Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Rocha and Cribari-Neto (2009) Beta Autoregressive Moving Average Models

Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Test (2009) 18: 529545

DOI 10.1007/s11749-008-0112-z
ORI GI NAL PAPER
Beta autoregressive moving average models
Andra V. Rocha Francisco Cribari-Neto
Received: 1 November 2007 / Accepted: 26 May 2008 / Published online: 13 June 2008
Sociedad de Estadstica e Investigacin Operativa 2008
Abstract We build upon the class of beta regressions introduced by Ferrari and
Cribari-Neto (J. Appl. Stat. 31:799815, 2004) to propose a dynamic model for con-
tinuous random variates that assume values in the standard unit interval (0, 1). The
proposed ARMA model includes both autoregressive and moving average dynam-
ics, and also includes a set of regressors. We discuss parameter estimation, hypothe-
sis testing, goodness-of-t assessment and forecasting. In particular, we give closed-
form expressions for the score function and for Fishers information matrix. An ap-
plication that uses real data is presented and discussed.
Keywords ARMA Beta distribution Beta ARMA Forecasts
Mathematics Subject Classication (2000) 62M10 91B84
1 Introduction
The beta distribution is commonly used for modeling experiments in which the vari-
able of interest is continuously distributed in the interval (a, b), where a and b are
known scalars, and a <b, since its density can assume quite different shapes depend-
ing on the values of the two parameters that index the distribution. A particularly use-
ful situation occurs when a =0 and b =1 so that the random variable assumes values
in the standard unit interval, (0, 1); this is the case, e.g., of rates or proportions.
The beta probability density function is given by
(y; p, q) =
(p +q)
(p)(q)
y
p1
(1 y)
q1
, 0 <y <1, (1)
A.V. Rocha () F. Cribari-Neto
Departamento de Estatstica, Universidade Federal de Pernambuco. Cidade Universitria, Recife,
PE, 50740-540, Brazil
e-mail: andrea@cox.de.ufpe.br
530 A.V. Rocha, F. Cribari-Neto
where p > 0, q > 0, and () is the gamma function. The mean and variance of y
are, respectively,
E(y) =
p
p +q
and Var(y) =
pq
(p +q)
2
(p +q +1)
.
The mode of the distribution exists when both p and q are greater than one, in which
case, mode(y) =(p1)/(p+q 2). The uniform distribution is a special case of (1)
when p =q =1. Estimation of p and q can be carried out by maximum likelihood.
Small sample bias adjustments to the maximum likelihood estimators of p and q
were obtained by Cribari-Neto and Vasconcellos (2002).
Ferrari and Cribari-Neto (2004) proposed a regression model in which the de-
pendent variable is beta distributed. Their parameterization is as follows.
1
Let =
p/(p+q) and =p+q, i.e., p = and q =(1); here, 0 < <1 and >0.
It then follows that the mean and the variance of y are, respectively,
E(y) = and Var(y) =
V()
1 +
,
where V() = (1 ). Note that can be interpreted as a precision parameter
in the sense that, for a given value of , the larger the value of , the smaller the
variance of y.
Using this parameterization, Ferrari and Cribari-Neto (2004) dened a regression
model which in many aspects resembles the class of generalized linear models (see,
for example, Nelder and Wedderburn 1972, and McCullagh and Nelder 1989). Their
model, however, is not a generalized linear model (GLM).
Our chief goal in this paper is to propose a time series model for random variables
that assume values in the standard unit interval. The approach is based on the class
of beta regression models of Ferrari and Cribari-Neto (2004). Our approach is also
similar to those of Benjamin et al. (2003) and Shephard (1995) (see also Li 1994,
and Fokianos and Kedem 2004), who have developed dynamic models for random
variables in the exponential family. We note that Zeger and Qaqish (1988) proposed
the so-called Markov regression models (which extends the class of GLMs) and that
Li (1991) developed goodness-of-t tests for such models. In this paper, we propose
the beta autoregressive moving average model (ARMA). It can be used to model
and forecast variates that assume values in the standard unit interval, such as rates
and proportions. The use of the ARMA model avoids the need to transform the data
prior to modeling. Moreover, the distributions of rates and proportions are typically
asymmetric and, hence, Gaussian-based inference is not appropriate. The ARMA
model naturally accommodates asymmetries and also non-constant dispersion.
The paper unfolds as follows. Section 2 introduces the proposed model, Sect. 3
focuses on parameter estimation, Sect. 4 considers further inference strategies and
prediction, and Sect. 5 illustrates the methodology by applying the model to real
data. Finally, concluding remarks are given in Sect. 6.
1
For an alternative formulation of the class of beta regressions, see Vasconcellos and Cribari-Neto (2005).
Beta autoregressive moving average models 531
2 The model
Our goal is to dene a dynamic model for beta distributed random variables observed
over time. For both regression and time series analysis it is typically more convenient
to work with the mean response and also with a precision (or dispersion) parameter.
Therefore, we shall employ the beta parameterization given in Ferrari and Cribari-
Neto (2004).
We shall assume that the response is continuous and takes values in the standard
unit interval (0, 1). We note, however, that the proposed model is also useful in sit-
uations where the response is restricted to the interval (a, b), where a and b are
known scalars (a < b). In this case, one can model (y a)/(b a) instead of mod-
eling y directly. We shall also assume that the covariates x
t
, t = 1, . . . , n, where
x
t
=(x
t 1
, . . . , x
t k
)

, are non-random. Here, n denotes the sample size and k <n.


Let y
t
, t =1, . . . , n, be random variables and assume that the conditional distribu-
tion of each y
t
, given the previous information set F
t 1
(i.e., the smallest -algebra
such that the variables y
1
, . . . , y
t 1
are measurable), follows the beta distribution.
That is, the conditional density of y
t
given F
t 1
is
f (y
t
| F
t 1
) =
()
(
t
)((1
t
))
y

t
1
t
(1 y
t
)
(1
t
)1
, 0 <y
t
<1, (2)
where E(y
t
| F
t 1
) =
t
and Var(y
t
| F
t 1
) = V(
t
)/(1 +) are, respectively, the
conditional mean and the conditional variance of y
t
; here, V(
t
) =
t
(1
t
).
In the class of beta regression models (see Ferrari and Cribari-Neto 2004),
t
is
related to a linear predictor,
t
, through a twice differentiable strictly monotonic link
function g : (0, 1) R. The most commonly used link functions are the logit, probit,
and complementary loglog links. Unlike the linear predictor of the beta regression
model, in the systematic component of the ARMA specication there is an addi-
tional component,
t
, which allows autoregressive and moving average terms to be
included additively. Thus, a general model for
t
is given by
g(
t
) =
t
=x

t
+
t
,
where = (
1
, . . . ,
k
)

is a set of unknown linear parameters and


t
is an ARMA
component which shall be described belowand is similar to what is given in Benjamin
et al. (2003).
We shall now motivate the denition of the ARMA component
t
. Consider an
ARMA(p, q) model initially as function of a term
t
, such that
t
= g(y
t
) x

t
.
Then,

t
= +
p

i=1

t i
+
q

j=1

j
r
t j
+r
t
, (3)
where r
t
denotes a random error and R is a constant. Although we have not
dened r
t
, it is assumed that E(r
t
| F
t 1
) =0. Taking conditional expectations with
respect to the -algebra F
t 1
in (3) we obtain the approximate model

t
= +
p

i=1

t i
+
q

j=1

j
r
t j
.
532 A.V. Rocha, F. Cribari-Neto
Note that
t i
with i > 0 is F
t 1
-measurable, and E(
t
| F
t 1
)
t
. Therefore, we
obtain the following expression for
t
:

t
= +
p

i=1

i
_
g(y
t i
) x

t i

_
+
q

j=1

j
r
t j
,
where x
t
R
k
, = (
1
, . . . ,
k
)

, k < n, and p, q N are, respectively, the au-


toregressive and moving average orders. The s and the s are the autoregres-
sive and moving average parameters, respectively, and r
t
is an error. Finally, since

t
=g(
t
) x

t
, we propose the following general model for the mean
t
:
g(
t
) = +x

t
+
p

i=1

i
_
g(y
t i
) x

t i

_
+
q

j=1

j
r
t j
. (4)
The ARMA(p, q) model is dened by (2) and (4). It is noteworthy that both
the tted values and the out-of-sample forecasts obtained using the ARMA model
will belong to the standard unit interval. There are several choices for the moving
average error terms; for example, errors measured on the original scale (i.e., y
t

t
),
on the predictor scale (i.e., g(y
t
)
t
), etc. What is required is that the error r
t
be
measurable with respect to F
t
.
Let us obtain the mean and variance of two errors. For y
t

t
, we have
E(y
t

t
| F
t 1
) =0 and Var(y
t

t
| F
t 1
) =
V(
t
)
1 +
,
where V(
t
) =
t
(1
t
). In particular, E(y
t

t
) = 0 and Var(y
t

t
) =
V(
t
)/(1 +). Note that the errors are orthogonal, since for i <j
E
_
(y
i

i
)(y
j

j
)
_
=E
_
(y
i

i
)E(y
j

j
| F
j1
)
_
=0. (5)
Since g() is continuously differentiable, we can Taylor-expand it as
g(y
t
) g(
t
) +g

(
t
)(y
t

t
) g(y
t
) g(
t
) g

(
t
)(y
t

t
).
Moreover,
t
=g(
t
), then, for the error g(y
t
)
t
,
E
_
g(y
t
)
t
| F
t 1
_
E
_
g

(
t
)(y
t

t
) | F
t 1
_
=0.
Given that g() is twice differentiable, it follows from the delta method that
Var
_
g(y
t
)
t
| F
t 1
_

_
g

(
t
)
_
2
V(
t
)
1 +
.
In particular, E(g(y
t
)
t
) 0 and Var(g(y
t
)
t
) (g

(
t
))
2
V(
t
)/(1+). With
an analogous argument to the one used with (5), we conclude that these errors are also
approximately orthogonal.
Beta autoregressive moving average models 533
3 Parameter estimation
The estimation of the parameters that index the ARMA model can be carried out by
maximumlikelihood. Let us denote the vector of parameters as =(,

, ,

,
where =(
1
, . . . ,
p
)

and =(
1
, . . . ,
q
)

. As noted earlier, we assume that the


covariates x
t
are non-stochastic.
The log-likelihood function for the parameter vector conditional on the rst
m observations, where m = max{p, q}, is =

n
t =m+1
logf (y
t
| F
t 1
), with
f (y
t
| F
t 1
) given in (2). Expectations are also taken in conditional fashion. Note
that, conditioned to F
m
, the rst m errors are zero (or approximately zero). Thus,
in the construction of the conditional log-likelihood function, the rst q errors are
assumed to equal zero.
3.1 Score vector
Let logf (y
t
| F
t 1
) =
t
(
t
, ). Then

t
(
t
, ) = log() log(
t
) log
_
(1
t
)
_
+(
t
1) logy
t
+
_
(1
t
) 1
_
log(1 y
t
).
Therefore, the conditional log-likelihood function is
=
n

t =m+1

t
(
t
, ).
Thus,

=
n

t =m+1

t
(
t
, )

t
d
t
d
t

.
Note that d
t
/d
t
=1/g

(
t
). We also have that

t
(
t
, )

t
=
_
log
y
t
1 y
t

_
(
t
)
_
(1
t
)
__
_
, (6)
where () is the digamma function, i.e., (z) =d log(z)/dz for z > 0. Let y

t
=
log{y
t
/(1 y
t
)} and

t
=(
t
) ((1
t
)).Then,

=
n

t =m+1
(y

t
)
1
g

(
t
)
.
Additionally, for l =1, . . . , k,

l
=
n

t =m+1

t
(
t
, )

t
d
t
d
t

l
.
534 A.V. Rocha, F. Cribari-Neto
Then,

l
=
n

t =m+1
(y

t
)
1
g

(
t
)
_
x
t l

p

i=1

i
x
(t i)l
_
.
Furthermore,

=
n

t =m+1
_

t
(y

t
) +log(1 y
t
)
_
(1
t
)
_
+()
_
.
Note also that, for i =1, . . . , p,

i
=
n

t =m+1

t
(
t
, )

t
d
t
d
t

i
,
which yields

i
=
n

t =m+1
(y

t
)
1
g

(
t
)
_
g(y
t i
) x

t i

_
.
Finally, for j =1, . . . , q,

j
=
n

t =m+1

t
(
t
, )

t
d
t
d
t

j
.
Therefore,

j
=
n

t =m+1
(y

t
)
1
g

(
t
)
r
t j
.
It is now possible to obtain the score vector U( ). Let y

= (y

m+1
, . . . , y

n
)

= (

m+1
, . . . ,

n
)

and T = diag{1/g

(
m+1
), . . . , 1/g

(
n
)}. Let also 1 be an
n 1 vector of ones, M be the (n m) k matrix with (i, j)th element given by
x
(i+m)j

p
l=1

l
x
(i+ml)j
, P be the (n m) p matrix whose (i, j)th element
equals g(y
i+mj
) x

i+mj
and R be the (n m) q matrix with (i, j)th element
given by r
i+mj
. Hence,
U

( ) = 1

T (y

),
U

( ) = M

T (y

),
U

( ) =
n

t =m+1
_

t
(y

t
) +log(1 y
t
)
_
(1
t
)
_
+()
_
,
U

( ) = P

T (y

),
and
U

( ) =R

T (y

).
Beta autoregressive moving average models 535
Therefore, the score vector is
U( ) =

( )
U

( )
U

( )
U

( )
U

( )

,
which is of dimension (k + p + q + 2) 1. The conditional maximum likelihood
estimator (CMLE) of is obtained as the solution of the system of equations given
by U( ) =0. Note that it does not have closed-form. Hence, it has to be numerically
obtained by maximizing the conditional log-likelihood function using a nonlinear
optimization algorithm, such as a Newton or quasi-Newton algorithm (see Nocedal
and Wright 1999).
3.2 Conditional Fishers information matrix
In what follows
i
and
i
will be used as surrogates for
i
,
i
or
i
. We have

j
=
n

t =m+1

t
_

t
(
t
, )

t
d
t
d
t

j
_
d
t
d
t

i
=
n

t =m+1
_

t
(
t
, )

2
t
d
t
d
t

j
+

t
(
t
, )

t
_
d
t
d
t

j
__
d
t
d
t

i
.
Since we are working with the conditional likelihood, we know, from the regu-
larity conditions, that E(
t
(
t
, )/
t
| F
t 1
) = 0; in particular, we have that
E(
t
(
t
, )/
t
) =0.
We also note that
t
/
l
=x
t l

p
i=1

i
x
(t i)l
,
t
/
i
=g(y
t i
) x

t i
and

t
/
j
=r
t j
are F
t 1
-measurable (since F
t
is a ltration). Thus, it follows from
the regularity conditions that
E
_

2

F
t 1
_
=
n

t =m+1
E
_

t
(
t
, )

2
t

F
t 1
__
d
t
d
t
_
2

i
.
From (6) we obtain

t
(
t
, )

2
t
=
_

(
t
) +

_
(1
t
)
__
.
Furthermore,
E
_

2

F
t 1
_
=
n

t =m+1
{

(
t
) +

((1
t
))}
g

()
2

i
.
536 A.V. Rocha, F. Cribari-Neto
Note that

=
n

t =m+1
_

t
(
t
, )

2
t
d
t
d
t

i
+

t
(
t
, )

t
_
d
t
d
t

i
__
d
t
d
t

.
Hence,
E
_

2

F
t 1
_
=
n

t =m+1
{

(
t
) +

((1
t
))}
g

()
2

i
.
Moreover,

2
=
n

t =m+1
_

t
(
t
, )

2
t
d
t
d
t

+

t
(
t
, )

t
_
d
t
d
t

__
d
t
d
t

.
Thus,
E
_

F
t 1
_
=
n

t =m+1
{

(
t
) +

((1
t
))}
g

()
2
.
We have that

j
=
n

t =m+1
(y

t
)
1
g

(
t
)

j
.
Therefore,

=
n

t =m+1
_
(y

t
)

_
1
g

(
t
)

i
.
It also follows from the regularity conditions that E(y

t
| F
t 1
) =

t
. Given that

t
/ =

(
t
)
t

((1
t
))(1
t
), we have
E
_

2

F
t 1
_
=
n

t =m+1

(
t
)
t

((1
t
))(1
t
)
g

(
t
)
d
t
d
i
.
We also have that

=
n

t =m+1
_
(y

t
)

_
1
g

(
t
)

,
which yields
E
_

2

F
t 1
_
=
n

t =m+1

(
t
)
t

((1
t
))(1
t
)
g

(
t
)
.
Beta autoregressive moving average models 537
Finally,
2
/
2
follows from the differentiation of U

( ) with respect to . We
obtain
E
_

F
t 1
_
=
n

t =m+1
_

(
t
)
2
t
+

_
(1
t
)
_
(1
t
)
2

()
_
.
Using
d
t
d
l
=x
t l

p

i=1

i
x
(t i)l
,
d
t
d
i
=g(y
t i
) x

t i
, and
d
t
d
j
=r
t j
,
we can obtain Fishers information matrix for . Let W =diag{w
m+1
, . . . , w
n
}, with
w
t
=
{

(
t
) +

((1
t
))}
g

(
t
)
2
,
c =(c
m+1
, . . . , c
n
)

, with c
t
={

(
t
)
t

((1
t
))(1
t
)}, and D =
diag{d
m+1
, . . . , d
n
}, with d
t
=

(
t
)
2
t
+

((1
t
))(1
t
)
2

(). Thus,
E
_

F
t 1
_
= tr(W), E
_

2

F
t 1
_
=M

W1,
E
_

2

F
t 1
_
= 1

T c, E
_

2

F
t 1
_
=P

W1,
E
_

2

F
t 1
_
= M

WM, E
_

2

F
t 1
_
=M

T c,
E
_

F
t 1
_
= tr(D), E
_

2

F
t 1
_
=P

WP,
E
_

2

F
t 1
_
= P

T c, E
_

2

F
t 1
_
=M

WM,
E
_

2

F
t 1
_
= R

T c, E
_

2

F
t 1
_
=M

WP,
E
_

2

F
t 1
_
= M

WR, E
_

2

F
t 1
_
=P

WR,
and
E
_

2

F
t 1
_
=R

W1.
538 A.V. Rocha, F. Cribari-Neto
Therefore, Fishers information matrix can be expressed as
K =K( ) =

,
where K

= tr(W), K

= K

= M

W1, K

= K

= 1

T c, K

= K

=
P

W1, K

= K

= R

W1, K

= M

WM, K

= K

= M

T c, K

=
tr(D), K

= K

= M

WP, K

= K

= R

WR, K

= P

WP, K

=
K

=P

T c, K

=R

WR, K

=K

=R

T c, and K

=K

=R

MP.
Note that Fishers information matrix is not block-diagonal, which implies that our
model is not a dynamic GLM. Under the usual regularity conditions for maximum
likelihood estimation and when the sample size is large,

N
(k+p+q+2)

, K
1

approximately, where N
r
denotes the r-dimensional normal distribution, and ,

,

, , and

are the CMLEs of , , , , and , respectively.


4 Hypothesis testing and prediction
Consider the following null and alternative hypotheses:
H
0
: t =0 and H
1
: t =0, (7)
where t is an r (k +p +q +2) matrix (r <k +p +q +2) of rank r. For instance,
consider the following partition of the (k + p + q + 2) 1 parameter vector: =
(

1
,

2
)

, where
2
is r 1 (r < k +p +q +2). Note that by letting t =
2
in (7)
one can test whether
2
equals zero.
Let be the CMLE of under the null hypothesis in (7) and let be the un-
restricted CMLE of . The test statistic commonly used to test H
0
: t = 0 is the
conditional log-likelihood ratio statistic (CLR):

n
=2
_
( ) ( )
_
,
where () is the conditional log-likelihood function. Under mild regularity condi-
tions and under H
0
,
n
D

2
r
, where
D
denotes convergence in distribution, so
Beta autoregressive moving average models 539
that the test can be performed using approximate critical values from the limiting
2
r
null distribution. One can also base the testing inference on the square root of the
CLR statistic, where the sign of the statistic is that of ( ), which is asymptoti-
cally standard normal under the null hypothesis. It is also noteworthy that by using
the asymptotic normality of the CMLE of , , one can easily construct approximate
condence intervals for the elements of .
We shall now consider tests of model adequacy. Li (1991) proposed Portmanteau
and score statistics for Markov regression models. We shall now follow his approach
to provide Portmanteau and score statistics for the ARMA model. At the outset,
consider the standardized score errors dened as
a
t
=
y

(
t
) +

((1
t
))
,
and note that it follows from the regularity conditions that E(a
t
|F
t 1
) = 0 and
Var(a
t
|F
t 1
) = 1/; also, E(a
i
a
j
) = 0 whenever i = j. Then, the lag k innovation
autocorrelation of a
t
is
C
k
=
1
n
n

t =k+1
a
t
a
t k
.
The corresponding kth residual autocorrelation,

C
k
, can be written as

C
k
=
1
n
n

t =k+1

a
t
a
t k
.
We shall work with the subsets of the score vector and Fishers information matrix
relative to , and . The following quantity will be useful:
V = lim
n
1
n
K,
where K denotes Fishers information matrix. Consider the following partition
of V
1
:
V
1
= lim
n
1
n

,
and let V

be the block of V
1
which corresponds to , , , and :
V

= lim
n
1
n

.
540 A.V. Rocha, F. Cribari-Neto
Let

C =(

C
1
, . . . ,

C
m
)

for some m>0. Then, following Sect. 2 and the Appendix


of Li (1991), it can be shown that, under correct model specication,

n

C is asymp-
totically normally distributed with mean zero and variance I
m
X

X, where
I
m
is the mm identity matrix and
X = lim
n
1
n

h
t
a
t 1


h
t
a
t m

(x
t 1

p
i=1

i
x
(t i)1
)h
t
a
t 1


(x
t 1

p
i=1

i
x
(t i)1
)h
t
a
t m
.
.
.
.
.
.
.
.
.

(x
t k

p
i=1

i
x
(t i)k
)h
t
a
t 1


(x
t k

p
i=1

i
x
(t i)k
)h
t
a
t m

(g(y
t 1
) x

t 1
)h
t
a
t 1


(g(y
t 1
) x

t 1
)h
t
a
t m
.
.
.
.
.
.
.
.
.

(g(y
t p
) x

t p
)h
t
a
t 1


(g(y
t p
) x

t p
)h
t
a
t m

r
t 1
h
t
a
t 1


r
t 1
h
t
a
t m
.
.
.
.
.
.
.
.
.

r
t q
h
t
a
t 1


r
t q
h
t
a
t m

,
with
h
t
=
(

(
i
) +

((1
i
)))
1/2
g

(
t
)
.
Hence, a test for the joint signicance of the rst m autocorrelations can be based on
n

C

(I
m


X)
1

C, which is asymptotically
2
m
under the null hypothesis
of no serial correlation.
Score tests on the parameter vector can be performed using the approach proposed
by Li (1991). Let
=(

1
,

2
)

, =(

1
,

2
)

, =(

1
,

2
)

, and =(

2
,

2
,

2
)

.
The null hypothesis is = 0 which is to be tested against the alternative that the
number of parameters is 1+k
1
+k
2
+p
1
+p
2
+q
1
+q
2
, where, k
i
, p
i
, and q
i
are the
number of parameters in
i
,
i
, and
i
, respectively, for i =1, 2.
2
The corresponding
score function is

U( ) =

( )
U

( )
U

( )
U

( )

,
where = (, , , , ). It is possible to show that

U/

n is asymptotically nor-
mally distributed with mean zero and variance V

when the null hypothesis


2
Note that we do not include in ; one can, however, consider the case where =(,

2
,

2
,

2
)

when
the null hypothesis also imposes =0.
Beta autoregressive moving average models 541
is true, where V

is the part of V that corresponds to , , , and . Then,


following Li (1991), and noting that under the null hypothesis

U(
1
) = 0, where

1
=(,

1
, ,

1
,

1
)

, a score test statistic is


Q=n
1

U( )

U( ),
where the estimates are obtained under the null hypothesis. Let a = ( a
1
, . . . , a
n
),
S
t
=h
t
Z
t
, S =(S
1
, . . . , S
n
)

and
Z
t
=

1
(x
t 1

p
i=1

i
x
(t i)1
)
.
.
.
(x
t k

p
i=1

i
x
(t i)k
)
(g(y
t 1
) x

t 1
)
.
.
.
(g(y
t p
) x

t p
)
r
t 1
.
.
.
r
t q

.
We can rewrite Q as
Q=

2
a

S

V

a
n
.
An asymptotically equivalent statistic is

Q=

2
a

SK

a,
where K

is the block of Fishers information matrix inverse which corresponds


to , , , and :
K

.
Under the null hypothesis, Q is asymptotically
2
k
2
+p
2
+q
2
. We note that the condi-
tional score test that we just developed only requires the estimation of the null model.
In order to produce forecasts, the CMLE of , , must be used to obtain estimates
for
t
, t =m+1, . . . , n, say
t
. By using
t
one can obtain the estimates of r
t
, r
t
,
for t =m+1, . . . , n (based on the functional structure of the error). For N > n, the
forecast of the error r
N
equals zero. Thus, to predict the mean value of the process at
T >n, one should use the CMLE of , ; the estimates of
t
, t =m+1, . . . , n; the
estimates of r
t
, t =m+1, . . . , n; replace r
t
by zero if t > n (these sufce to obtain
542 A.V. Rocha, F. Cribari-Neto

n+1
, and one can then proceed analogously to obtain
n+2
, and so on); and replace
y
t
by
t
if n <t <T . For instance, the mean response estimate at n +1 is

n+1
=g
1
_
+x

n+1

+
p

i=1

i
_
g(y
n+1i
) x

n+1i

_
+
q

j=1

j
r
n+1j
_
.
At time n +2, we obtain

n+2
= g
1
_
+x

n+2

+
p

i=2

i
_
g(y
n+2i
) x

n+2i

_
+
1
_
g(
n+1
) x

n+1

_
+
q

j=2

j
r
n+2j
_
,
and so on.
Finally, we note that model selection can be performed using the Akaike informa-
tion criterion (AIC) introduced by Akaike (1973, 1974) or, alternatively, the Bayesian
information criterion (BIC) of Schwarz (1978). For a detailed discussion of informa-
tion criteria and their properties, see Choi (1992).
5 An application
This section contains an application of the ARMA model proposed in Sect. 2. The
estimations and computations were carried out using the free statistical software R;
see http://www.r-project.org. We used the quasi-Newton algorithm known as BFGS
to maximize the conditional log-likelihood function. The data refers to the rate of hid-
den unemployment due to substandard work conditions in So Paulo, Brazil (TDOP-
RMSP). Hidden unemployment due to substandard work conditions relates to people
who work illegally, who perform unpaid work for relatives, and also who have been
seeking employment for the past 12 months. The data were obtained from the data-
base of the Applied Economic Research Institute (IPEA) from the Brazilian Federal
Government
3
and covers a period of 179 months (January 1991 through November
2005). The maximum and minimum values are 0.057 and 0.024, respectively, and the
average unemployment rate equals 0.044. A time series plot of the data is given in
Fig. 1.
We shall consider four AR models (p = 1, . . . , 4); see Table 1. The link func-
tion is logit and model selection is carried out using the AIC (Akaike information
criterion) and the BIC (Bayesian information criterion):
AIC =2

+2p and BIC =2

+plog(n),
where

denotes the log-likelihood function evaluated at the maximum likelihood
estimates, p is the number of autoregressive parameters, and n is the sample size.
3
See http://www.ipeadata.gov.br or obtain directly from http://beta.arma.googlepages.com/beta-arma-
data.txt.
Beta autoregressive moving average models 543
Fig. 1 Rate of hidden unemployment in So Paulo, Brazil
Table 1 AR models
Model 1
t
= +
1
g(y
t 1
)
Model 2
t
= +
1
g(y
t 1
) +
2
g(y
t 2
)
Model 3
t
= +
1
g(y
t 1
) +
2
g(y
t 2
) +
3
g(y
t 3
)
Model 4
t
= +
1
g(y
t 1
) +
2
g(y
t 2
) +
3
g(y
t 3
) +
4
g(y
t 4
)
Fig. 2 Sample autocorrelation functions of the standardized residuals obtained from Models 1 and 5
The AIC selects Model 4 whereas the BIC picks Model 1. We note that the co-
efcient of g(y
t 2
) (Model 4) is not statistically signicant at the usual signicance
levels, since the corresponding p-value equals 0.898. We thus consider a new model,
namely, the model with AR terms g(y
t 1
), g(y
t 3
), and g(y
t 4
) (Model 5). Figure 2
shows the residual correlograms corresponding to Models 1 and 5. It is clear that the
residuals from Model 1 are serially correlated, unlike the residuals obtained using
Model 5. We thus select Model 5 as the best model.
544 A.V. Rocha, F. Cribari-Neto
The estimated model is

t
=
exp{ +
1
g(y
t 1
) +
3
g(y
t 3
) +
4
g(y
t 4
)}
1 +exp{ +
1
g(y
t 1
) +
3
g(y
t 3
) +
4
g(y
t 4
)}
,
where (,
1
,
3
,
4
) = (0.16726, 1.18317, 0.57566, 0.33718), and the respec-
tive asymptotic standard errors obtained from the inverse of Fishers information
matrix are (0.0611, 0.0479, 0.0918, 0.0692).
As a nal step in the analysis, we turn to forecasting. We remove the nal six
observations from the series, t the model (Model 5) and produce six out-of-sample
forecasts. The observed values are 0.051, 0.052, 0.050, 0.049, 0.046, and 0.046, and
the corresponding forecasts are 0.052, 0.052, 0.051, 0.050, 0.050, and 0.049. The
AR forecasts are, overall, quite accurate.
6 Concluding remarks
In this paper we proposed a dynamic beta regression model: the ARMA model.
It can be used to model random variates that are continuous, assume values in the
standard unit interval (0, 1) and are observed over time. The proposed model is par-
ticularly useful for the time series modeling of rates and proportions. The model is
built upon the assumption that the conditional distribution of the variable of interest
given its past behavior is beta. As is well known, the beta distribution is very exible
for modeling data that are restricted to the standard unit interval, since the beta den-
sity can display quite different shapes depending on the values of the parameters that
index the distribution. Parameter estimation is performed by maximum likelihood,
and we derived closed-form expressions for the score function and Fishers informa-
tion matrix. Hypothesis testing inference can be carried out using standard asymptotic
tests. The proposed ARMA yields tted values and out-of-sample forecasts which
belong to the standard unit interval, unlike the standard ARMA model tted to rates
and proportion time series data.
Acknowledgements We thank two referees for their comments and suggestions. We also gratefully
acknowledge partial nancial support fromCoordenao de Aperfeioamento de Pessoal de Nvel Superior
(CAPES) and Conselho Nacional de Desenvolvimento Cientco e Tecnolgico (CNPq).
References
Akaike H (1973) Information theory and an extension of the maximum likelihood principle. In: Petroc BN,
Kaski F (eds) Second international symposium in information theory. Akademiai Kiado, Budapest,
pp 267281
Akaike H (1974) A new look at the statistical model identication. IEEE Trans Automat Control AC-
19:716723
Benjamin MA, Rigby RA, Stasinopoulos M (2003) Generalized autoregressive moving average models.
J Am Stat Assoc 98:214223
Choi B (1992) ARMA model identication. Springer, New York
Cribari-Neto F, Vasconcellos KLP (2002) Nearly unbiased maximum likelihood estimation for the beta
distribution. J Stat Comput Simul 72:107118
Beta autoregressive moving average models 545
Ferrari SLP, Cribari-Neto F (2004) Beta regression for modelling rates and proportions. J Appl Stat
31:799815
Fokianos K, Kedem B (2004) Partial likelihood for time series following generalized linear models. J Time
Ser Anal 25:173197
Li WK (1991) Testing model adequacy for some Markov regression models for time series. Biometrika
78:8389
Li WK (1994) Time series models based on generalized linear models: some further results. Biometrics
50:506511
McCullagh P, Nelder JA (1989) Generalized linear models, 2nd edn. Chapman and Hall, London
Nelder JA, Wedderburn RWM (1972) Generalized linear models. J R Stat Soc A 135:370384
Nocedal J, Wright SJ (1999) Numerical optimization. Springer, New York
Schwarz G (1978) Estimating the dimension of a model. Ann Stat 6:461464
Shephard N (1995). Generalized linear autoregressions. Technical report, Nufeld College, Oxford Uni-
versity. Manuscript available at http://www.nu.ox.ac.uk/economics/papers/1996/w8/glar.ps
Vasconcellos KLP, Cribari-Neto F (2005) Improved maximum likelihood estimation in a new class of beta
regression models. Braz J Probab Stat 19:1331
Zeger SL, Qaqish B (1988) Markov regression models for time series: a quasi-likelihood approach. Bio-
metrics 44:10191031

You might also like