnc time shift of the concurrent trend filter
usrdef regARIMA user-defined regression component
xcal final calendar factors from irregular component regression
xeastr X-11 Easter adjustment factors
xhol final holiday factors from irregular component regression
xtd final trading day factors from irregular component regression
xtrm final extreme value adjustment factors
yyss sliding spans of the year-to-year changes in the seasonally adjusted series
B Print and Save Tables
Contents
B.1 Print and Save Tables Available for X-13ARIMA-SEATS specs . . . . . . . . . . . . . . 239
B.2 Special Output Related to the seats Spec . . . . . . . . . . . . . . . . . . . . . . . . 248
B.3 Special tables Used to Save X-13ARIMA-SEATS Output as Percentages . . . . . . . . 249
Tables
B.2 Output Tables for Seats spec that can Only be Saved . . . . . . . . . . . . . . . . . . 248
B.3 X-13ARIMA-SEATS File Extensions for Special SEATS Saved Output . . . . . . . . . . 249
B.4 Tables That Can Be Saved as Percentages in the save Argument . . . . . . . . . . . 250
This appendix contains listing of the various output tables that can be printed or saved using the X-13-
ARIMA-SEATS program.
Table B.1 contains a list of tables that are available for printing and saving using the print and save
arguments of the individual specs.
A listing of special output tables associated with the seats spec is given in Appendix B.2.
A listing of special tables that can be saved as percentages is given in Table B.4 in Appendix B.3.
B.1 Print and Save Tables Available for X-13ARIMA-SEATS specs

Short gives a short name for the tables of the print and short arguments. This name is also used as a
file extension if the table is saved.
Save? indicates which tables can be saved (+) into a separate file with the save argument.
Brief indicates which tables are printed (+) when the brief print level is specified. See Section 3.2 for
more information on print levels.
Default indicates which tables are printed (+) by default.
Spec indicates which spec the tables are defined for.
Table B.1: Print and Save Tables
Name Short Save? Brief Default Spec

autochoice ach no yes yes automdl
239
240 APPENDIX B. PRINT AND SAVE TABLES
Table B.1: Print and Save Tables (continued)

autochoicemdl amd no no no automdl
autodefaulttests adt no no no automdl
autofinaltests aft no no no automdl
autoljungboxtest alb no no no automdl
bestfivemdl b5m no no no automdl
header hdr no yes yes automdl
unitroottest urt no yes yes automdl
unitroottestmdl urm no no no automdl
acf acf yes no yes check
acfplot acp no no yes check
acfsquared ac2 yes no yes check
acfsquaredplot ap2 no no yes check
histogram hst no no yes check
normalitytest nrm no no yes check
pacf pcf yes no no check
pacfplot pcp no no no check
adjcompositeplot b1p no no no composite
adjcompositesrs b1 yes yes yes composite
calendaradjcomposite cac yes no no composite
compositeplot cmp no no no composite
compositesrs cms yes yes yes composite
header hdr no yes yes composite
indadjsatot iaa yes yes yes composite
indadjustfac iaf yes no yes composite
indadjustmentratio i18 yes no no composite
indaoutlier iao yes no yes composite
indcalendar ica yes no yes composite
indcalendaradjchanges ie8 yes no yes composite
indcratio cri yes no yes composite
indforcefactor iff yes yes yes composite
indftestd8 idf no no yes composite
indirregular iir yes no yes composite
indirregularplot iip no no no composite
indlevelshift ils yes no yes composite
indmcdmovavg if1 yes no no composite
indmodirr ie3 yes no no composite
indmodoriginal ie1 yes no no composite
indmodsadj ie2 yes no no composite
indmovseasrat ims no no yes composite
indqstat if3 no yes yes composite
indreplacsi id9 no no yes composite
indresidualseasf irf no no yes composite
indrevsachanges i6a yes no yes composite
B.1. PRINT AND SAVE TABLES AVAILABLE FOR X-13ARIMA-SEATS SPECS 241

indrndsachanges i6r yes no yes composite
indrobustsa iee yes no no composite
indrratio rri yes no yes composite
indsachanges ie6 yes no yes composite
indsadjround irn yes yes yes composite
indseasadj isa yes yes yes composite
indseasadjplot iap no no no composite
indseasonal isf yes yes yes composite
indseasonaldiff isd yes yes yes composite
indseasonalplot isp no no no composite
indtest itt no yes yes composite
indtotaladjustment ita yes no no composite
indtrend itn yes no yes composite
indtrendchanges ie7 yes no yes composite
indtrendplot itp no no no composite
indunmodsi id8 yes no yes composite
indx11diag if2 no yes yes composite
indyrtotals ie4 yes no no composite
origchanges ie5 yes no yes composite
origwindsaplot ie0 no no no composite
outlieradjcomposite oac yes no no composite
prioradjcomposite ia3 yes no no composite
ratioplotindsa ir2 no no no composite
ratioplotorig ir1 no no no composite
armacmatrix acm yes no no estimate
averagefcsterr afc no no yes estimate
estimates est yes yes yes estimate
iterationerrors ite no no no estimate
iterations itr yes no no estimate
lformulas lkf no no no estimate
lkstats lks yes yes yes estimate
model mdl yes yes yes estimate
options opt no no yes estimate
regcmatrix rcm yes no no estimate
regressioneffects ref yes no no estimate
regressionresiduals rrs no no no estimate
residuals rsd yes no no estimate
roots rts yes no no estimate
forcefactor ffc yes yes yes force
revsachanges e6a yes no yes force
rndsachanges e6r yes no yes force
saround rnd yes yes yes force
seasadjtot saa yes yes yes force

backcasts bct yes no no forecast
forecasts fct yes no yes forecast
transformed ftr yes no yes forecast
transformedbcst btr yes no no forecast
variances fvr yes no no forecast
chngestimates che yes no no history
chngrevisions chr yes no yes history
chngsummary chs no yes yes history
fcsterrors fce yes yes yes history
fcsthistory fch yes no no history
header hdr no yes yes history
indsaestimates iae yes no no history
indsarevisions iar yes no yes history
indsasummary ias no yes yes history
lkhdhistory lkh yes yes yes history
outlierhistory rot yes yes yes history
saestimates sae yes no no history
sarevisions sar yes no yes history
sasummary sas no yes yes history
seatsmdlhistory smh yes yes yes history
sfestimates sfe yes no no history
sfilterhistory sfh yes no no history
sfrevisions sfr yes no yes history
sfsummary sfs no yes yes history
trendchngestimates tce yes no no history
trendchngrevisions tcr yes no yes history
trendchngsummary tcs no yes yes history
trendestimates tre yes no no history
trendrevisions trr yes no yes history
trendsummary trs no yes yes history
acf iac yes yes yes identify
acfplot acp no no yes identify
pacf ipc yes yes yes identify
pacfplot pcp no no yes identify
regcoefficients rgc no no no identify
finaltests fts yes no no outlier
header hdr no no yes outlier
iterations oit yes no no outlier
temporaryls tls no yes yes outlier
tests ots no no no outlier
header hdr no yes yes pickmdl
pickmdlchoice pch no yes yes pickmdl
usermodels umd no yes yes pickmdl

aictest ats no yes yes regression
aoutlier ao yes yes yes regression
chi2test cts no yes yes regression
dailyweights tdw no no no regression
holiday hol yes yes yes regression
levelshift ls yes yes yes regression
outlier otl yes yes yes regression
regressionmatrix rmx yes no no regression
regseasonal a10 yes yes yes regression
seasonaloutlier so yes yes yes regression
temporarychange tc yes yes yes regression
tradingday td yes yes yes regression
transitory a13 yes yes yes regression
userdef usr yes yes yes regression
adjustfac s16 yes no no seats
adjustmentratio s18 yes no no seats
diffseasonaladj dsa yes no no seats
difftrend dtr yes no no seats
irregular s13 yes no no seats
seasadjconst sec yes no no seats
seasonal s10 yes no no seats
seasonaladj s11 yes no no seats
seasonaladjfcstdecomp afd yes no no seats
seasonalfcstdecomp sfd yes no no seats
seasonalsum ssm yes no no seats
seriesfcstdecomp ofd yes no no seats
totaladjustment sta yes no no seats
transitory s14 yes no no seats
transitoryfcstdecomp yfd yes no no seats
trend s12 yes no no seats
trendconst stc yes no no seats
trendfcstdecomp tfd yes no no seats
adjoriginal b1 yes yes yes series
adjorigplot b1p no no no series
calendaradjorig a18 yes no no series
header hdr no yes yes series
outlieradjorig a19 yes no no series
savefile sav no yes yes series
seriesmvadj mv yes yes yes series
seriesplot a1p no no no series
span a1 yes yes yes series
specfile spc yes yes yes series
chngspans chs yes no no slidingspans

factormeans fmn no no yes slidingspans
header hdr no yes yes slidingspans
indchngspans cis yes no no slidingspans
indfactormeans fmi no yes yes slidingspans
indpercent pci no yes yes slidingspans
indsaspans ais yes no no slidingspans
indsfspans sis yes no no slidingspans
indsummary smi no no yes slidingspans
indychngspans yis yes no no slidingspans
indyypercent piy no no no slidingspans
indyysummary siy no no no slidingspans
percent pct no yes yes slidingspans
saspans sas yes no no slidingspans
sfspans sfs yes no no slidingspans
ssftest ssf no no yes slidingspans
summary sum no no yes slidingspans
tdspans tds yes no no slidingspans
ychngspans ycs yes no no slidingspans
yypercent pcy no no no slidingspans
yysummary suy no no no slidingspans
speccomposite is0 yes yes yes spectrum
specindirr is2 yes yes yes spectrum
specindsa is1 yes yes yes spectrum
specirr sp2 yes yes yes spectrum
specorig sp0 yes yes yes spectrum
specresidual spr yes no yes spectrum
specsa sp1 yes yes yes spectrum
specseatsextresiduals ser yes yes yes spectrum
specseatsirr s2s yes yes yes spectrum
specseatssa s1s yes yes yes spectrum
aictransform tac no yes yes transform
permprior a2p yes no no transform
permprioradjusted a3p yes no no transform
permprioradjustedptd a4p yes no no transform
prior a2 yes yes yes transform
prioradjusted a3 yes no no transform
prioradjustedptd a4d yes no no transform
seriesconstant a1c yes yes yes transform
seriesconstantplot acp no no no transform
tempprior a2t yes no no transform
transformed trn yes no no transform
adjoriginalc c1 yes no no x11
adjoriginald d1 yes no no x11

adjustdiff fad yes yes yes x11
adjustfac d16 yes yes yes x11
adjustmentratio e18 yes no yes x11
autosf asf no no no x11
biasfactor bcf yes no no x11
calendar d18 yes yes yes x11
calendaradjchanges e8 yes no yes x11
combholiday chl yes yes yes x11
extreme c20 yes no no x11
extremeb b20 yes no no x11
ftestb1 b1f no no no x11
ftestd8 d8f no yes yes x11
irregular d13 yes no yes x11
irregularadjao iao yes no no x11
irregularb b13 yes no no x11
irregularc c13 yes no no x11
irregularplot irp no no no x11
irrwt c17 yes no yes x11
irrwtb b17 yes no no x11
mcdmovavg f1 yes no no x11
modirregular e3 yes no no x11
modoriginal e1 yes no no x11
modseasadj e2 yes no no x11
modsic4 c4 yes no no x11
modsid4 d4 yes no no x11
movseasrat d9a no no yes x11
origchanges e5 yes no yes x11
origwsaplot e0 no no no x11
qstat f3 no yes yes x11
ratioplotorig ra1 no no no x11
ratioplotsa ra2 no no no x11
replacsi d9 yes no yes x11
replacsib4 b4 no no no x11
replacsib9 b9 no no no x11
replacsic9 c9 yes no no x11
residualseasf rsf no no yes x11
robustsa e11 yes no no x11
sachanges e6 yes no yes x11
seasadj d11 yes yes yes x11
seasadjb11 b11 yes no no x11
seasadjb6 b6 yes no no x11
seasadjc11 c11 yes no no x11
seasadjc6 c6 yes no no x11

seasadjconst sac yes yes yes x11
seasadjd6 d6 yes no no x11
seasadjplot sap no no no x11
seasonal d10 yes yes yes x11
seasonaladjregsea ars yes yes yes x11
seasonalb10 b10 yes no no x11
seasonalb5 b5 yes no no x11
seasonalc10 c10 yes no no x11
seasonalc5 c5 yes no no x11
seasonald5 d5 yes no no x11
seasonaldiff fsd yes yes yes x11
seasonalplot sfp no no no x11
sib3 b3 yes no no x11
sib8 b8 yes no no x11
tdadjorig c19 yes no no x11
tdadjorigb b19 yes no no x11
tdaytype tdy no yes yes x11
totaladjustment tad yes no yes x11
trend d12 yes no yes x11
trendadjls tal yes no no x11
trendb2 b2 yes no no x11
trendb7 b7 yes no no x11
trendc2 c2 yes no no x11
trendc7 c7 yes no no x11
trendchanges e7 yes no yes x11
trendconst tac yes no no x11
trendd2 d2 yes no no x11
trendd7 d7 yes no no x11
trendplot trp no no no x11
unmodsi d8 yes no yes x11
unmodsiox d8b yes no yes x11
x11diag f2 no yes yes x11
yrtotals e4 yes no yes x11
calendar xca yes yes yes x11regression
calendarb bxc yes no no x11regression
combcalendar xcc yes yes yes x11regression
combcalendarb bcc yes no no x11regression
combtradingday c18 yes yes yes x11regression
combtradingdayb b18 yes no no x11regression
extremeval c14 yes no yes x11regression
extremevalb b14 yes no no x11regression
holiday xhl yes yes yes x11regression
holidayb bxh yes no no x11regression

outlierfinaltests xft no no no x11regression
outlierhdr xoh no no yes x11regression
outlieriter xoi yes no no x11regression
outliertests xot no no no x11regression
priortd a4 yes yes yes x11regression
tradingday c16 yes yes yes x11regression
tradingdayb b16 yes no no x11regression
x11reg c15 yes no yes x11regression
x11regb b15 yes no no x11regression
xaictest xat no yes yes x11regression
xregressioncmatrix xrc yes no no x11regression
xregressionmatrix xrm yes no no x11regression
B.2 Special Output Related to the seats Spec

Tables B.2 and B.3 give examples of special types of output files that can be produced by the seats spec.
Table B.2 gives a listing of tables that can only be saved by the program by using the save argument within
the seats spec.
Table B.2: Output Tables for Seats spec that can Only be Saved

componentmodels mdc models for the components
filtersaconc fac concurrent finite seasonal adjustment filter
filtersasym faf symmetric finite seasonal adjustment filter
filtertrendconc ftc concurrent finite trend filter
filtertrendsym ftf symmetric finite trend filter
pseudoinnovtrend pic pseudo-innovations of the trend component
pseudoinnovseasonal pis pseudo-innovations of the seasonal component
psuedoinnovtransitory pit pseudo-innovations of the transitory component
psuedoinnovsadj pia pseudo-innovations of the final SEATS seasonal adjustment
squaredgainsaconc gac squared gain for finite concurrent seasonal adjustment filter
squaredgainsasym gaf squared gain for finite symmetric seasonal adjustment filter
squaredgaintrendconc gtc squared gain for finite concurrent trend filter
squaredgaintrendsym gtf squared gain for finite symmetric trend filter
timeshiftsaconc tac time shift for finite concurrent seasonal adjustment filter
timeshifttrendconc ttc time shift for finite concurrent trend filter
wkendfilter wkf end filters of the semi-infinite Wiener-Kolmogorov filter
Note that the out argument can control whether one of these tables can be saved; Section 7.14 has more
information.
X-13ARIMA-SEATS can also save other output files that were produced by the original SEATS program.
These output files can contain forecasts, components or diagnostics generated from the SEATS model-based
adjustment performed. Table B.3 shows the file extensions that are used to save the corresponding special
output file from SEATS in the same way the short table names are used as file extensions in storing individual
tables to separate files.
B.3. SPECIAL TABLES USED TO SAVE X-13ARIMA-SEATS OUTPUT AS PERCENTAGES 249
Table B.3: X-13ARIMA-SEATS File Extensions for Special SEATS Saved Output
SEATS X-13ARIMA-SEATS Produced when Contents

file name extension out=2? of file
f.txt tbf no Forecasts of the series, seasonally
adjusted series, and components of
the model-based seasonal adjustment,
saved in columns separated by white
space
o.txt tbo no Series, seasonally adjusted series, and
components of the model-based sea-
sonal adjustment, saved in columns
separated by white space
rogtable.out rog yes Selected statistics from the growth
rate output
summarys.txt sum yes Summary information and diagnostics
from SEATS adjustment
table-s.out tbs yes Annotated listing of the series, the sea-
sonally adjusted series, and compo-
nents of the model-based seasonal ad-
justment, saved in columns separated
by white space
z.txt tbz no Log of the series, seasonally adjusted
series, and components of the model-
based seasonal adjustment, saved in
columns separated by white space
SEATS file name gives the file name saved by the SEATS program.
X-13ARIMA-SEATS extension gives the file extension used to save the output from the corre-
sponding SEATS output file.
B.3 Special tables Used to Save X-13ARIMA-SEATS Output as Percent-

ages
Table B.4 gives table names and abbreviations that can be used with the save argument to save certain tables as
percentages rather than ratios. The percentages are only produced when multiplicative or log-additive seasonal
adjustment is specified by the user in the mode argument of the x11 spec (or a log transformation is specified
in the transform spec in the case of the tables from the seats spec).
Table B.4: Tables That Can Be Saved as Percentages in the save Argument
name short spec description of table

indadjustfacpct ipa composite indirect combined adjustment factors expressed as
percentages if appropriate
indcalendaradjchangespct ip8 composite percent changes in original series adjusted for calen-
dar effects
indirregularpct ipi composite indirect irregular component expressed as percent-
ages if appropriate
indrevsachangespct ipf composite percent changes for indirect seasonally adjusted se-
ries with forced yearly totals
indrndsachangespct ipr composite percent changes for rounded indirect seasonally ad-
justed series
indsachangespct ip6 composite percent changes for indirect seasonally adjusted se-
ries
indseasonalpct ips composite indirect seasonal component expressed as percent-
ages if appropriate
indtrendchangespct ip7 composite percent changes for indirect trend component
origchangespct ip5 composite percent changes for composite series
revsachangespct p6a force percent changes in seasonally adjusted series with
forced yearly totals
rndsachangespct p6r force percent changes in rounded seasonally adjusted series
adjustfacpct psa seats combined adjustment factors, expressed as percent-
ages if appropriate
irregularpct psi seats final irregular component, expressed as percentages
if appropriate
transitorypct psc seats final transitory component, expressed as percentages
if appropriate
seasonalpct pss seats final seasonal factors, expressed as percentages if ap-
propriate
adjustfacpct paf x11 combined adjustment factors, expressed as percent-
ages if appropriate
calendaradjchangespct pe8 x11 percent changes in original series adjusted for calen-
dar factors
irregularpct pir x11 final irregular component, expressed as percentages
if appropriate
origchangespct pe5 x11 percent changes in the original series
sachangespct pe6 x11 percent changes in seasonally adjusted series
seasonalpct psf x11 final seasonal factors, expressed as percentages if ap-
propriate
trendchangespct pe7 x11 percent changes in final trend cycle
Spec indicates which spec the tables are defined for.
C Irregular-component Regression
Models Used
Contents
C.1 Irregular regression models for multiplicative decompositions. . . . . . . . . . . . 251
C.1.1 Obtaining separate trading day and holiday factors . . . . . . . . . . . . . . . . . . . . 253
C.1.2 Estimating only holiday effects or stock trading day effects. . . . . . . . . . . . . . . . 254
C.1.3 Estimating user-defined flow trading day and/or holiday effects . . . . . . . . . . . . . 254
C.2 Irregular regression models for other decomposition modes . . . . . . . . . . . . . 255
C.2.1 Additive Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
C.2.2 Pseudo-Additive Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
C.2.3 Log-Additive Decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
C.3 When tdprior is used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
This appendix gives details of the models applied by the X-13ARIMA-SEATS x11regression spec to the
irregular component to estimate calendar effects. The regression models detailed in this appendix are used to
estimate a calendar effect, or other effect, from the irregular component It of a preliminary seasonal adjustment
that did not adjust for the effect. The estimation is done by ordinary least squares (OLS) applied to a regression
model for the effect. In the simplest cases detailed below, the model has the form
It − 1.0 = β 0 Xt + et ,
where Xt is a regression vector with variables that describe the basic effect of interest.
C.1 Irregular regression models for multiplicative decompositions.

The irregular component is presumed to have no seasonality or trend (beyond a constant level of 1.0, in the case
of a multiplicative decomposition). Hence, the regressors that are used in regression models for the irregulars
should usually not have a seasonal or trend component. For this reason, most trading day and Easter regression
functions used in the regression spec (see Tables 4.1 and 7.27) are modified for use in the x11regression
spec (see Table 7.56). The modifications for trading day variables for the various types of seasonal adjustment
decompositions are derived in section 1.4 of Findley, Monsell, Bell, Otto, and Chen (1998). We will indicate the
nature of this modification with a combined monthly flow trading day and holiday regression function of the
form
X6
γ0 mt + γj (dj,t − d7,t ) + δ 0 Ht , (C.1)
j=1
where dj,t = no. of weekdays of type j in month t (with j = 1, . . . 7 denoting Monday,. . ., Sunday, respectively),
P7
mt = j=1 dj,t (the length on month t in days), and Ht denotes a (column) vector of holiday regressors.
251
252 APPENDIX C. IRREGULAR-COMPONENT REGRESSION MODELS USED
Because of the definition of the calendar, over most time intervals of interest these variables are periodic,
mt+48 = mt, dj,t+336 = dj,t , and Ht+P = Ht with P depending on the holiday variables included in Ht . (If
all proposed corrections to the Gregorian calendar are used, the Easter calendar has a period of 38, 000 years
= 456, 000 months. For this reason it is often more practical to choose P so that approximate periodicity holds,
Ht+P ' Ht ).
If ft is an approximately periodic function of period 12p months, ft+12p ' ft , then its (approximate) com-
bined seasonal and level component is given by its calendar month means
p
1X
ft∗ = ft+12j ,
p j=1
∗
which is approximately periodic with period 12 months, ft+12 ' ft∗ . If seasonal and level effects are removed
from ft by division, the resulting deseasonalized, level-neutral component of ft is ft /ft∗ . To apply these ideas
to the function (C.1) above, we note that if p is a multiple of 28, then d∗j,t = d∗7,t , 1 ≤ j ≤ 6, with the result that
the seasonal and level component of this calendar effect function is
γ0 m∗t + δ 0 H∗t ,
with (
mt , mt = 30, 31
m∗t = .
28.25 , mt = 28, 29
Therefore, if a time series contains a trading day and holiday component of the form (C.1), then its irregular
component from multiplicative deseasonalization and detrending can be expected to have a trading day and
holiday component close to
mt P6 Ht
+ j=1 αj ((dj,t − d7,t )/m∗t ) + β 0 m
P6
γ0 mt + j=1 γj (dj,t − d7,t ) + δ 0 Ht m∗
t
∗
t
∗ 0 ∗ = H ∗ . (C.2)
γ0 m t + δ H t 0
1 + β m∗ t
t
The expression on the right is a nonlinear function of αj = γj /γ0 and β = δ/γ0 . However, because trading
day effects and holiday effects are usually in the range of a few percent, the approximation
−1
H∗t H∗t

1 + β0 ' 1 − β0
m∗t m∗t
can be applied to (C.2). After multiplying the numerator on the right in (C.2) by this factor, the terms
that involve products of coefficients are generally small enough to be ignored. This yields the following linear
approximation to (C.2),
6
Ht − H∗t

mt X dj,t − d7,t 0
+ αj ( ) + β . (C.3)
m∗t j=1
m∗t m∗t
In obtaining this approximation, we have also made use of

mt 1
=1+ (mt − m∗t ),
m∗t 28.25
C.1. IRREGULAR REGRESSION MODELS FOR MULTIPLICATIVE DECOMPOSITIONS. 253
and have treated the term involving the leap year variable LYt = mt − m∗t as one whose product with
β 0 (Ht − H∗t ) /m∗t is negligible. The formula (C.3) suggests a linear regression model for the irregular com-
ponent It of the form
6
Ht − H∗t

mt X dj,t − d7,t 0
It − ∗ = αj +β + κ0 AOt + et ,
mt j=1
m∗t m∗t
where AOt , denotes a regression vector containing any needed additive outlier variables. Instead of using this
model, X-13ARIMA-SEATS, for conformity with the X-11and X-11-ARIMAtrading day regression models, obtains
the coefficients in (C.3) from ordinary least squares estimation (OLS) of the rescaled model
6
X
m∗t It − mt = αj (dj,t − d7,t ) + β 0 (Ht − H∗t ) + κ0 AOt + εt (C.4)
j=1
whenever td is specified in the variables argument of x11regression, with one or more of the holiday effect
specifications easter[w], labor[w], and thank[w]. As explained in the footnote of Table 4.1, the regressors
associated with these holiday variables also have the deseasonalized form Ht − H∗t when they are estimated from
the regression spec. This is done so that seasonal effects occur only in the seasonal part of the model, and only
in the seasonal factors of the decomposition. For conformity with X-11-ARIMA/88, the regressors associated
with sceaster[w] are never deseasonalized. In effect, the entries of H∗t in (C.4) associated with any specified
sceaster[w] regressors are set to zero.
C.1.1 Obtaining separate trading day and holiday factors
The calendar factors (C.3) can be approximately factored as the product of holiday factors
Ht − H∗t

0
1+β (C.5)
m∗t
and trading day factors

6 P7
j=1 (1 + αj )dj,t

mt X dj,t − d7,t
+ αj = , (C.6)
m∗t j=1
m∗t m∗t
P6
(with α7 = − j=1 αj ). The numbers 1 + αj are called the daily weights. The trading day factor formula (C.6)
can also be written as P(5)
mt j αj
∗ + ,
mt m∗t
P(5)
where j denotes the sum of the j for which djt = 5. This formula shows that, apart from length of month
effects, the trading day effects depend only on the effects of the days that occur five times in the month. When
only trading day effects are estimated, the formulas above apply with β = 0.
If one or more trading day ”weights” 1 + αj are negative and the option reweight has been specified, then,
for the trading day factor calculation all αj < −1 are replaced by αj0 = −1 and all αj ≥ −1 are replaced by
αj0 = (1 + αj ) w − 1, where
 −1
 X 
w=7 (1 + αi ) ,
 
αi ≥−1
assuming no αj > −1 have been assigned fixed values using the b argument. If there are fixed values, only
unfixed αj > −1 are replaced, and in the replacement formula w is defined by
( )( )−1
X X
w= 7− (1 + αi ) (1 + αi ) ,
αi fixed αi notfixed
for all αi > −1.
C.1.2 Estimating only holiday effects or stock trading day effects.
If only holiday effects, or stock trading day effects, are specified in the variables argument of x11regression,
then X-13ARIMA-SEATS estimates these effects by OLS applied to a model of the form
It − 1 = β 0 (Ht − H∗t ) + κ0 AOt + et . (C.7)
respectively,
6
X
It − 1 = αj Dj,t + κ0 AOt + et . (C.8)
j=1
where the Dj,t are the regressors associated with the specified tdstock[w] in Table 4.1. These models lead to
calendar effect adjustment factors of the form
1 + β 0 (Ht − H∗t ) . (C.9)
respectively,
6
X
1+ αj Dj,t = 1 − αj(t) . (C.10)
j=1
where αj(t) is the coefficient of the w-th day of month t.
C.1.3 Estimating user-defined flow trading day and/or holiday effects
The regression model (C.4) yields mt /m∗t as the component of the mean function for the irregulars It that is
known independently of the estimated coefficients. This is also the default specification of the known component
when user-defined variables are used. If this default is accepted, then the OLS regression model with at least
one user-defined trading day or holiday variable has the form
m∗t It − mt = α0 TDt + β 0 H̃t + κ0 AOt + εt , (C.11)

C.2. IRREGULAR REGRESSION MODELS FOR OTHER DECOMPOSITION MODES 255
with TDt and H̃t denoting the vectors of trading day and holiday variables specified. User-defined variables
are input by way of file or data arguments. The program does not deseasonalize user-defined variables. They
should be input to the program in an appropriately deseasonalized form. X-13ARIMA-SEATS calculates calendar
factors
mt TDt H̃t
+ α0 ∗ + β 0 ∗ ,
m∗t mt mt
that are approximately factored into holiday factors and trading day factors in analogy with (C.5), and (C.6).
If only holiday effects are estimated, then the default known mean function component is the constant 1.0, and
the model and resulting holiday factors are the analogues of (C.7) and (C.9). Similarly, if only stock trading
day effects are estimated, then the default known mean function component is the constant 1.0, and the model
and resulting holiday factors are the analogues of (C.8) and (C.10).
When the default known mean functions just described are not appropriate, the user can input a mean
function µt by means of the umfile or umdata arguments. In this case, the regression model estimated is
It − µt = α0 TDt + β 0 H̃t + κ0 AOt + et , (C.12)
and only the calendar factors
µt + α0 TDt + β 0 H̃t
are produced. The coefficients α,β, estimated from (C.12) are on a different scale from those obtained from
(C.11), being smaller by roughly the factor
47
1 X 1
' .03288.
48 j=0 m∗t
The same approximate scale difference holds for calendar coefficients calculated from regression instead of
x11regression, or from (C.7) or (C.8) instead of (C.4).
C.2 Irregular regression models for other decomposition modes

We present below the models used with additive, pseudo-additive, and log-additive decompositions for the
case of combined flow trading day and holiday effect estimation with predefined regressors. The appropriate
modifications to these models for the case of user-defined, stock trading day or holiday regressors are analogous
to those described above for multiplicative decompositions.
C.2.1 Additive Decompositions

If mode=add in the x11 spec, calendar effects are estimated by OLS from a model of the form
6
X
It = α0 LYt + αj (dj,t − d7,t ) + β 0 (Ht − H∗t ) + κ0 AOt + et .
j=1
P6
The calendar effect is thus exactly the sum of the trading day effect α0 LYt + j=1 αj (dj,t − d7,t ) and the holiday
effect β 0 (Ht − H∗t ).
C.2.2 Pseudo-Additive Decompositions
If mode=pseudoadd in the x11 spec, then, with m = 30.4375 and LYt = mt −m∗t , calendar effects are estimated
by OLS from a model of the form
6
X
m (It − 1) − LYt = αj (dj,t − d7,t ) + β 0 (Ht − H∗t ) + κ0 AOt + εt .
j=1
The calendar effect factors

6
Ht − H∗t

1 X dj,t − d7,t 0
1 + LYt + αj +β
m j=1
m m
can be approximately factored as

 
6
Ht − H∗t

 1 X dj,t − d7,t
1 + LYt + αj 1 + β0
 m j=1
m  m
to obtain trading day and holiday factors.
C.2.3 Log-Additive Decompositions
If mode=logadd in the x11 spec, calendar effects are estimated by OLS from a model of the form
6
X
m∗t (log It + 1) − mt = αj (dj,t − d7,t ) + β 0 (Ht − H∗t ) + κ0 AOt + εt . (C.13)
j=1
These can be exactly factored into trading day and holiday factors,
 
6
Ht − H∗t

 mt X dj,t − d7,t 0
exp −1 + ∗ + αj exp β .
 mt j=1
m∗t  m∗t
Two other useful forms for the trading day factors appear in the identity
 
6 
mt  X dj,t − d7,t
exp −1 + ∗ exp αj
mt 
j=1
m∗t 
 
6
 X dj,t 
exp −1 + (1 + αj ) ∗ , (C.14)

j=1
mt 
which emphasizes the leap year factors exp {−1 + mt /m∗t } ' mt /m∗t and the daily weights (1 + αj ).
C.3. WHEN TDPRIOR IS USED 257
C.3 When tdprior is used

Any of the coefficients in the models above can be assigned fixed values by an appropriate specification of the b
argument. Sometimes users have prior information that suggests values for the seven daily weights associated
with the trading factors (C.6) or (C.13) of multiplicative, respectively, log-additive adjustment. When “prior”
(p)
daily weights 1 + αj , 1 ≤ j ≤ 7 are assigned values by means of the tdprior argument, the series is preadjusted
by  
P7 (p) 6
(1 + α )d

j=1 j j,t m t
 X (p) d j,t − d7,t
= ∗ 1+ αj (C.15)
m∗t mt  mt
j=1

when mode=mult in x11, or by

 
6
mt X
(p) dj,t − d7,t 
exp −1 + ∗ exp αj (C.16)
mt 
j=1
m∗t 
when mode=logadd. One advantage of using tdprior instead of b is that the user can also invoke aictest to
automatically test whether significant trading day effects still occur in the irregular component of the preadjusted
series and to calculate adjustment factors for removing any remaining effects. However, the fact that prior
adjustment by (C.15) or (C.16) removes the leap year effect mt /m∗t , respectively exp {−1 + mt /m∗t }, makes
it necessary, when td is specified in the variables argument, to modify the models (C.4) and (C.13) used by
x11regression for estimating remaining effects. When mode=mult, the model
6
X
mt It − mt = αj (dj,t − d7,t ) + β 0 (Ht − H∗t ) + κ0 AOt + εt ,
j=1
is used in place of (C.4), and, when mode=logadd,

6
X
m∗t log It = αj (dj,t − d7,t ) + β 0 (Ht − H∗t ) + κ0 AOt + εt .
j=1
instead of (C.13). The first model yields the calendar factors

6
Ht − H∗t

X dj,t − d7,t
1+ αj + β0 ,
j=1
mt mt
from which combined calendar factors are formed by multiplication with (C.15). The result is approximately
 
7 ∗ 

mt X (p) d j,t Ht − H
(1 + αj + αj ) + β0 t
. (C.17)
m∗t  j=1
mt mt 
The second model yields the calendar factors

 
6 ∗ 

X dj,t − d7,t H t − H
exp αj ∗ + β0 ∗
t
,

j=1
m t mt 
and multiplication by (C.16) yields the combined factors

 
6
 X (p) dj,t 
exp −1 + (1 + αj + αj ) ∗ . (C.18)

j=1
mt 
The formulas (C.17) and (C.18) show that a statistically significant t-statistic for a coefficient αj can be inter-
(p)
preted as meaning that the prior weight 1 + αj needs significant revision.
Bibliography
Abraham, B. and G. Box (1978). Deterministic and forecast-adaptive time-dependent models. Applied Statis-
tics 27, 120–130.
Abraham, B. and J. Ledolter (1983). Statistical Methods for Forecasting. New York: John Wiley and Sons.
Akaike, H. (1973). Information theory and an extension of the likelihood principle. In B. Petrov and F. Czaki
(Eds.), Second International Symposium on Information Theory, pp. 267–287. Budapest: Akademia Kiado.
Akaike, H. (1980). Seasonal adjustment by a bayesian modeling. Journal of Time Series Analysis 1, 1–14.
Akaike, H. and M. Ishiguro (1980). BAYSEA, a bayesian seasonal adjustment program. Computer Science
Monographs No. 13, Tokyo: The Institute for Statistical Mathematics.
Baxter, M. A. (1994). A guide to seasonal adjustment of monthly data with X-11 (third edition). Central
Statistical Office, United Kingdom.
Bednarek, M. (2007). The dates of easter for 500 years. [Online]. Available: http://home.swiftdsl.com.
au/~mbednarek//easter.php [2007, November 18].
Bell, W. R. (1983). A computer program for detecting outliers in time series. Proceedings of the American
Statistical Association, Business and Economic Statistics Section, 634–639.
Bell, W. R. (1984). Seasonal decomposition of deterministic effects. SRD Research Report No. 84/01, U. S.
Census Bureau http://www.census.gov/srd/papers/pdf/rr84-1.pdf.
Bell, W. R. (1987). A note on overdifferencing and the equivalence of seasonal time series models with
monthly means and models with (0,1,1)12 seasonal parts when θ = 1. Journal of Business and Economic
Statistics 5, 383–387. http://www.census.gov/srd/papers/pdf/rr84-32.pdf.
Bell, W. R. (1995). Correction to seasonal decomposition of deterministic effects. SRD Research Report No.
95/01, U. S. Census Bureau http://www.census.gov/srd/papers/pdf/rr95-01.pdf.
Bell, W. R. (2004). On RegComponent time series models and their applications. In A. Harvey, S. J. Koopman,
and N. Shephard (Eds.), State Space and Unobserved Component Models: Theory and Applications (Pro-
ceedings of a Conference in Honour of James Durbin), pp. 248–283. Cambridge: Cambridge University
Press.
Bell, W. R. and S. Hillmer (1988). A matrix approach to likelihood evaluation and signal extraction for
ARIMA component time series models. Research Report Number 1988/22, Statistical Research Division,
U.S. Census Bureau http://www.census.gov/srd/papers/pdf/rr88-22.pdf.
Bell, W. R. and S. C. Hillmer (1983). Modeling time series with calendar variation. Journal of the American
Statistical Association 78, 526–534.
259
260 BIBLIOGRAPHY
Bell, W. R. and S. C. Hillmer (1984). Issues involved with the seasonal adjustment of economic time series.
Journal of Business and Economic Statistics 2, 299–320. (invited paper with discussion).
Bell, W. R. and S. C. Hillmer (1985). A reply (to comments by Sims in Bell and Hillmer(1984)). Journal of
Business and Economic Statistics 3, 95–97.
Berk, K. N. (1974). Consistent autoregressive spectral estimates. Annals of Statistics 2, 489–502.
Bobbitt, L. and M. C. Otto (1990). Effects of forcasts on the revisions of seasonally adjusted values using the
X-11 seasonal adjustment procedure. Proceedings of the American Statistical Association, Business and
Economic Statistics Section, 449–453. http://www.census.gov/srd/papers/pdf/rr90-09.pdf.
Box, G. and D. Cox (1964). An analysis of transformations. Journal of Royal Statistical Society B 26, 211–252.
Box, G. E. P. and G. M. Jenkins (1976). Time Series Analysis: Forecasting and Control (2nd ed.). San
Francisco, CA: Holden-Day.
Box, G. E. P. and G. C. Tiao (1975). Intervention analysis with applications to economic and environmental
problems. Journal of the American Statistical Association 70, 70–79.
Bozik, J. E. and M. C. Otto (1988). Benchmarking: Evaluating methods that preserve month-to-month
changes. Research Report Number 88/07, Statistical Research Division, U.S. Census Bureau http://
www.census.gov/srd/papers/pdf/rr88-07.pdf.
Brillinger, D. R. (1975). Time Series–Data Analysis and Theory. New York: Holt Rinehart and Winston.
Brockwell, P. J. and R. A. Davis (1991). Time Series: Theory and Methods. New York: Springer-Verlag.
Burnham, K. P. and D. R. Anderson (2004). Multimodal inference: Understanding AIC and BIC in model
selection. Sociological Methods & Research 33, 261–304.
Causey, B. and M. L. Trager (1982). Derivation of solution to the benchmarking problem: Trend revision.
Unpublised research notes, U.S. Census Bureau, reproduced in Appendix C of Bozik and Otto (1988) -
http://www.census.gov/srd/papers/pdf/rr88-07.pdf.
Chang, I. and G. C. Tiao (1983). Estimation of time series parameters in the presence of outliers. Technical
Report No. 8, Statistics Research Center, University of Chicago.
Chang, I., G. C. Tiao, and C. Chen (1988). Estimation of time series parameters in the presence of outliers.
Technometrics 30, 193–204.
Chen, Z. (1985). The asymptotic efficiency of a linear procedure of estimation for ARMA models. Journal of
Time Series Analysis 6, 52–62.
Chen, Z.-G. and P. Durk (2005). A therapy for ill-seasonality in X-11 seasonal adjustment. Statistics Canada
Working Paper BSMD-2005-009E.
Cholette, P. and E. B. Dagum (1994). Benchmarking time series with autocorrelated survey erros. Interna-
tional Statistical Review 62, 367–377.
Cholette, P. A. (1978). A comparison and assessment of various adjustment methods of sub-annual series to
yearly benchmarks. Research Paper, Seasonal Adjustment and Time Series Staff, Statistics Canada.
Cholette, P. A. (1984). Adjusting sub-annual series to benchmarks. Survey Methodology 10, 35–49.
Chow, J. and K. Moore (2009). Constructing an Easter regressor for a stock series in X-12-ARIMA. Proceedings
of the American Statistical Association, Business and Economic Statistics Section. [CD-ROM] (to appear).
BIBLIOGRAPHY 261
Cleveland, W. P. and M. R. Grupe (1983). Modeling time series when calendar effects are present. In Pro-
ceedings of the Conference on Applied Time Series Analysis of Economic Data, pp. 57–73. Washington,
DC: U. S. Census Bureau.
Cleveland, W. S. and S. J. Devlin (1980). Calendar effects in monthly time series: Detection by spectrum
analysis and graphical methods. Journal of the American Statistical Association 75, 487–496.
Dagum, E. B. (1980). The X-11-ARIMA seasonal adjustment method. Statistics Canada.
Dagum, E. B. (1988). X-11-ARIMA/88 seasonal adjustment method - Foundations and users manual. Statis-
tics Canada.
Dagum, E. B. (1996). A new method to reduce unwanted ripples and revisions in trend-cycle estimates from
X-11-ARIMA. Survey Methodology 22, 77–83.
Dagum, E. B. and P. A. Cholette (2006). Benchmarking, Temporal Distribution, and Reconciliation Methods
for Time Series. New York: Springer. Lecture Notes in Statistics, Vol. 186.
Dagum, E. B. and A. Luati (2009). A cascade linear filter to reduce revisions and false turning points for real
time trend-cycle estimation. Econometric Reviews 28 (1), 40–59. [Online]. Available: http://www2.stat.
unibo.it/beedagum/Papers/Dagum_Luati_2009_ER.pdf [2009, January 9].
Darne, O. and E. B. Dagum (2009). Performance of short-term trend predictors for current economic
analysis. Economics Bulletin. [Online]. Available: http://www2.stat.unibo.it/beedagum/Papers/
Darne-Dagum_EB09.pdf [2009, February 10].
den Butter, F. A. G. and M. M. G. Fase (1991). Seasonal Adjustment as a Practical Problem. Amsterdam:
North Holland.
Doornik, J. A. and D. Hendry (2001). GiveWin: An Interface to Empirical Modelling (3rd ed.). London:
Timberlake Consultants Press.
Duffet-Smith, P. (1981). Practical Astronomy With Your Calculator (2nd ed.). Cambridge University Press.
Feldpausch, R. M., C. C. H. Hood, and K. C. Wills (2004). Diagnostics for model-based seasonal adjustment.
Proceedings of the American Statistical Association, Business and Economic Statistics Section. [CD-ROM]
http://www.census.gov/ts/papers/jsm2004rmf.pdf.
Findley, D., T. S. McElroy, and K. C. Wills (2005). Modifications of SEATS diagnostic for detecting
over- and underestimation of seasonal adjustment components. http://www.census.gov/ts/papers/
findleymcelroywills2005.pdf.
Findley, D. F. (1985). On the unbiasedness property of AIC for exact or approximating linear stochastic time
series models. Journal of Time Series Analysis 6, 229–252.
Findley, D. F. (1999). Akaike’s Information Criterion II. In S. Kotz, C. B. Read, and D. L. Banks (Eds.),
Encyclopedia of Statistical Science, Update Volume 3, pp. 2–6. New York: John Wiley and Son.
Findley, D. F. (2005). Asymptotic stationarity properties of out-of-sample forecast errors of misspecified
regARIMA models and the optimality of GLS for one-step-ahead forecasting. Statistica Sinica 15 (2),
447–476.
Findley, D. F. (2007). Optimality of GLS for one-step-ahead forecasting with regARIMA and related models
when the regression is misspecified. Econometric Theory 23 (6), 1083–1107.
Findley, D. F. (2009). Stock series holiday regressors generated by flow series holiday regressors. SRD Research
Report RRS 2009-4, U. S. Census Bureau http://www.census.gov/srd/papers/pdf/rrs2009-04.pdf.
262 BIBLIOGRAPHY
Findley, D. F. and C. C. Hood (1999). X-12-ARIMA and its application to some Italian indicator series. In
Seasonal Adjustment Procedures – Experiences and Perspectives, pp. 231–251. Rome: Istituto Nazionale
di Statistica (ISTAT). http://www.census.gov/ts/papers/x12istat.pdf.
Findley, D. F. and D. E. K. Martin (2003). Frequency domain analysis of SEATS and X-11/X-12-ARIMA sea-
sonal adjustment filters for short and moderate length time series. Research Report Number 2003/02, Sta-
tistical Research Division, U.S. Census Bureau http://www.census.gov/srd/papers/pdf/rrs2003-02.
pdf.
Findley, D. F. and D. E. K. Martin (2006). Frequency domain analysis of SEATS and X-11/X-12-ARIMA
seasonal adjustment filters for short and moderate length time series. Journal of Official Statistics 22 (1),
1–34.
Findley, D. F. and B. C. Monsell (1986). New techniques for determining if a time series can be seasonally
adjusted reliably, and their application to U. S. foreign trade series. In M. Perryman (Ed.), Regional
Economic Modeling, pp. 195–228. Amsterdam: Kluwer-Nijhoff.
Findley, D. F. and B. C. Monsell (2009). Modeling stock trading day effects under flow day-of-week constraints.
Journal of Official Statistics 25 (3), 415–430. http://www.jos.nu/Articles/abstract.asp?article=
253415 [Online; accessed 14-October-2009].
Findley, D. F., B. C. Monsell, W. R. Bell, M. C. Otto, and B. C. Chen (1998). New capabilities of the X-12-
ARIMA seasonal adjustment program (with discussion). Journal of Business and Economic Statistics 16,
127–77. http://www.census.gov/ts/papers/jbes98.pdf.
Findley, D. F., B. C. Monsell, H. B. Shulman, and M. G. Pugh (1990). Sliding spans diagnostics for seasonal
and related adjustments. Journal of the American Statistical Association 85, 345–355.
Findley, D. F. and R. J. Soukup (2000). Modeling and model selection for moving holidays. Proceedings of
the American Statistical Association, Business and Economic Statistics Section, 102–107. http://www.
census.gov/ts/papers/asa00_eas.pdf.
Findley, D. F. and C. Z. Wei (2002). AIC, overfitting principles, and the boundedness of moments of inverse
matrices for autoregressions and related models. Journal of Multivariate Analysis 83, 415–450.
Findley, D. F., K. C. Wills, J. A. D. Aston, R. Feldpausch, and C. C. Hood (2003). Diagnostics for ARIMA
model-based seasonal adjustment. Proceedings of the American Statistical Association, Business and Eco-
nomic Statistics Section. [CD-ROM].
Fuller, W. A. (1976). Introduction to Statistical Time Series. New York: John Wiley and Son.
Gastwirth, J. L. and M. G. B. Owens (1977). On classical test of normality. Biometrika 64, 135–139.
Geary, R. C. (1936). Moments of the ratio of mean deviation to the standard deviation as a test of normality.
Biometrika 28, 295–305.
Geweke, J. (1978). Revision of seasonally adjusted time series. SSRI Report No. 7822, University of Wisconsin,
Department of Statistics.
Gómez, V. (1998). Automatic model identification in the presence of missing observations and outliers. Work-
ing Paper D-98009, Ministerio de Economı́a y Hacienda, Dirrectión General de Análisis y Programación
Presupuestaria, Madrid.
Gómez, V. and A. Maravall (1996). Programs TRAMO and SEATS : Instructions for the user (beta version:
June 1997). Banco de España, Servicio de Estudios, DT 9628. (Updates and additional documentation
can be found at http://www.bde.es/webbde/es/secciones/servicio/software/econom.html.).
BIBLIOGRAPHY 263
Gómez, V. and A. Maravall (2001a). Automatic modeling methods for univariate series. In D. Pena, G. C.
Tiao, and R. S. Tsay (Eds.), A Course in Time Series Analysis. New York, NY: J. Wiley and Sons.
Gómez, V. and A. Maravall (2001b). Seasonal adjustment and signal extraction in economic time series. In
D. Pena, G. C. Tiao, and R. S. Tsay (Eds.), A Course in Time Series Analysis. New York, NY: J. Wiley
and Sons.
Griswold, R. E. and M. T. Griswold (1997). The ICON Programming Language (3rd ed.). San Jose: Peer-to-
Peer Communications.
Hampel, F. R., E. M. Ronchetti, P. J. Rousseeuw, and W. A. Stahel (1986). Robust Statistics: The Approach
Based on Influence Functions. New York, NY: J. Wiley and Sons.
Hannan, E. J. and B. G. Quinn (1979). The determination of the order of an autoregression. Journal of Royal
Statistical Society B 41, 190–195.
Hannan, E. J. and J. Rissanen (1982). Recursive estimation of mixed autoregressive-moving average models.
Hillmer, S. C. and G. C. Tiao (1979). Likelihood function of stationary multiple autoregressive moving average
models,. Journal of the American Statistical Association 74, 652–660.
Hood, C. C. (2002a). Comparing the automatic ARIMA model selection procedures of TRAMO and X-
12-ARIMA Version 0.3 and the seasonal adjustments of SEATS and X-12-ARIMA. unpublished work
presented at the Eurostat Working Group on Seasonal Adjustment Meeting, Luxembourg, April 2002.
Hood, C. C. (2002b). Comparison of time series characteristics for seasonal adjustments from SEATS and
X-12-ARIMA. American Statistical Association, Proceedings of the Business and Economic Statistics
Section [CD-ROM] http://www.census.gov/ts/papers/choodasa2002.pdf.
Hood, C. C. (2002c). X-12-Graph: A SAS/GRAPH program for X-12-ARIMA output, user’s guide for the X-
12-Graph interactive for PC/Windows, version 1.2. U. S. Census Bureau, U. S. Department of Commerce.
Hood, C. C., J. D. Ashley, and D. F. Findley (2000). An empirical evaluation of the performance of
TRAMO/SEATS on simulated series. American Statistical Association, Proceedings of the Business and
Economic Statistics Section http://www.census.gov/ts/papers/asa00_ts.pdf.
Hood, C. C. H. (2005). An empirical comparison of methods for benchmarking seasonally adjusted series
to annual totals. Proceedings of the American Statistical Association, Business and Economic Statistics
Section. [CD-ROM] http://www.census.gov/ts/papers/chood_asa2005.pdf.
Huot, G. (1975). Quadratic minimization adjustment of monthly or quarterly series to annual totals. Research
Paper, Seasonal Adjustment and Time Series Staff, Statistics Canada.
Hurvich, C. M. and C. Tsai (1989). Regression and time series model selection in small samples. Biometrika 76,
297–307.
Kaiser, R. and A. Maravall (2001). Measuring Business-Cycles in Economic Time Series. New York: Springer.
Lecture Notes in Statistics, Vol. 154.
Klein, J. L. (1991). Statistical Visions in Time – A History of Time Series Analysis 1662–1938. Cambridge:
Cambridge University Press.
Kohn, R. and C. F. Ansley (1985). Efficient estimation and prediction in time series regression models.
Ladiray, D. and B. Quenneville (2001). Seasonal Adjustment with the X-11 Method. New York: Springer.
Lecture Notes in Statistics, Vol. 158.
264 BIBLIOGRAPHY
Lehman, E. L. (1986). Testing Statistical Hypothesis (Second ed.). New York, NY: J. Wiley and Sons.
Lin, J.-L. and T.-S. Liu (2002). Modeling lunar calendar holiday effects in Taiwan. Taiwan Economic Forecast
and Policy 33 (1), 1–37. http://www.census.gov/ts/papers/lunar.pdf.
Ljung, G. M. (1993). On outlier detection in time series. Journal of Royal Statistical Society B 55, 559–567.
Ljung, G. M. and G. E. P. Box (1978). On a measure of lack of fit in time series models. Biometrika 65,
297–304.
Ljung, G. M. and G. E. P. Box (1979). The likelihood function of stationary autoregressive-moving average
models. Biometrika 66, 265–270.
Lothian, J. (1984). The identification and treatment of moving seasonality in X-11. Proceedings of the Business
and Economic Statistics Section of the American Statistical Association, 166–171.
Lothian, J. and M. Morry (1978). A test of quality control statistics for the X-11-ARIMA seasonal adjustment
program. Research Paper, Seasonal Adjustment and Time Series Staff, Statistics Canada.
Lytras, D. P. (2012a). X-13-Graph: A SAS/GRAPH Program for X-13ARIMA-SEATS Output, User’s Guide
for the Batch Program on the PC or Unix, Version 2.0. U. S. Census Bureau, U. S. Department of
Commerce http://www.census.gov/srd/www/x13graph/x13gbdoc/x13gbat_doc.html.
Lytras, D. P. (2012b). X-13-Graph Java, Version 2.0. U. S. Census Bureau, U. S. Department of Commerce
http://www.census.gov/srd/www/x13graph/x13gjavadoc/x13graph_java_doc.html.
Lytras, D. P., R. M. Feldpausch, and W. R. Bell (2007). Determining seasonality: A comparison of diag-
nostics from X-12-ARIMA. Proceedings of the International Conference on Establishment Surveys III .
(CD-ROM).
Maravall, A. (1995). Unobserved components in economic time series. In M. H. Pesaran and M. Wickens
(Eds.), The Handbook of Applied Econometrics. Basil Blackwell.
McElroy, T. S. (2006). Statistical properties of model-based signal extraction diagnostic tests. Research Re-
port No. RRS2006/06, Statistical Research Division, U.S. Census Bureau http://www.census.gov/srd/
papers/pdf/rrs2006-06.pdf.
McElroy, T. S. (2008a). Matrix formulas for nonstationary signal extraction. Econometric Theory 24 (4), 1–22.
McElroy, T. S. (2008b). Statistical properties of model-based signal extraction diagnostic tests. Communica-
tions in Statistics: Theory and Methods 37 (4), 591–616.
McElroy, T. S. and R. Gagnon (2006). Finite sample revision variances for ARIMA model-based signal
extraction. Research Report No. RRS2006/05, Statistical Research Division, U.S. Census Bureau http:
//www.census.gov/srd/papers/pdf/rrs2006-05.pdf.
McElroy, T. S. and R. Gagnon (2008). Finite sample revision variances for ARIMA model-based signal
extraction. Journal of Official Statistics 24 (3), 451–467.
Monsell, B. C. (2002). An update on the development of the X-12-ARIMA seasonal adjustment program. In
Proceedings of the 3rd International Symposium on Frontiers of Time Series Modeling, pp. 1–11. Tokyo:
Institute of Statistical Mathematics.
Monsell, B. C. (2006). Recent developments in seasonal adjustment software at the Census Bureau. Pro-
ceedings of the Eurostat Conference on Seasonality, Seasonal Adjustment and Their Implications for
Short-term Analysis and Forecasting.
BIBLIOGRAPHY 265
Montes, M. J. (1997a). Frequency of the date of Easter 1875 to 2124. http://www.smart.net/~mmontes/

freq3.html [Online; accessed 8-December-1999].
Montes, M. J. (1997b). Frequency of the date of Easter over one 400 year Gregorian cycle. http://www.
smart.net/~mmontes/freq2.html [Online; accessed 8-December-1999].
Montes, M. J. (2001). Calculation of the ecclesiastical calendar. http://www.smart.net/~mmontes/ec-cal.
html [Online; accessed 31-July-2001].
Mood, A. M., F. A. Graybill, and D. C. Boes (1974). Introduction to the Theory of Statistics, 3rd Edition.
New York: McGraw-Hill.
More, J. J., B. S. Garbow, and K. E. Hillstrom (1980). User guide for MINPACK-1. Report ANL-80-74,
Argonne National Laboratory, Argonne, Illinois.
Otto, M. and W. R. Bell (1993). Detecting temporary changes in level in time series. Proceedings of the
American Statistical Association, Business and Economic Statistics Section, 170–174.
Otto, M. C. and W. R. Bell (1990). Two issues in time series outlier detection using indicator variables.
Proceedings of the American Statistical Association, Business and Economic Statistics Section, 182–187.
Otto, M. C., W. R. Bell, and J. P. Burman (1987). An iterative GLS approach to maximum likelihood
estimation of regression models with ARIMA errors. Proceedings of the American Statistical Association,
Business and Economic Statistics Section, 632–637.
Ozaki, T. (1977). On the order determination of ARIMA models. Applied Statistics 26, 290–301.
Pearson, E. S. (1938). The probability integral transformation for testing goodness of fit and combining
independent tests of significance. Biometrika 30, 134–148.
Pearson, E. S. and H. O. Hartley (1954). Biometrika Tables for Statisticians, Volume 1. Cambridge: Cam-
bridge University Press.
Pierce, D. (1971). Least squares estimation in the regression model with autoregressive-moving average errors.
Planas, C. and R. Depoutot (2002). Controlling revisions in ARIMA-model-based seasonal adjustment. Jour-
nal of Time Series Analysis 23, 193–214.
Priestley, M. (1981). Spectral Analysis and Time Series. London: Academic Press.
Quenneville, B., P. Cholette, G. Huot, K. Chiu, and T. D. Fonzo (2004). Adjustment of seasonally adjusted
series to annual totals. Research Report, Statistics Canada.
SAS Institute Inc. (1990). SAS/GRAPH Software: Reference, Version 6, First Edition, Volume 1. Cary, NC:
SAS Institute.
Schwarz, G. (1978). Estimating the dimension of a model. Annals of Statistics 6, 461–464.
Shiskin, J., A. H. Young, and J. C. Musgrave (1967). The X-11 variant of the Census Method II seasonal
adjustment program. Technical Paper No. 15, U.S. Department of Commerce, U. S. Census Bureau.
Snedecor, G. W. and W. G. Cochran (1980). Statistical Methods (7th ed.). Ames: The Iowa State University
Press.
Soukup, R. J. and D. F. Findley (1999). On the spectrum diagnostics used by X-12-ARIMA to indicate
the presence of trading day effects after modeling or adjustment. Proceedings of the American Statistical
Association, Business and Economic Statistics Section, 144–149. http://www.census.gov/ts/papers/
rr9903s.pdf.
266 BIBLIOGRAPHY
Taniguchi, M. and Y. Kakizawa (2000). Asymptotic Theory of Statistical Inference for Time Series. New
York: Springer-Verlag.
Thomson, P. and T. Ozaki (2002). Transformation and seasonal-trend decomposition. In Proceedings of the
3rd International Symposium on Frontiers of Time Series Modeling, pp. 197–212. Tokyo: Institute of
Statistical Mathematics.
Titova, N. and B. C. Monsell (2009). Detecting stock calendar effects in U. S. Census Bureau inventory
series. SRD Research Report RRS 2009-6, U. S. Census Bureau http://www.census.gov/srd/papers/
pdf/rrs2009-06.pdf.
Vandaele, W. (1983). Applied Time Series and Box-Jenkins Models. New York: Academic Press.
Wikipedia (2009). Stock and flow — Wikipedia, the free encyclopedia. http://en.wikipedia.org/wiki/
Stock_and_flow [Online; accessed 15-January-2009].
Wilson, G. T. (1983). The estimation for time series models, part I. Yet another algorithm for the exact
likelihood of ARMA models. Technical Report No. 2528, Mathematics Research Center, University of
Wisconsin-Madison.
Index
additive decomposition, 193, 199, 205, 213 irregular regression

sliding spans, 180, 183 critical value, 221
aggregate series, see composite series revisions history, 111–112
AIC, 37, 43, 60, 92, 148 revisions history, 111
AIC test, 44, 45, 136, 145–147 automdl spec
AIC difference, 44, 45, 136, 145–146, 193, 199 arguments, 63–65
irregular regression, 221, 231 acceptdefault, 63
AIC difference, 221, 231 checkmu, 63
select transformation, 195, 199 diff, 63
AICC, 3, 37, 43, 92, 136, 145, 199, 221, 231 fcstlim, 63
airline model, 199 ljungboxlimit, 64
Akaike, Hirtugu, 3, 4 maxdiff, 64
arima spec maxorder, 64
arguments, 59–60 mixed, 64
ar, 59 print, 64
ma, 59 rejectfcst, 64
model, 59 savelog, 64
title, 60 details, 66–72
details, 60 examples, 73
examples, 61–62 rarely used arguments, 65–66
usage, 59 usage, 63
ARIMA model, 1, 26
airline model, 62 backcasts, 1, 101, 136, 140, 194, 214, 221, 227, 231
cancellation of AR and MA factors, 42 modelspan argument, 172
convergence, 36, 40–41, 88, 90 BIC, 37, 43, 92
overdifferencing, 42–43
asymmetric, 1, 212 calendar effects, 1
autocovariance, 41 irregular regression, 2, 229
automatic model selection, 3 change-of-regime regressors, 144–145
automdl spec, 66–72 check spec
balanced model, 65, 70 arguments, 74
pickmdl spec, 6, 132–133 maxlag, 74
backcasts, 131, 133 print, 74
default model, 133 qtype, 74
forecasts, 133 save, 74
automatic outlier identification, 34, 38–39, 71, 127–128 savelog, 74
critical value, 38, 64, 66, 72, 124, 127, 128 details, 74–77
267
268 INDEX
examples, 77–78 maxiter, 87

usage, 74, 115 outofsample, 87
Chi-squared test, 136 print, 88
collinearity, 29 save, 88
composite spec savelog, 88
arguments, 79–80 tol, 88
appendbcst, 79 details, 90–92
appendfcst, 79 examples, 92–93
decimals, 79 rarely used arguments, 88–90
modelspan, 79 usage, 87
name, 80 extreme values
print, 80 exclude from irregular regression, 225
save, 80 X-11, 2, 18, 204, 210, 213–214
savelog, 80 calendar sigma, 211
title, 80
type, 80 flags, 11–15
details, 83–85 -c, 15, 84
examples, 85–86 -d, 9, 11
rarely used arguments, 80–83 -g, 14
usage, 79 -i, 11
composite series, 15, 58, 83 -m, 8, 11, 15, 84
concurrent seasonal adjustment, 110, 113 -n, 14, 15, 191
convergence tolerance, 41, 90 -p, 14, 15, 171
-q, 14
Dagum, Estela, 4 -r, 14
deterministic, 43, 91, 92 -s, 13–14, 84, 184
detrended series, see SI values -v, 15
differencing operators, 42, 48 -w, 14, 15
direct seasonal adjustment, 84 force spec
arguments, 94–96, 97
Easter lambda, 94
AIC test, 45, 136 mode, 94
irregular regression, 221, 231 print, 94
regARIMA model, 146 rho, 95
Bateman and Mayes procedure, 211, 214, 231 round, 95
regressor, 230 save, 94
flow, 31, 142 start, 95
Statistics Canada, 32, 142, 230 target, 96
stock, 31, 142 type, 96
removal of mean from regressor, 31, 140, 147–148, usefcst, 96
226, 232 details, 97–99
effective number of observations, 43, 47 examples, 99–100
error messages, 5–7, 60, 127 rarely used arguments, 96
estimate spec usage, 94
arguments, 87–88 forecast spec
exact, 87 arguments, 101
INDEX 269
exclude, 101 print, 120

lognormal, 101 save, 120
maxback, 101 sdiff, 120
maxlead, 101 details, 120–121
print, 101 examples, 122–123
probability, 101 usage, 120
save, 101 inadmissable decomposition, 72
details, 102–103 indirect seasonal adjustment, 84
examples, 103–105 revision history diagnostics, 85, 112
usage, 101 sliding spans diagnostics, 85
forecasts, 39, 136, 140, 194, 221, 227, 231 unadjusted series as component, 83
forecast extension, 1, 21, 103, 214 intervention effects, 34
prediction interval, 39 invertible, 37, 41–42
irregular component, 212–214, 220
Gómez, Victor, 2, 4, 18, 66, 157 irregular component regression, 211
graphics metafile, 14 irregular regression
AIC test, 231
handling spaces in file names, 10, 12–13 outlier critical value, 221
Hannan-Quinn criterion, 37, 43, 92 revisions history, 107, 110–112
Hannan-Rissanen estimation method, 65, 68 sliding spans, 179, 180
history spec Ishiguro, Makio, 4
arguments, 106–110
endtable, 106 Jacobian transformation adjustment, 47
estimates, 106
fixmdl, 106 kurtosis, 77
fixreg, 107
fstep, 107 leap year adjustment, 28, 143, 193, 198
print, 107 length-of-month adjustment, 28, 144, 193, 198, 199
sadjlags, 107 length-of-quarter adjustment, 144, 193, 199
save, 107 Ljung-Box Q-statistic, 37, 63, 64, 66, 68, 71, 74, 77
savelog, 109 log file, 11
start, 109 log-additive decomposition, 199, 205, 213
target, 109 sliding spans, 183
trendlags, 110
details, 112–113 Maravall, Agustı́n, 2, 4, 18, 66, 157
examples, 113–114 metadata spec
rarely used arguments, 110–112 arguments, 115
usage, 106 keys, 115
holiday effects, 33–34, 58, 149, 227, 232 values, 115
keep in seasonally adjusted series, 211 details, 116–117
remove from nonseasonal series, 148, 232 examples, 118–119
metafile, 7–11, 15, 79, 83, 84, 86
identify spec data metafile, 7, 9–11, 17, 167, 172
arguments, 120 input metafile, 8
diff, 120 Minimum AIC criterion (MAIC), 44, 45
maxlag, 120 missing value, 171, 173, 176
270 INDEX
model selection criteria, 43–48 file, 130

model span, 79–80, 112, 169, 179 identify, 130
moving average, X-11 method, 131
seasonal, 205, 211, 214 mode, 131
trend, 210 outofsample, 131
moving seasonality ratio (MSR), 205, 214 overdiff, 131
multiplicative decomposition, 199, 205, 213 print, 131
sliding spans, 183 qlim, 132
savelog, 132
nested models, 44, 45 details, 132–133
noninvertible model, 41 examples, 133–134
nonnested models, 45 usage, 130
preadjustments, 47
out-of-sample forecast error, 48, 104 prediction interval, 39, 102
outlier spec pseudo-additive decomposition, 193, 205, 212, 213
arguments, 124–126
critical, 124 regARIMA model, 1, 27
lsrun, 124 revisions history, 107
method, 125 sliding spans, 179
print, 125 regression spec
save, 125 arguments, 136–139
savelog, 125 aicdiff, 136
span, 125 aictest, 136
types, 126 chi2test, 136
details, 127–128 chi2testcv, 136
examples, 128–129 data, 136
rarely used arguments, 126–127 file, 137
usage, 124 format, 137
outlier regressor, 23, 34, 46, 107 print, 137
level shift (LS), 32, 34, 38, 142 save, 137
point outlier (AO), 32, 142, 230 savelog, 137
ramp, 33, 34, 143, 214 start, 139
seasonal, 32, 34 user, 139
seasonal outlier (SO), 143 usertype, 139
temporary change (TC), 32, 143 variables, 139
rate of decay, 32, 127, 140 details, 140–149
temporary level shift, 33, 143 examples, 149–156
use in revisions history, 113 rarely used arguments, 139–140
output file, 5–6, 8, 9, 22 usage, 135
alternate output filename, 7, 8, 10 regression variables
constant, 28, 30, 141
permanent prior adjustment, 198, 200 fixed seasonal, 29
pickmdl spec seasonal indicator, 30, 141
arguments, 130–132 trigonometric, 30, 141
bcstlim, 130 holiday, 107
fcstlim, 130 Easter, 33, 142, 230
INDEX 271
flow, 33–34 examples, 165–166

Labor Day, 32, 33, 142, 230 rarely used arguments, 160–161
Statistics Canada Easter, 32, 142, 230 usage, 157
stock, 34 series spec
Stock Easter, 142 arguments, 167–171
Thanksgiving, 32, 33, 142, 230 appendbcst, 167
leap year, 30, 141 appendfcst, 167
leap year regressor, 143 comptype, 168
length-of-month, 30, 141 compwt, 168
length-of-month regressor, 143 data, 168
length-of-quarter, 30, 141 decimals, 168
outlier, 107 file, 168
outliers, 34 format, 168
level shift (LS), 32, 34, 38, 142 modelspan, 169
point outlier (AO), 32, 142, 230 name, 169
ramp outlier, 33, 34, 143 period, 169
seasonal outlier, 32, 34 precision, 169
seasonal outlier (SO), 143 print, 170
temporary change (TC), 32, 143 save, 170
temporary level shift outlier, 33, 143 span, 170
trading day, 29, 107 start, 170
flow, 29, 30, 141, 229 title, 171
one coefficient, 30, 33, 141, 229 type, 171
one coefficient stock, 31, 142 details, 171–173
stock, 29, 31, 142, 230 examples, 173–177
rescaling, 144 rarely used arguments, 171
revision history usage, 167
AICC, 3 SI values, 214
forecast error, 3 singularity, 29, 121, 125, 144
revision history diagnostics, 85 skewness, 76
roots, 42, 60, 90–92 sliding spans
threshold, 178–179
saved table, 22 sliding spans diagnostics, 85, 183–185
seasonal component, 212, 213 slidingspans spec
seats spec arguments, 178–180
arguments, 157–159 cutchng, 178
appendfcst, 157 cutseas, 178
hpcycle, 157 cuttd, 178
out, 157 fixmdl, 179
print, 158 fixreg, 179
qmax, 159 length, 179
save, 158 numspans, 179
savelog, 159 outlier, 179
signifsc, 159 print, 180
statseas, 159 save, 180
details, 161–165 savelog, 180
272 INDEX
start, 180 type, 198

details, 183–185 details, 198–201
examples, 185–187 examples, 201–203
rarely used arguments, 180–182 rarely used arguments, 198
usage, 178 usage, 193
spectrum spec transformation, 28, 103, 143, 199, 213
arguments, 188 AIC test, 195, 199
print, 188 Box-Cox power transformation, 46, 47, 195
save, 188 inverse, 195
savelog, 188 log transform, 46
start, 188 logistic, 28, 47, 195
type, 188 model comparisons, 46–47
details, 191 square root, 195
examples, 191–192 transformation adjustment, 44
rarely used arguments, 188–191 trend
usage, 188 Henderson filter, 2, 210–212, 214
spectrum diagnostic, 19, 49–52, 188 level shift outliers, 214
periodogram, 188 trend-cycle component, 212, 214
start date, 188
visually significant, 191 user-defined regressors, 107, 147
AIC test, 136
temporary prior adjustment, 198, 200 irregular regression, 221
Trading Day seasonal, 147
AIC test type specification, 147, 231
regARIMA model, 146
trading day effects, 58, 149, 227, 232 x11 spec
AIC test, 136, 146, 231 arguments, 204–210
irregular regression, 221, 231 appendbcst, 204
remove from nonseasonal series, 148, 232 appendfcst, 204
transform spec final, 204
arguments, 193–198 mode, 205
adjust, 193 print, 205
aicdiff, 193 save, 205
data, 194 savelog, 205
file, 194 seasonalma, 205
format, 194 sigmalim, 205
function, 195 title, 210
mode, 195 trendma, 210
name, 196 type, 210
power, 196 details, 212–215
precision, 196 examples, 215–219
print, 196 rarely used arguments, 211–212
save, 196 usage, 204
savelog, 196 x11regression spec
start, 196 arguments, 221–226
title, 198 aicdiff, 221
INDEX 273
aictest, 221
critical, 221
data, 221
file, 221
format, 221
outliermethod, 222
outlierspan, 222
print, 222
prior, 223
save, 222
savelog, 223
sigma, 225
span, 225
start, 225
tdprior, 225
user, 225
usertype, 225
variables, 226
details, 229–232
examples, 232–234
rarely used arguments, 226–229
usage, 220

docX13AS PDF

Uploaded by

Copyright:

Available Formats

docX13AS PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

docX13AS PDF

Uploaded by

Copyright:

Available Formats

X-13ARIMA-SEATS Reference Manual

Time Series Research Staff

January 15, 2013

2.3 Input errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.4 Specifying an alternate output filename . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.5 Running X-13ARIMA-SEATS on more than one series . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.5.1 Running X-13ARIMA-SEATS in multi-spec mode . . . . . . . . . . . . . . . . . . . . . . . . 8

2.5.2 Running X-13ARIMA-SEATS in single spec mode . . . . . . . . . . . . . . . . . . . . . . . . 9

2.5.3 Special Case: File Names Containing Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.6 Log Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.8 Program limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3 The Specification File and Its Syntax 17

3.1 Examples of Input Specification Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.2 Print and save . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.4 General rules of input syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4 RegARIMA modeling Capabilities 26

5 Points Related to regARIMA Model Estimation 40

6 Points Related to Seasonal Adjustment and Modeling Diagnostics 49

7 Documentation for Individual Specs 56

7.7 FORECAST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

A Graphics Codes 235

B Print and Save Tables 239

C Irregular-component Regression Models Used 251

2.1 X-13ARIMA-SEATS Program Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.1 X-13ARIMA-SEATS Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.1 Predefined Regression Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

6.1 Revision Measure Calculated for Revision Lag Analysis . . . . . . . . . . . . . . . . . . . 55

7.1 Available Output Tables for Automdl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

7.14 Tables Saved As Percentages in the save Argument . . . . . . . . . . . . . . . . . . . . . 95

7.44 Transformations Available Using the function Argument . . . . . . . . . . . . . . . . . . 195

A.1 Codes Associated with the X-13ARIMA-SEATS Graphics Metafile . . . . . . . . . . . . . . . 235

2.3 Input errors

2.4 Specifying an alternate output filename

path\x13as path\filename path\outname (2.1)

2.5 Running X-13ARIMA-SEATS on more than one series

2.5.1 Running X-13ARIMA-SEATS in multi-spec mode

To run X-13ARIMA-SEATS with a input metafile, use the following syntax:

2.5.2 Running X-13ARIMA-SEATS in single spec mode

To run X-13ARIMA-SEATS with a data metafile, use the following syntax:

x13as specfile -d metafile

x13as default -d exports

2.5.3 Special Case: File Names Containing Spaces

"c:\export data\xuu1.dat" "c:\export output\xuu1 a"

2.6 Log Files

path\x13as arg1 arg2 . . . argN

Table 2.1: X-13ARIMA-SEATS Program Flags

flags description of flags

x13as test test2

x13as "c:\My Spec Files\test" "c:\My Output\test2"

the diagnostics summary file will be stored in test.udg, and for

x13as test -s -o testout

the diagnostics summary file will be stored in testout.udg.

(2) a graphics metafile containing the names of these files;

x13as test -g c:\sagraph

x13as test -g c:\sagraph -o testout

2.8 Program limits

Table 2.2: X-13ARIMA-SEATS Program Limits

Table 3.1: X-13ARIMA-SEATS Specifications

3.1 Examples of Input Specification Files