Dr.

Stefan Linz
is an economist and head of the
Short-term Economic Indices for
Industry, Methods Development

SEASONAL ADJUSTMENT OF for Short-term Statistics, Seasonal

Adjustment section of the Federal

SHORT-TERM STATISTICS USING

Statistical Office.

X-12-ARIMA AND X13 IN JDEMETRA+ Dr. Claudia Fries

is an economist and has been assis-
tant head of the above section since
2016. She coordinates both sea-
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker sonal adjustment and implementa-
tion of the seasonal adjustment
software JDemetra+ at the Federal
Statistical Office.

 Keywords: seasonal – adjustment – calendar adjustment – JDemetra+ –

X13 – X-12-ARIMA Julia Völker
is an economist and has been a
ABSTRACT member of the above section’s aca-
demic staff since January 2018. The
The results of economic time series – for instance monthly data on turnover in in- main focus of her work is on further
dustry or quarterly gross domestic product figures – are affected by annually recur- development projects in the area of
ring seasonal effects and calendar variations. This makes an assessment of current seasonal adjustment.

development more difficult. Such effects can be eliminated using statistical methods
of seasonal adjustment in order to permit a better assessment of current trends. This
article describes the use of the seasonal adjustment methods X-12-ARIMA and X13
in JDemetra+ for purposes of seasonal adjustment of the short-term statistics of the
Federal Statistical Office.

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 1
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

1 In addition to values that are both calendar and season-

ally adjusted, purely calendar adjusted and also unad-
justed values are generally available for all statistics for
Why seasonal adjustment?  | 1 which adjusted results are published.

Short-term statistics provide monthly or quarterly data

on the development of various economic sectors and the  Excursus: year-on-year comparison
economy as a whole. The variables include, for instance,
turnover, the number of persons employed and the vol- Year-on-year comparisons are a simple method
ume of output in the different economic branches, and to eliminate the usual seasonal fluctuations from
the results of the quarterly domestic product calcula- observation. In this context, the unadjusted value
tions for Germany. Regarding these statistics, the focus or the purely calendar adjusted value of the current
of interest is more often on the development of results period is compared with the relevant value of the
over time than on the absolute values of individual same quarter or month of the previous year.
statistical results. For this reason, the relevant publica-
Figure 1
tions primarily present growth rates on a comparative
Year-on-year and month-on-month comparison
period. However, changes are often strongly affected by
seasonal effects which make an assessment of current
Value of the previous month
trends difficult.

The aim of seasonal adjustment is to eliminate sea-

sonal variations which occur with similar intensity every Current
value
year from the time series values. An example is the
effects of typical seasonal fluctuations in weather or
average decreases recorded in holiday months. | 2 Sea-
sonal adjustment allows it to be assessed, for instance,
whether a month-on-month increase in construction Value of the previous year
2019 - 01 - 0628
output recorded in March is to be regarded as extraordi-
narily strong or as being in line with the typical seasonal
trend. However, the results of year-on-year comparisons
may point in a different direction to that of month-
Seasonal adjustment often includes calendar adjust- on-month or quarter-on-quarter comparisons based
ment. | 3 This is aimed at adjusting effects which arise on seasonally adjusted results. This is for instance
from different numbers of working or trading days or cer- the case when a longer upswing is nearly over and
tain days of the week in a specific quarter or month due the month-on-month comparison already reveals
to the position of weekends or statutory holidays. a downward trend, while the current result is still
above the value of the previous year. As a rule, com-
parisons with the previous period using seasonally
  1 We would like to thank Jonas Flechsig (mathematics student) for his adjusted results allow economic analysis that is
research assistance for this article.
  2 The effects of bridging days (days lying between a public holiday
based on a comparison with a period closer to the
and a weekend) or movable school holidays are not part of adjust- one under review and is therefore more relevant than
ment. Similarly, no adjustment is made for “extraordinary” weather year-on-year comparisons.
effects, that is weather effects other than those that can typically be
expected during the course of a year. As agreed in the European Sta-
tistical System, the effects of extraordinary concentrations of leave It is common at the international level to publish sea-
or weather conditions that are atypical of a season can and should sonally adjusted results. This facilitates economic analy-
be reflected in the seasonally adjusted results as they do not occur
with similar intensity at the same time each year (Eurostat, 2015,
sis based on month-on-month or quarter-on-quarter
here: page 20 f.). comparisons - especially if the relevant analysis covers
  3 In this article, the concept of seasonal adjustment generally covers many economic branches and the seasonal fluctuations
the adjustment for both seasonal fluctuations and calendar effects.
Where seasonal adjustment explicitly excludes calendar adjustment,
common in the individual branches cannot be presumed
it is referred to as “seasonal adjustment in the narrow sense”. to be known. However, seasonal adjustment has the

2 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

disadvantage that mathematical-statistical methods of In the adjustment process, only the calendar and sea-
analysis are used which require various assumptions to sonal components are excluded from the time series so
be made and parameters to be set. Therefore the results that both the trend-cycle component and the irregular
obtained by means of different adjustment procedures component are still included in the calendar and sea-
and by various individuals engaged in seasonal adjust- sonally adjusted result. The following equation shows
ment may differ. To increase the objectivity of seasonally this relationship:
adjusted results and enhance their comparability within
Xt
the European Statistical System, harmonised seasonal (2) X KSB
t = = Tt · It
St · Kt
adjustment methods are applied and both nationally
and internationally coordinated standards of seasonal

adjustment are observed (see Chapter 3).

The underlying time series model

Seasonal adjustment is based on the assumption that
a time series can be decomposed into several compo-
nents. The trend-cycle component reflects long-term ten-
dencies and cyclical movements. The seasonal compo-
nent covers movements which recur with similar intensity
each year. The calendar component includes the aver-
age impact of calendar constellations, for instance due
to the varying number of working days between months
or quarters of the same name. The irregular component
comprises effects that are incidental or can be explained
economically, but do not belong to the other compo-
nents. Examples are the effects of an exceptionally warm
winter on construction activity or the impact of strikes on
the turnover of a given branch. In applying X-12-ARIMA
and X13 in JDemetra+, the bodies of official statistics
mostly use a multiplicative model for the decomposition
of time series into individual components: | 4

(1) Xt = St . Kt . Tt . It

Xt (unadjusted) value

St seasonal component

Kt calendar component

Tt trend-cycle component

It irregular component

t time

4 In some cases, however, additive models or combinations of additive

and multiplicative components are used. The following explanations
always refer to the multiplicative model.

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 3
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

“The downside of seasonal adjustment is that seasonal- regulation (Framework Regulation Integrating Busi-
ity cannot be precisely defined and different approaches ness Statistics). | 8
– such as the signal extraction approach […] and the
> Regarding the labour cost index, the compulsory
semi-parametric approach based on a set of predefined
transmission of seasonally and working-day adjusted
moving averages […] – may result in different outcomes.
results is governed and substantiated in an imple-
The expertise of an analyst will also impact on the qual-
menting regulation: “Seasonal and working-day
ity of seasonal adjustment, although the primary drivers
adjustment of the labour cost index is an essential part
are the quality of the unadjusted time series and the pro-
of the compilation of the index. Adjusted series make it
duction timetable.” (Eurostat, 2015, here: page 6)
possible to compare results and to interpret the index
The guidelines on seasonal adjustment are intended to in a comprehensible manner.” | 9
enhance comparability between statistics in the ESS,
which is to be achieved by adherence to given standards
of seasonal adjustment and the application of certain 4
seasonal adjustment methods. Individual stages of the
seasonal adjustment process are detailed and different Mathematical-statistical methods
options are described for each case. They are classified
as Alternative A (best alternative), B (acceptable) and C
and software
(to be avoided). In addition to the guidelines, EU regu-
lations regarding the relevant statistics specify further
4.1 Mathematical-statistical methods
European requirements for seasonal adjustment. Some
examples are given below:
Regarding the mathematical-statistical methods of sea-
> In national accounting, the requirement to provide sonal adjustment, three “families of methods” are rel-
calendar and seasonally adjusted quarterly data is evant in the context of German and European official
set out in the Regulation on the European System of statistics.  Chart 1
Accounts (ESA 2010). The following notes are pro-
Chart 1
vided, for example, concerning the data of the table of
Families of seasonal adjustment methods used in official
main aggregates to be supplied: “Quarterly data are to statistics in Germany and the EU
be provided in non-seasonally adjusted form, as well (1) Census X11, X-12-ARIMA, X13 in JDemetra+
as in seasonally adjusted form (including calendar
(2) SEATS, TRAMO/SEATS, Tramoseats in JDemetra+
adjustments, where relevant).” | 6 The ESA 2010 regu-
(3) Berliner Verfahren, BV4.1
lation also specifies how seasonal adjustment is to be
made.
Census X11, X-12-ARIMA, X13 in JDemetra+
> In the area of business statistics, the Regulation con-
cerning short-term statistics, as it is called, plays a The family shown at the top of Chart 1 became popular
central role. In accordance with this regulation, data through a software program developed by the United
for the variables of “production”, “turnover” and partly States Census Bureau in the 1960s (Census X-11). The
“hours worked” in industry, construction, retail trade underlying mathematical-statistical method is based
and repair as well as other services are to be transmit- on weighted moving averages (so-called filters) that are
ted in a working-day adjusted form. | 7 The obligation iteratively applied to determine the trend-cycle and the
to provide seasonally adjusted data is to be extended seasonal component of a time series. X-12-ARIMA is the
in the context of the newly created FRIBS framework extended method where both a pre-treatment and fore-
cast of the time series using RegARIMA models are made
before trend and seasonal filters are applied. These are
  6 Regulation (EU) No 549/2013 of the European Parliament and of
the Council of 21 May 2013 on the European system of national and   8 See Waldmüller/Weisbrod (2015) for FRIBS.
regional accounts in the European Union, Annex B, Table 1.   9 Commission Regulation (EC) No 1216/2003 of 7 July 2003 imple-
  7 Council Regulation (EC) No 1165/98 of 19 May 1998 concerning menting Regulation (EC) No 450/2003 of the European Parliament
short-term statistics, consolidated version. and of the Council concerning the labour cost index.

4 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

regression approaches that are combined with time in JDemetra+) are based on procedures that are classi-
series models (so-called ARIMA models) in order to fied as (A) alternative in the guidelines. Besides, the lat-
estimate calendar effects, to identify, model and esti- ter classify the use of seasonal adjustment procedures
mate outliers and to reduce methodological problems not mentioned under (A) or (B) - including the Berliner
in applying the filters. | 10 X-12-ARIMA is also available Verfahren- as a (C) course of action. However, the rel-
in the software JDemetra+ where it is referred to as X13 evant classification is only based on the criterion of
(see Chapter 4.2). international comparability rather than the quality of the
methods.
SEATS, TRAMO/SEATS, Tramoseats in JDemetra+ For reasons of international comparability, the Census
The mathematical-statistical method SEATS, promoted family methods were introduced at the Federal Statisti-
by the Bank of Spain, is based on a signal extrac- cal Office in the late 1990s in addition to the Berliner
tion technique for modelling the trend-cycle and sea- Verfahren that had exclusively been used until that
sonal components. As a result of further development, time. | 13 Today the seasonally adjusted results based on
TRAMO/SEATS appeared in 2001. This includes time X-12-ARIMA or X13 in JDemetra+ are typically shown at
series pre-treatment and forecasting using the TRAMO the top of the relevant publications of the Federal Statis-
algorithm (which is also based on RegARIMA models) tical Office.
and seasonal adjustment in the narrow sense using
SEATS. | 11 The method is also available in the JDemetra+
software where it is referred to as Tramoseats.
4.2 Software

So far the Federal Statistical Office has carried out sea-

Berliner Verfahren (Berlin procedure), BV4.1 sonal adjustment according to X-12-ARIMA on the basis
The Berliner Verfahren is designed for modelling time of a software with the same name (version 0.2.8). This
series components through moving linear regressions. software was developed by the United States Census
The outliers of a time series are identified beforehand Bureau. Since the end of 2017, the Federal Statistical
using probability models. The mathematical bases of Office has gradually changed over to the program JDeme-
this method were developed together by Technische tra+ (at present version 2.2) which was developed within
Universität Berlin and the German Institute for Eco- the European Statistical System and the System of Euro-
nomic Research in the late 1960s. The method has been pean Central Banks. Eurostat generally recommends the
used at the Federal Statistical Office since 1972. From use of JDemetra+ for the seasonal adjustment of official
1983 onwards, the Office applied the fourth version of statistics in the European Union. | 14 The platform-inde-
the Berlin Verfahren (BV4), which was the result of fur- pendent open source software is available, for instance,
ther development by the Federal Statistical Office. That as a client application; it permits seasonal adjustment
version was replaced by BV4.1 in 2004. BV4.1 differs using X-12-ARIMA or TRAMO/SEATS. | 15 As currently the
from BV4 to the extent that it comprises methodologi- use of JDemetra+ is largely based on the same elements
cal improvements regarding the treatment of calendar as X-12-ARIMA at the Federal Statistical Office, the
effects, outliers and user-defined components. | 12 changeover to X13 in JDemetra+ (abbreviated: X13 JD+)
does not involve fundamental methodological changes.
The mentioned guidelines of the European Statistical
System also contain recommendations as to the sea- The Federal Statistical Office uses the software BV4.1
sonal adjustment method to be used. The methods for seasonal adjustment according to the Berliner Ver-
X-12-ARIMA and TRAMO/SEATS (or X13 and Tramoseats fahren. A client solution is available to external users for
free download (www.destatis.de).
10 The X-12-ARIMA method is described in more detail in the fifth chap-
ter of this article.
11 See Gómez/Maravall (2001) for the TRAMO/SEATS method. The acro- 13 See Hauf (2002) for a detailed description of the respective back-
nym TRAMO stands for “Time Series Regression with ARIMA Noise, ground.
Missing Observations and Outliers” and SEATS for “Signal Extraction 14 See Eurostat (2018) for methodological descriptions and download
in ARIMA Time Series”. options.
12 See Speth (2004) for the Berlin method. 15 As mentioned above, X-12-ARIMA is referred to as X13 in JDemetra+.

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 5
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

For the sake of completeness, it should be mentioned The term ARIMA refers to a set of modelling tools which
that, in recent years, the United States Census Bureau is based on visible interdependencies between the val-
has provided software called X-13-ARIMA-SEATS as a cli- ues of different periods of the same time series (see
ent solution, too, which enables seasonal adjustment excursus on ARIMA models). ARIMA modelling permits
based on X-12-ARIMA and TRAMO/SEATS. | 16 the interrelations identified in this time series – based
on the data available - to be applied to the future. So
part of the future development can be anticipated,
5 which facilitates forecasting in order to extend the time
series.
Setting parameters for seasonal
 Excursus: ARIMA-models
adjustment
The term ARIMA (Autoregressive Integrated Moving
Average) originates from time series analysis. | 18
5.1 Why to set parameters? ARIMA models contain three different modelling
options: differencing, modelling of autoregressive
The X-12-ARIMA and X13 JD+ procedures provide a wide processes (AR) and of so-called moving average
range of options to consider both the specific character- processes (MA). These three options are applied to
istics of a time series and the time series development the respective preceding periods in non-seasonal
over certain periods in determining the components modelling or to the relevant previous-year periods in
of the series (seasonality, trend-cycle, etc.). | 17 To use seasonal modelling. On the whole, six specification
these options, various specification parameters relating parameters are used for the presentation of ARIMA
to the RegARIMA model to be applied, the trend and sea- processes; these are shown as follows:
sonal filters and other features have to be determined.
The specification of parameters required during the vari- (3) Outline of the ARIMA model:
ous calculation steps will be briefly explained below. As ARIMA(p,d,q)(P,D,Q)S
the names X-12-ARIMA and X13 JD+ refer to the same (p,d,q) specification parameters of the
procedure, only the first designation will be used for rea- non-seasonal part of the ARIMA model
sons of simplification.
(P,D,Q)S specification parameters of the
seasonal part of the ARIMA model
5.2 First stage of the method:
First the parameters of non-seasonal modelling
pre-treatment and forecasting which are indicated in small letters in the first
brackets are explained. Parameter d relates to dif-
The first stage of the method comprises a pre-treatment ferencing. In the case of non-seasonal first-order
of the time series using RegARIMA models, during which differencing, further calculations are based on the
calendar and outlier effects are identified, modelled and period-to-period difference of the unadjusted values
estimated. In addition, the time series of unadjusted rather than the unadjusted values themselves.
values is extended at both ends by means of forecasts.
Extending the time series is required, for instance, for (4) Non-seasonal first-order difference:
the second stage of the procedure. During that stage,
weighted moving averages are calculated which cannot
Δ1Xt first-order difference of the unadjusted
be determined as such at the ends of the time series.
value at time t compared with the pre-
16 See the website of the United States Census Bureau for methodologi- vious period
cal descriptions and download options. [Accessed on 30 May 2018].
Available at www.census.gov/srd/www/x13as Xt unadjusted value at time t
17 See Findley et al. (1998) for the X-12-ARIMA method. A detailed
description with reference to its application at Deutsche Bundesbank 18 See Box/Jenkins (1970) for ARIMA models in time series analysis.
is given by Kirchner (1999). Gericke/Seidel (2014) provide an over- An easy-to-read and application-oriented introduction is given by
view of X-12-ARIMA. Nazmen (1988).

6 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

Differencing is required if the time series includes (6) Moving average process:
a trend so that AR and MA processes could not be

applied. AR and MA processes are based on the
assumption that current values only depend on
preceding values and on random variables and do
not follow a general trend. In most time series of
economic statistics, and especially with multiplica-
tive time series models, the unadjusted values are
subjected to logarithmic transformation before dif-
ferencing. Parameter d for period-to-period differ-
encing is used to determine how many reiterations of
differencing should be made. Usually, parameter d
takes value 0 or 1, and seldom 2. | 19 No differencing
is applied for value 0, which means that calculation
continues directly with the unadjusted values.

Parameter p relates to autoregressive processes.

The assumption is that the current values of a time
series depend on the previous values of the same
time series:

(5) Autoregressive process:

Xt = a1 Xt–1 + a2 Xt–2 + … + ap Xt–p + εt

Δ2Xt = Δ1Xt – Δ1Xt–1. ∆1s (∆1 Xt ) = ∆1 Xt – ∆1 Xt–12 = (Xt – Xt–1) – (Xt–12 – Xt–13 )

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 7
 Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

ysis. The regressors are intended to reflect calendar and is subtracted from the number of working days in May
also outlier effects. Equation (7) refers to a multiplica- 2018. The effect of this approach which is called calen-
tive time series model with a logarithmic transformation dar centring is that the overall level of the time series is
of the unadjusted values, where Zt represents the ARIMA not shifted by calendar adjustment. | 27 Another desired
part: | 23 advantage of calendar centring is, for instance, that the
annually recurring low number of working days in Febru-
(7) RegARIMA-model: ary has no calendar effect, but is covered by the later
calculation of seasonal components. In certain cases,
ln (Xt ) = � αl · klt + � m ⋅ LSmt + � n ⋅ AOnt + � v ⋅ TCvt + Zt
l m n vseveral working day regressors can be used for different
periods of the year, for instance, a regressor for January
) = � αl · klt + � m ⋅ LSmt + � n ⋅ AOnt + � v ⋅ TCvt + Zt
l                 m n v to November and another for December.

klt calendar regressors The outlier regressors LSmt, AOnt, TCvt can be used to
explicitly model extraordinary time series develop-
LSmt level shift
ments such as the impact of strikes. There are three dif-
AOnt additive outlier ferent outlier models. The level shift regressor (abbre-
TCvt temporary change viated: LS) is used to model a remaining structural
break. The additive outlier regressor (AO) represents
a, β, g, l estimation coefficients the shift of an individual data point of the time series
l, m, n, v running indices for regressors that returns to the previous level as early as during
of a regressor type the next period. The temporary change regressor (TC)
is designed to model a development where a sudden
Zt ARIMA modelling change decreases again gradually in the following data
When applying X-12-ARIMA in the short-term statistics points. Not all types of outliers are always applied in
of industry, the calculation of calendar regressors klt is RegARIMA models.
based on the number of working days which can vary
RegARIMA modelling has two functions in X-12-ARIMA.
from month to month. | 24 Typically, working days are all
First, the calendar factor Kt required in equation (2) can
days from Monday to Friday. To calculate a working day
be calculated from the estimation coefficients of the
regressor, first the number of working days in the current
calendar regressors. This factor is applied in calculating
month is determined after deducting statutory or quasi
the calendar and seasonally adjusted result. Here the
statutory holidays. | 25 Public holidays that are statutory
ARIMA part serves to measure the effects of the calen-
only in some Länder are weighted using the number of
dar regressors more precisely; a regression across time
employees subject to social insurance contributions in
series values without the ARIMA part could possibly
the industry of those Länder as a proportion of the total
provide biased results. At this point, the function of the
number in Germany as a whole. | 26 Afterwards the dif-
outlier regressors is that of control variables, which also
ference is calculated between the number of working
improve estimation. To calculate the calendar factor, the
days of the relevant month and the month-specific long-
estimation coefficients are multiplied by the regressor
term average. For instance, the average number of work-
values, and afterwards the logarithmic transformation is
ing days in all months of May between 1991 and 2030
reversed:

23 See for instance Bee Dagum/Bianconcini (2016) for the RegARIMA (8) Kt = exp ( � αl ·klt )
model. l

24 See Deutsche Bundesbank (2012) for a detailed description of calen- In the multiplicative model with previous logarith-
dar adjustment using X-12-ARIMA in the German official statistics.
mic transformation the estimation coefficients can be
25 Although quasi statutory holidays are not prescribed by law, many
employers offer them to their employees on a voluntary basis. For interpreted as semi-elasticities. They indicate, for the
instance, carnival is a quasi statutory holiday in some Länder.
26 For reasons of data availability, the number of employees subject to 27 Deutsche Bundesbank provides a plug-in called “TransReg” (down-
social insurance contributions in industry is sometimes also used for load at https://github.com/bbkrd/TransReg/releases) for calendar
weighting purposes in other than industry statistics. centring when applying the JDemetra+ software.

8 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

Table 1 calendar factors are therefore estimated on the basis

Semi-elasticity of the calendar regressors in of monthly time series which are closely related to the
the production index for intermediate goods quarterly national accounts aggregate to be analysed.
Estimation coefficients When adjusting gross value added in manufacturing, for
% instance, the monthly production index for manufactur-
January to November 3.2 ing is used to estimate the calendar factors. The monthly
December 2.1 calendar factors are then aggregated to obtain quar-
terly values which are used for the calendar adjustment
monthly data, the average percentage increase in the in national accounting (Deutsche Bundesbank, 2012,
unadjusted values if a given month has one working day here: page 55).
more than the long-term average.  Table 1 shows the
estimation coefficients for the calendar regressors in the A second function of RegARIMA modelling in X-12-ARIMA
production index for intermediate goods. | 28 includes the “pre-treatment” of unadjusted values and
extensions at the ends. Pre-treatment is designed to
The data available show that an additional working day temporarily remove both calendar and outlier effects
in the months from January to November will lead to an from the data; the pre-treated and extended series is
average 3.2 % increase in the production of intermediate then used as the basis for performing the second stage
goods in the German industry. An additional working day of the procedure (see Chapter 5.4).  Figure 2 shows
in December will only yield a 2.1 % increase. In December the month-on-month changes in the production index
it is more likely that additional working days are offset by for intermediate goods in March with and without pre-
employees taking leave. If, for instance, the Christmas treatment. To distinguish the pre-treated values from the
holidays fall on a weekend in years with an “employer- unadjusted values Xt , the former are written in small let-
friendly” calendar, there is a trend towards taking more ters (xt).
time off around the public holidays although, arithmeti-
cally, more working days are available. While the month-on-month change rates of the unad-
justed values of the production for intermediate goods
The big challenge in adjusting quarterly data is that the show clear variations over the years, the values pre-
different calendar effects of the months in a quarter may treated for calendar and outlier effects increase rather
superimpose each other. In the national accounts, the regularly by roughly 10 % in March. The thus more clearly
emerging seasonal pattern can be eliminated during the
28 See Bald-Herbel (2013) for the production index. second stage of the procedure.

Figure 2
Month-on-month change rates of the production index for industrial intermediate goods, March
Percent

Unadjusted values (Xt) Pre-treated values (xt)

0.20 0.20

0.15 0.15

0.10 0.10

0.05 0.05

0 0
2010 11 12 13 14 15 16 17 2018 2010 11 12 13 14 15 16 17 2018

2019 - 01 - 0629

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 9
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

It should be noted that, due to pre-treatment, outliers Determining the trend-cycle

are not incorporated in the seasonal component calcula-
tion. This in turn means that outlier effects are fully vis- The second stage of X-12-ARIMA begins with the estima-
ible in the calendar and seasonally adjusted time series. tion of a trend-cycle component to be used as the basis
The final calculation of the calendar and seasonally for measuring the seasonal variation. | 31 The trend filters
adjusted result according to equation (2) is performed applied here are weighted moving averages:
using again values which have not been pre-treated. t+a
(9) Tt = � θi ∙xi
i=t–a

5.3 Second stage of the method: applica-

tion of trend and seasonal filters

The above calculation of calendar factors and pre-

treatment of unadjusted data are followed by the sec-
ond stage of the X-12-ARIMA procedure. This stage
is designed for seasonal adjustment in the narrow
sense. | 29 As mentioned earlier, this refers to smoothing
the pre-treated time series using trend and seasonal fil-
ters which are calculated by means of weighted moving
averages. The trend and seasonal filters of the second
stage of the X-12-ARIMA procedure are applied step by
step using various iterations and partial iterations. | 30

29 A detailed description of the second stage of X-12-ARIMA using an

example is given by Ladiray/Quenneville (2001).
30 A reason for the iterative approach is that, although extreme values
are smoothed through trend and seasonal filters, the filtered results
remain, to a certain extent, affected by extreme values. For this
reason, the extreme values are replaced by substitute values during
the various iteration steps. However, extreme values can eventually
be identified and replaced only when the final filtered results are
available. Regarding the role of extreme values, it should be noted
that these are only excluded from the seasonal component and are
consequently included in the irregular component of the time series.
This means that they are not eliminated from the seasonally adjusted
result which also contains the irregular component.

Figure 3
Weights (θ) of the moving average for calculating the provisional trend-cycle in the first
iteration

1/12

1/24

0
t–6 t–5 t–4 t–3 t–2 t–1 t t+1 t+2 t+3 t+4 t+5 t+6

2019 - 01 - 0630

10 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

Figure 4
Unadjusted values and trend-cycle component in the production index for industrial intermediate goods

120

Unadjusted values Trend-cycle component

110

100

90

80

70

60
2005 06 07 08 09 10 11 12 13 14 15 16 17 18

2019 - 01 - 0631

engaged in seasonal adjustment may adjust the support The seasonal pattern of the production index for inter-
span of the Henderson filter as a specification parameter. mediate goods (e.g. base chemicals, metal products)
becomes already apparent in Figure 5. Production tends
 Figure 4 shows the unadjusted values and the trend- to increase during the year. In winter the conditions are
cycle component eventually determined after comple- less favourable for production and related transports,
tion of the iterations, using the production index for
and there is less demand for intermediate goods than
intermediate goods as an example.
during the rest of the year. Clear deviations from this
tendency are observed in March (upwards) as well as in
Determining the seasonal-irregular ratio (SI ratio) August and December (both downwards). The produc-
tion peak in March can be expected to be due to the
As a next step, the differences are determined between
regularly high number of calendar days in that month
the pre-treated values and the trend-cycle values
obtained. These are referred to as the seasonal-irregular which, because of calendar centring, is reflected in the
ratios (SI ratio). Regarding the multiplicative time series seasonal figure rather than the calendar factor. Although
model, the SI ratios are calculated as follows: July and August also have 31 calendar days and nearly
no statutory holidays, the fact that there is a regular
xt extraordinary concentration of leave during the summer
(10) Srt =
Tt holidays has a noticeable effect especially in August. As
mentioned above, December is characterised by a very

low production level, which is attributable to the public
holidays and the extraordinary concentration of leave in
this month.

Determining the final seasonal component

Figure 5 also shows the variance of the seasonal compo-
nent from year to year. As a next step, smoothed values
which are representative of the typical seasonal varia-

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 11
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

Figure 5
SI Ratio in the production index for industrial intermediate goods, 2010 to 2017

1,1

1,0

2010 2016 2017

0,9

0,8
Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sept. Oct. Nov. Dec.
2010 to 2017

The data points are sorted into months and into years. The points for a month show the raw seasonal factors of that month in all eight years of the presentation period.
2019 - 01 - 0632

tion of a month have to be generated from the fluctuat- Final seasonal components are determined as month-
ing SI ratios. Neither the causes of seasonal fluctuations specific weighted moving averages by applying seasonal
nor their effects on the time series can be fully observed. filters to the SI ratios of one month each throughout the
Since, however, the seasonal effects are assumed to years:
affect the time series to a similar extent every year, a j+b
ഥ
certain relatively stable level of the month-specific sea- m
(12) Sj = ෍ i ·Sr,m
i
i=j–b
sonal variation can be presumed. The final seasonal –
components are intended to adequately reflect the sup- Sm j value of the month-specific final seasonal
posed trend of the seasonal variations that cannot be component in year j
directly observed, on the one hand, by being as close m – specific month for which the weighted
as possible to the SI ratios and, on the other, by main- moving average is determined
taining a relatively stable level over the years. In this
context, the month-specific seasonal variation is to be j year
allowed to change gradually from year to year. To calcu- δ weight
late a month-specific final seasonal component, moving
b half range of the support span (not includ-
averages with specific weights, also called seasonal fil-
ing year j itself)
ters, are therefore applied again. In the equation, period
t for the SI ratios is expressed in months m and years j: i running index over the support span

r,m Applying the X-12-ARIMA seasonal adjustment method,

(11) Sj : = Srt
the person engaged in seasonal adjustment selects
the support span and the type of seasonal filter with a
view to covering the typical seasonal amplitude in the
best-possible way while, at the same time, taking into
account gradual changes in the seasonal variation.

12 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

Figure 6
SI ratios and final seasonal components for the production index for industrial intermediate goods, 2010 to 2017

1,1

1,0

0,9

0,8
Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sept. Oct. Nov. Dec.
2010 to 2017

The data points are sorted into months and into years. The points for a month show the seasonal-irregular component of that month in all eight years of the presentation period.
The final seasonal factors are shown as red lines.
2019 - 01 - 0633

Past experience has shown that, for a large number of  Excursus: seasonal filters
time series, the so-called 3x9 filter is very well suited to
reflect a gradually changing seasonal pattern. It covers To illustrate the names of the seasonal filters, these
a support span of eleven years and is regarded as the can be shown as overlapping simple, composite
standard filter in adjusting economic time series. If the moving averages.  Figure 7 provides an exemplary
seasonal patterns change more quickly, seasonal filters calculation of the value of a 3x5 filter for the month
with shorter support spans can be used. An example is of January 2013 where the filter is shown in the form
the 3x5 filter which covers no more than seven periods of three overlapping unweighted moving averages
(see the following excursus on seasonal filters for the across five periods each. The entire support span of
relevant names). If the seasonal patterns of individual the filter covers the January values of the years 2010
months change at different rates, different seasonal to 2016. Since, however, the January values located
filters can be chosen for the various months of a time towards the middle are included in the calculation
series.  Figure 6 shows the final seasonal components several times, their implicit weight is higher than
in the form of lines. Here the final seasonal components that of the values at the ends. The chart below the
were obtained by using 3x9 seasonal filters. computation formula shows the weights of the 3x5
seasonal filter resulting from the overlapping of the
unweighted moving averages.

As can be seen, the three values in the middle are

included in the calculation with a weight of 20 %
each. The weights decline towards the two ends, and
the weights of the values at the ends are just under
7 % each.

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 13
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

Figure 7
3x5 seasonal filter shown as overlapping composite moving averages

11 11 , Jan , Jan , Jan , Jan

S r2010 S r2011 S r2012 S r2013
Jan , Jan
SS Jan
2013
= · [
2013 =
33 · [ 55
( + + + + S r2014 )

1 1
S Jan
2013 = · [
. ( , Jan
S r2011 + , Jan
S r2012 + , Jan
S r2013 + , Jan
S r2014 +
, Jan
S r2015 )
3 5
1 1
S Jan
2013 = · [
. ( , Jan
S r2012 + , Jan
S r2013 + , Jan
S r2014 +
, Jan
S r2015 +
, Jan
S r2016 )]
3 5

Implicit weights (δ i ) of the 3x5 seasonal filter

1/5
20.0% 20.0% 20.0%

1/10 13.3% 13.3%

6.7% 6.7%
0
2010 2011 2012 2013 2014 2015 2016

2019- 01 - 0634

Calculation of seasonally adjusted results Xt

KSB
(13) Xt = , mit St : = Sm
j
To calculate the calendar and seasonally adjusted time St · Kt
series using the multiplicative model option, the unad- XKtSB calendar und seasonally adjusted time
justed values Xt that were not pre-treated are finally series
divided by both the final seasonal components and the Gruppierte
K calendar factor of formula (8)
t
calendar factors. The final seasonal components are
again sorted chronologically. Grafik S
für final seasonal components (sorted
t
chronologically)
Umbruch
Figure 8
Unadjusted values and calendar and seasonally adjusted results of the production index for industrial intermediate goods

120

Unadjusted values Calendar and seasonally adjusted results

110

100

90

80

70

60
2005 06 07 08 09 10 11 12 13 14 15 16 17 18

2019 - 01 - 0635

14 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

 Figure 8 shows the calendar and seasonally adjusted features of the seasonality. | 33 Here various quality indi-
time series together with the unadjusted values of cators, tests for residual seasonality and distribution
the production index for intermediate goods for the characteristics of the model residuals can be consid-
period from January 2005 to March 2018. The season- ered as being helpful. Nonetheless, these can serve as
ally adjusted time series shows less variations than the a rough orientation only. | 34 The complete specifications
unadjusted values, which is the reason why the trend- used at the Federal Statistical Office can be provided
cycle component can be better identified. Nevertheless, upon request. Thus data users can reproduce the sea-
it shows considerable fluctuations due to the irregular sonally adjusted results using the published unadjusted
component included. data.  Chart 2 shows central specification parameters
for three examples of statistics.

5.4 Specification as the total of

parameter settings

The previous chapters have shown that seasonal adjust-

ment using X-12-ARIMA requires various parameters to
be set when modelling the RegARIMA regression and
applying trend and seasonal filters. The total of parame-
ter settings is referred to as the specification. The speci- 33 The basic problem in determining the specification is that the time
series analysis is based on a decomposition of unadjusted values
fication is determined individually for each time series to into unobservable components whose values cannot be measured
be adjusted. It is made in a way to take due account of all objectively. Therefore a specification must be sought which provides
an overall plausible solution to the decomposition problem.
relevant information on the characteristics of the given
34 For instance, the “M4” indicator refers to autocorrelation of the
time series, extraordinary effects and the characteristic irregular component. High autocorrelation indicates that the irregular
component still has systematic patterns that should probably be
assigned to the other components, which could improve the specifi-
cation. However, the indicator does not indicate which of the specifi-
cation parameters should be adjusted. See Kirchner et al. (2018) for
the use of quality indicators.

Chart 2
Selected specification parameters of seasonal adjustment using three sets of statistics as examples
Production index for intermediate goods Nominal turnover in accommodation and Quarterly price adjusted gross domestic
food service activities product
Method X13 JD+ X-12-ARIMA X-12-ARIMA
Time series model multiplicative multiplicative multiplicative
Calendar regressors (1) Working days, January to November (1) Number of days in the month (no direct application of
(2) Working days, December (2) Number of Mondays calendar regressors; see Chapter 5.2 in
(3) Number of Tuesdays the text)
…
(7) Number of Saturdays
Outliers LS (09.2008) LS (05.2009) LS (01.2009) LS(Q4.2008)
LS (11.2008) LS (08.2009) LS(Q1.2009)
LS (12.2008) LS (03.2010) LS(Q2.2010)
LS (01.2009) LS (04.2010)
LS (03.2009) LS (01.2011)
ARIMA model ARIMA(013)(011)S ARIMA(011)(011)S ARIMA(010)(011)S
Trend filter Henderson 17 Henderson 17 Henderson 13
Saisonal filter 3x9 3x9 3x5: third quarter
3x9: other quarters
Support span January 2003 until current reference month January 1994 until current reference month 1st quarter of 1991 until current
(see Chapter 6.3 in the reference quarter
text)

As at May 2018.

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 15
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

6 made before each monthly or quarterly data publication.

The results of the calculation based on forecast seasonal
factors are then compared with those of the alternative
Setting a seasonal adjustment calculation; the seasonal factors are only updated if the
framework accuracy gains of the recalculation justify this.

This approach requires greater effort than an annual

6.1 Preliminary note adjustment of the seasonal factors involving an interim
use of forecast factors. For this reason the seasonal fac-
tors of statistics that are not so much in the user focus
Seasonal adjustment using X-12-ARIMA and X13 JD+
are adjusted only once per year. This often applies to
requires specification parameters to be set and also
time series at lower levels of detail; partly many hun-
various decisions to be taken that refer to the general
dreds of adjusted series are provided in great detail
framework of seasonal adjustment and apply to whole
every month (see Chapter 7).
sets of statistics or statistical domains. A number of
re­commendations regarding this framework are included
in the above-mentioned ESS guidelines on seasonal 6.3 Support span and routine revisions
adjustment (Eurostat, 2015).
To calculate calendar and seasonal factors using X-12-
ARIMA, the time series must cover a period of at least
6.2 Revision regimes
three years. Longer data series are useful, although time
series should not be too long (Eurostat, 2015, here:
As the seasonal patterns of time series and the relevant page 39 f.).
extent of seasonal effects can change over time, sea-
sonal factors should be reestimated regularly, at least Regarding the industrial production index, for exam-
once a year (Eurostat, 2015, here: page 33). | 35 Adjusting ple, data from 2003 onwards are currently used as the
and recalculating the seasonal factors leads to a higher support span for deriving the relevant factors. In the
accuracy of the estimated seasonality. At the same national accounts and in other statistics, the support
time, however, this means that the results of previous span begins in 1991. The revision policy permits time
periods of the seasonally adjusted time series may also series to be revised over the entire support span when
change so that revisions of the adjusted results would the seasonal factors are recalculated. Usually, however,
normally be necessary. An all too frequent adjustment the extent of revisions of individual time series values is
may therefore have disadvantages for users, especially small if these values refer to a time in the more distant
if the accuracy gains of recalculating the relevant fac- past. To avoid minor revisions of such back data, older
tors are not considerable, while extensive revisions seasonally adjusted results can be “frozen”, that is, they
would be required. In German official statistics, “Con- are no longer revised (Eurostat, 2015, here: page 34).
trolled current adjustment”, as it is called, is used for Regarding the production index, for instance, the sea-
adjustments according to X-12-ARIMA or X13 JD+. This sonally adjusted results have been frozen for the whole
is aimed at both achieving greatest possible accuracy of period before January 2015. | 36 In the quarterly national
the estimated seasonality and avoiding revisions where accounts however revisions of the whole time series
they are not worthwhile due to small accuracy gains. The are permitted when the final seasonal components are
seasonal factors are calculated once per year and fore- recalculated.
casted for at least one year. In addition, an alternative
calculation including reestimated seasonal factors is

35 This refers to changes in the seasonal pattern which go beyond devel-

opments that can be considered in the seasonal filters as these are
calculated as weighted moving averages and therefore can reflect,
to a certain extent, a gradual development of the seasonal pattern.
Calendar effects are typically rather stable so that, generally, the cal- 36 This approach is linked to the rebasing of the unadjusted data and
culated calendar factors have to be reviewed only once a year. their recalculation from 2015. See Linz et al. (2018) for rebasing.

16 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

6.4 Direct and indirect seasonal In the national accounts, however, the nominal data are
always obtained on the basis of indirect adjustment.
adjustment

The unadjusted values of the statistics to be adjusted 7

are typically provided as a set of time series subdivided,
for instance, by the economic branches covered by the
relevant statistics. In addition, aggregates are published Scope and critical assessment of the
which relate to several economic branches of the statis- seasonal adjustment processes
tics or to other breakdowns. In the national accounts, for
instance, aggregates are also calculated for institutional In the process of regular seasonal adjustment, the
sectors. Federal Statistical Office produces a large number of
adjusted time series at monthly and quarterly intervals.
The question which arises in relation to seasonal adjust- As far as adjustments based on X-12-ARIMA and X13
ment is whether the aggregates themselves should be JD+ are concerned, “Controlled current adjustment” is
adjusted (direct adjustment of aggregates) in addition used for nearly 800 directly adjusted series where the
to the time series of the economic branches, or whether seasonal factors and control features are reviewed every
the seasonally adjusted aggregates should be calcu- month or quarter. For another nearly 2,000 industry and
lated from the seasonally adjusted results of the indi- retail trade time series for which detailed results are
vidual branches of economic activity (indirect adjust- available, calendar and seasonal factors are forecasted
ment of aggregates; Deutsche Bundesbank, 2010). The once per year and then applied to the monthly data. | 38
solutions implemented in practice differ depending  Table 2 provides an overview of the number of directly
on the statistics concerned. Regarding the production adjusted time series of the statistical domains.
index for industry, direct seasonal adjustment is made
at a medium level of breakdown, while aggregates at Table 2
higher levels are adjusted indirectly. Adjustments in the Directly adjusted time series, X-12-ARIMA or X13 JD+, by
statistical domain
national accounts context are normally direct adjust-
ments at a very detailed level, while adjustments at Controlled cur- Current Adjust-
rent adjustment ment
higher levels are typically made indirectly. However, the
National accounts 212 –
total gross domestic product is adjusted in a direct man-
Time series in industry (industrial sector:
ner. | 37 Turnover in wholesale trade is directly adjusted at new orders, turnover, production, stock
all levels of aggregation. of orders, hours worked; construction:
production and new orders) 338 1,821
Labour cost index 54 –
In addition to nominal unadjusted values, price adjusted
Retail trade, retail sales of motor vehicles,
values (so-called volume data) are published for most of wholesale trade, accommodation and
the statistics. Again several practical seasonal adjust- food service activities 99 105
ment approaches are used in this respect. Regarding Services (turnover, persons employed,
wages and salaries, hours worked) 59 –
the production index for industry, first the nominal
ILO unemployment statistics 8 –
unadjusted values are adjusted directly. Dividing the Foreign trade 4 –
nominal by the price adjusted original values, implicit All domains 774 1,926
price series are generated that are subjected to direct
ILO: International Labour Organization.
seasonal adjustment. The seasonally adjusted volume
data are then obtained in an indirect manner by dividing
the adjusted nominal data by the adjusted price series. Since the beginning of 2018, the software used for the
regular seasonal adjustment of industrial indicators
of short-term statistics in industry has been gradually

37 To ensure coherence in the direct adjustment of the gross domestic 38 Apart from seasonal adjustment using X-12-ARIMA and X13 in JD+,
product (GDP), taxes on products are indirectly adjusted using a deri- many sets of statistics are additionally adjusted based on the BV4
vation scheme: taxes on products (indirect) = GDP (direct) – gross procedure. And there are statistics which are adjusted using exclu-
value added (indirect) + subsidies on products (direct) sively BV4.

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 17
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

changed over to JDemetra+ (incl. X13). In the national relevant economic branches. Although, as mentioned
accounts, changing over to this software is scheduled above, the Federal Statistical Office always provides the
for autumn 2019. Corresponding stepwise changes are unadjusted values in addition to the seasonally adjusted
also intended for the remaining short-term statistics. results, for clarity reasons the focus is only on one type
of results in first releases. Taking into account the needs
The regular calculation of seasonally adjusted results of the different target groups, and in accordance with
is an integral part of the statistical production process. the recommendations of the European Statistical Sys-
In the Generic Statistical Business Process Model, it is tem and the Organisation for Economic Co-operation
assigned to Phase 6 whose subprocesses are summa- and Development (OECD), it was finally decided for most
rised under “Analyse”. | 39  Chart 3 short-term statistics to list the month-on-month or quar-
ter-on-quarter change rates of the seasonally adjusted
Chart 3
results generally at the top of press releases. | 40
Phases of the Generic Statistical
Business Process Model
In assessing economic developments, other aspects
Phase 1 – Specify needs
than changes of seasonally adjusted data on the previ-
Phase 2 – Design
ous period may play a role and be more relevant to certain
Phase 3 – Build user needs. Let us look again at the example of Figure 1.
Phase 4 – Collect Here the fact that, although the results were slightly
Phase 5 – Process down in the recent past, they were still at a rather high
Phase 6 – Analyze level at the time of observation might be important in
Phase 7 – Disseminate describing the economic situation from a user perspec-
Phase 8 – Evaluate tive. In some of these cases, a year-on-year comparison
is additionally used. However, it could also be useful to
compare the relevant level with a longer-term average
Regarding short-term statistics, for instance indices and that does not change from period to period. This could
growth rates are determined during this phase. Calculat- for instance be achieved by providing information on the
ing seasonally adjusted results constitutes an additional (positive or negative) extent to which the current results
step of analysis in this phase, which increases the data differ from such a longer-term average which could be
relevance further. In more concrete terms, user needs calculated, for example, from a long-term trend. | 41
are met regarding current results of economic develop-
ment which are not masked by seasonal fluctuations And finally, there is a competitive relationship between
that can be expected to recur and calendar effects, and the relevance of the results and other quality aspects. As
which therefore provide a better understanding of cycli- mentioned above, seasonal and calendar components
cal developments. However, various conflicting goals are unobservable components which need to be made
may occur in this respect. measurable by operationalisation approaches (selecting
the adjustment procedure, setting parameters). Usually,
First the relevance of statistical results can only be different operationalisation approaches lead to different
assessed in relation to a target group or a certain pur- results. To avoid, as far as possible, potential losses of
pose of use. For instance, some business associations
and enterprises (which, as reporting units, often pro- 40 The ESS guidelines on seasonal adjustment require above all sea-
vide individual data as the basis for the unadjusted val- sonally adjusted results to be presented in press releases (Eurostat,
2015, here: page 46): “Seasonally adjusted data are the most
ues) regard the fact that seasonally adjusted results do appropriate figures to be presented in press releases“. And the OECD
not reflect the movements that are typical, known and recommendations even apply to countries beyond Europe: “When
applicable, the focus of press releases [...] concerning the main
expected by the operators in a given branch as a disad- sub-annual indicators should be on their appropriately seasonally
vantage. This target group sometimes prefers to use the adjusted version.“ (OECD, 2007, here: page 20 f.).
unadjusted data for analysing the development of the 41 The Business Cycle Monitor of the Federal Statistical Office (www.
destatis.de > Visualised statistics > Business Cycle Monitor) for
instance combines the development at the current end with the
difference between the current results and a long-term trend. Here
39 See Gehle/Lüüs (2017) for the Generic Statistical Business Process a distinction is made between four different phases of short-term
Model, here: page 48. development.

18 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

objectivity and clarity of the results, the Federal Statis-

tical Office uses X-12-ARIMA and X13 JD+ for seasonal
adjustment, which are internationally recognised and
harmonised methods. Seasonal adjustment continues
to be carried out in close collaboration with Deutsche
Bundesbank, and best-possible transparency is pro-
vided in presenting the underlying methods.

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 19
Dr. Stefan Linz, Dr. Claudia Fries, Julia Völker

LITERATURE

Bald-Herbel, Christiane. Umstellung der Konjunkturindizes im Produzierenden

Gewerbe auf das Basisjahr 2010. In: Wirtschaft und Statistik. Edition 3/2013,
pages 185 ff.

Bee Dagum, Estela/Bianconcini, Silvia. Saisonal Adjustment Methods and Real Time
Trend-Cycle Estimation. 2016.

Box, George E. P./Jenkins, Gwilym M. Time Series Analysis: Forecasting and Control.
San Francisco 1970.

Box, George E. P./Jenkins, Gwilym M./Reinsel, Gregory C./Ljung, Greta M. Time Series
Analysis: Forecasting and Control. 5th edition. 2015.

Deutsche Bundesbank/Statistisches Bundesamt. Memorandum of Understanding.

Frankfurt 2014. [Accessed on 27 June 2018]. Available at: www.destatis.de

Deutsche Bundesbank. Kalendarische Einflüsse auf das Wirtschaftsgeschehen.

In: Monatsbericht Dezember 2012, pages 53 ff.

Deutsche Bundesbank. Das Ganze und seine Teile: Aggregationsprobleme saison-

bereinigter Daten. In: Monatsbericht Juni 2010, pages 63 ff.

Eurostat. ESS guidelines on seasonal adjustment. Luxembourg 2015.

[doi: 10.2785/317290]

Eurostat. JDemetra+ officially recommended as software for the seasonal adjustment of

official statistics. [Accessed on 27 June 2018]. Available at:
http://ec.europa.eu/eurostat

Findley, David F./Monsell, Brian C./Bell, William R./Otto, Mark C./Chen, Bor-Chung.
New Capabilities and Methods of the X-12-ARIMA Seasonal-Adjustment Program. In:
Journal of Business & Economic Statistics. Volume 16. 1998, pages 127 ff.

Gehle, Christian/Lüüs, Hans-Peter. Prozessmanagement im Statistischen Bundesamt.

In: WISTA Wirtschaft und Statistik. Edition 5/2017, pages 46 ff.

Gericke, Pierre-André/Seidel, Gerald. Saisonbereinigung. Methodenbericht der

Statistik der Bundesagentur für Arbeit. Nürnberg 2014.

Gómez, Victor/Maravall, Agustín. Seasonal Adjustment and Signal Extraction in Eco-

nomic Time Series, Chapter 8. In: Peña, Daniel/Tiao, George C./Tsay, Ruey S. (editors).
A Course in Time Series Analysis. New York 2001.

Hauf, Stefan. Saisonbereinigung des Statistischen Bundesamtes im Bereich der Volks­

wirtschaftlichen Gesamtrechnungen. In: Allgemeines Statistisches Archiv. Volume 86
(2002), pages 119 ff.

20 Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq
Seasonal adjustment of short-term statistics using X-12-ARIMA and X13 in JDemetra+

LITERATURE

Kirchner, Robert. Auswirkungen des neuen Saisonbereinigungsverfahrens Census

X-12-ARIMA auf die aktuelle Wirtschaftsanalyse in Deutschland. Diskussionspapier
7/99 der Volkswirtschaftlichen Forschungsgruppe der Deutschen Bundesbank.
Frankfurt am Main 1999.

Kirchner, Robert/Ladiray, Dominique/Mazzi, Gian L. Quality Measures and Reporting

for Seasonal Adjustment. In: Eurostat (editor). Handbook on Seasonal Adjustment.
Edition 2018.

Ladiray, Dominique/Quenneville, Benoit. Seasonal Adjustment with the X-11 Method.

Lecture Notes in Statistics. Volume 158. New York 2001.

Linz, Stefan/Möller, Hans-Rüdiger/Mehlhorn, Peter. Umstellung der Konjunkturindizes

im Produzierenden Gewerbe auf das Basisjahr 2015. In: WISTA Wirtschaft und Statis-
tik. Edition 2/2018, pages 49 ff.

Nazmen, Sufi M. Applied time series analysis for business and economic forecasting.
In: Owen, Donald B. Statistics: textbooks and monographs. Volume 93. 1988.

OECD. Data and Metadata Reporting and Presentation Handbook. Paris 2007.

Speth, Hans-Theo. Komponentenzerlegung und Saisonbereinigung ökonomischer

Zeitreihen mit dem Verfahren BV4.1. Methodenberichte des Statistischen Bundes­
amtes. Issue 3. Wiesbaden 2004.

Waldmüller, Bernd/Weisbrod, Joachim. Neuere Entwicklungen in den Unternehmens­

statistiken. In: WISTA Wirtschaft und Statistik. Edition 5/2015, pages 33 ff.

LEGAL BASIS

Council Regulation (EC) No 1165/98 of 19 May 1998 concerning short-term statistics

(Official Journal of the EC L 162, page 1), consolidated version, last amended by Com-
mission Regulation (EU) No 461/2012 of 31 May 2012 (Official Journal of the
EU L 142, page 26).

Commission Regulation (EC) No 1216/2003 of 7 July 2003 implementing Regula-

tion (EC) No 450/2003 of the European Parliament and of the Council concerning the
labour cost index (Official Journal of the EC L 169, page 37), consolidated version, last
amended by Commission Regulation (EC) No 973/2007 of 20 August 2007(Official
Journal of the EU L 216, page 10).

Regulation (EU) No 549/2013 of the European Parliament and of the Council of

21 May 2013 on the European system of national and regional accounts in the Euro-
pean Union (Official Journal of the EU L 174, page 1).

Statistisches Bundesamt (Federal Statistical Office) | German version published in WISTA | 4 | 2018, p. 59 et seq 21
Extract from the journal
WISTA Wirtschaft und Statistik
Published by:
Statistisches Bundesamt (Federal Statistical Office)
www.destatis.de
You may contact us at
www.destatis.de/kontakt

Abbreviations Explanation of symbols

WISTA = Wirtschaft und Statistik – = no figures or magnitude zero
JD = annual average 0 = less than half of 1 in the last digit
D = average (for values which cannot be added occupied, but more than zero
up) . = numerical value unknown or not to be
Vj = quarter of a year disclosed

Hj = half-year ... = data will be available later

a. n. g. = not elsewhere classified X = cell blocked for logical reasons

o. a. S. = no main economic activity I or — = fundamental change within a series affect-

ing comparisons over time
St = piece
/ = no data because the numerical value is not
Mill. = million sufficiently reliable
Mrd. = billion () = limited informational value because
numerical value is of limited statistical
reliability

© Statistisches Bundesamt 2019 Figures have in general been roundes without taking account
of the totals, so that there may be an apparent slight dis-
Reproduction and distribution, also of parts, are permitted crepancy between the sum of the constituent items and the
provides that the source is mentioned. total as shown.