Cointegrated VAR Models with Weak
Form of Common Cyclical Features
+
Carlos Enrique Carrasco Gutierrez
,
Reinaldo Castro Souza
)
Osmani Teixeira de Carvalho Guillén
¸
Abstract
An important aspect of empirical research based on the vector autoregres
sive (VAR) model is the choice of the lag order, since all inferences in this
model depend on the correct model speci…cation. There have been many
studies of how to select the lag order of a nonstationary VAR model subject
to cointegration restrictions. In this work, we consider in the model an addi
tional weak form (WF) restriction of common cyclical features to analyze the
appropriate way to select the correct lag order. We use two methodologies:
the traditional information criteria (AIC, HQ and SC) and an alternative cri
terion (IC(p; s)) that selects the lag order p and the rank structure s due to
the WF restriction. We use a MonteCarlo simulation in the analysis. The
results indicate that the cost of ignoring additional WF restrictions in vector
autoregressive modeling can be high, especially when the SC criterion is used.
Keywords: Cointegration; Common Cyclical Features; Reduced Rank Model;
Information Criteria.
JEL Codes: C32, C53.
We are grateful for the comments and suggestions by João Victor Issler, Wagner Gaglianone,
Ricardo Cavalcanti, Luiz Renato Lima and many seminar participants at the 2006 Brazilian Econo
metric Meeting and 2007 European Econometric Meeting in Hungary. We extend special thanks
to Alain Hecq for solving doubts and providing helpful comments. We are responsible for any
remaining errors in this paper. Carlos Enrique also acknowledges …nancial support from Getulio
Vargas Foundation and the O¢ce to Improve University Research (CAPES).
y
FUCAPE Business School, Av. Fernando Ferrari, 1358. Boa Vista,VitóriaES, Sala 12, CEP:
29075505. Email: carlos.gutierrez@fucape.br.
z
reinaldo@ele.pucrio.br, DEEPUCRJ, Brazil
x
osmani.guillen@bcb.gov.br, Central Bank of Brazil and IbmecRJ, Brazil.
1
1 Introduction
In the modeling of economic and …nancial time series, the vector autoregressive
(VAR) model has become the standard linear model used in empirical works. An
important aspect of empirical research on the speci…cation of VAR models is deter
mination of the lag order of the autoregressive lag polynomial, since all inferences
in the VAR model depend on the correct model speci…cation. Several works have
demonstrated the e¤ect of lag length selection. Lutkepohl (1993) indicated that
selecting a higher order lag length than the true one causes an increase in the
mean square forecast errors of the VAR and that under…tting the lag length often
generates autocorrelated errors. Braun and Mittnik (1993) showed that impulse
response functions and variance decompositions are inconsistently derived from the
estimated VAR when the lag length di¤ers from the true length. When cointegra
tion restrictions are considered in the model, the e¤ect of lag length selection on
the cointegration tests has been demonstrated. For example, Johansen (1991) and
Gonzalo (1994) pointed out that VAR order selection can a¤ect proper inference
about cointegrating vectors and rank.
Recently empirical works have considered other kinds of restrictions in the VAR
model (e.g., Engle and Issler, 1995; Caporale, 1997; Mamingi and Sunday, 2003). En
gle and Kozicki (1993) showed that VAR models can have other types of restrictions,
called common cyclical features, which are restrictions on the shortrun dynamics.
These restrictions are de…ned in the same way as cointegration restrictions, but while
cointegration refers to relations among variables in the longrun, common cyclical
restrictions refer to relations in the shortrun. Vahid and Engle (1993) proposed the
serial correlation common feature (SCCF) as a measure of common cyclical features.
SCCF restrictions might be imposed in a covariance stationary VAR model or in a
cointegrated VAR model. The concept of serial correlation common features appears
to be useful. It means that stationary time series move together in a way such that
there are linear combinations of these variables which yield white noise processes
and that their impulse response functions are collinear. In several practical applica
2
tions the existence of shortrun comovements between stationary time series (e.g.,
between …rst di¤erenced cointegrated I(1)) has been analyzed. For instance, Engle
and Issler (1995) found common cycle comovement in U.S. sectoral output data;
Hecq (2002) and Engle and Issler (1993) found common cycles in Latin American
countries; and Carrasco and Gomes (2009) found common international cycles in
GNP data for Mercosur countries.
When shortrun restrictions are imposed in cointegrated VAR models, it is pos
sible to de…ne a weak version of SCCF restrictions. Hecq, Palm and Urbain (2006)
de…ned a weak version of SCCF restrictions, which they denominated weakform
(WF) common cyclical restrictions. A fundamental di¤erence between SCCF and
WF restrictions is the form in which each one imposes restrictions on the represen
tation of the vector error correction model (VECM)
1
. When SCCF are imposed, all
matrices of a VECM have rank less than the number of variables analyzed. On the
other hand, with WF restrictions, all matrices except the longrun matrix have rank
less than the number of variables being analyzed. Hence, WF restrictions impose less
constraint on VECM parameters. Some advantages emerge when WF restrictions
are considered. First, due to the fact that the weak form common cyclical method
does not impose constraints on the cointegration space; the rank of common cyclical
features is not limited by the choice of cointegrating rank.
The literature has shown how to select an adequate lag order of a covariance
stationary VAR model and an adequate lag order of a VAR model subject to coin
tegration restrictions. Among the classical procedures are information criteria, such
as Akaike (AIC), Schwarz (SC) and HannanQuinn (HQ) (Lutkepohl, 1993). Kil
ian (2001) studied the performance of traditional AIC, SC and HQ criteria of a
covariance stationary VAR model. Vahid and Issler (2002) analyzed the standard
information criteria in a covariance stationary VAR model subject to SCCF re
striction and more recently Guillén, Issler and Athanasopoulos (2005) studied the
standard information criteria in VAR models with cointegration and SCCF restric
1
When a VAR model has cointegration restriction it can be represented as a VECM. This
representation is also known as Granger Representation Theorem (Engle and Granger 1987).
3
tions. However, when cointegrated VAR models contain an additional weak form
of common cyclical features, there are no reported works on how to appropriately
determine the VAR model order.
The objective of this paper is to investigate the performance of information
criteria in selecting the lag order of a VAR model when the data are generated
from a true VAR with cointegration and WF restrictions, referred to as the correct
model. We carry out the following two procedures: a) the use of standard criteria, as
proposed by Vahid and Engle (1993), referred to here as IC(j), and /) the use of an
alternative model selection criterion (see Vahid and Issler, 2002 and Hecq, Palm and
Urbain, 2006), consisting of simultaneously selecting the lag order p and the number
of weak forms of common cyclical features, :, which is referred to as IC(j. :)
2
. The
most relevant results can be summarized as follows. The information criterion that
selects the pair (j. :) performs better than the model chosen by conventional criteria,
especially the AIC(j. :) criterion. The cost of ignoring additional WF restrictions in
vector autoregressive modeling can be high, particularly when the SC(j) criterion
is used.
The rest of this work is organized as follows. Section 2 presents the economet
ric model. Section 3 discusses the information criteria. Section 4 shows a Monte
Carlo simulation and Section 5 presents the results. Finally, Section 6 contains our
conclusions.
2 The Econometric Model
We show the VAR model with shortrun and longrun restrictions. First, we consider
a Gaussian vector autoregression of …nite order j, called VAR(j), such that:
¸
t
=
p
i=1
¹
i
¸
ti
÷
t
(1)
where, ¸
t
is a vector of : …rst order integrated series, 1(1), ¹
i
, i = 1. . . . . j are
matrices of dimension : :,
t
~ `o::c (0. !) . 1 (
t
) = 0 and 1 (
t
0
) ={! , if
2
This is quite recent in the literature (see, Hecq et al., 2006).
4
t = t and 0
nn
, if t ,= t, where ! is nonsingular}. The model (1) can be written
equivalently as; H(1) ¸
t
=
t
where 1 represents the lag operator and H(1) =
1
n
÷
p
i=1
¹
i
1
i
such that when 1 = 1, H(1) = 1
n
÷
p
i=1
¹
i
. If cointegration is
considered in (1) the (: :) matrix H() satis…es two conditions: a) rank (H(1)) =
:, 0 < : < :, such that H(1) can be expressed as H(1) = ÷c,
0
, where c and ,
are (: :) matrices with full column rank, :; and b) the characteristic equation
[H(1)[ = 0 has :÷: roots equal to 1 and all other are outside the unit circle. These
assumptions imply that ¸
t
is cointegrated of order (1. 1). The elements of c are the
adjustment coe¢cients and the columns of , span the space of cointegration vectors.
We can represent a VAR model as a Vector Error Correction Model (VECM). By
decomposing the polynomial matrix H(1) = H(1) 1÷H
(1) ^, where ^ = (1 ÷1)
is the di¤erence operator, a VECM is obtained:
^¸
t
= c,
0
¸
t1
÷
p1
i=1
I
i
^¸
ti
÷
t
(2)
where: c,
0
= ÷H(1), I
j
= ÷
p
k=j+1
¹
k
for , = 1. ..... j ÷ 1 and I
0
= 1
n
. The
VAR(j) model can include additional shorthorizon restrictions as shown by Vahid
and Engle (1993). We consider an interesting WF restriction (as de…ned by Hecq,
Palm and Urbain, 2006) that does not impose constraints on longrun relations.
De…nition 1 The weak form (WF) holds in (2) if, in addition to cointegration
restriction, there exists an (: :) matrix
, of rank :, whose columns span the
cofeature space, such that
,
0
(^¸
t
÷ c,´¸
t1
) =
,
0
t
. where
,
0
t
is an sdimensional
vector that constitutes an innovation process with respect to information prior to
period t, given by ¦¸
t1
. ¸
t2
. .... ¸
1
¦ .
Equivalent to de…nition 1, we consider WF restrictions in the VECM if there
exists a cofeature matrix
, that satis…es the following assumption:
Assumption 1 :
,
0
I
j
= 0
sn
for , = 1. ..... j ÷ 1.
Therefore this is a naturally weaker alternative assumption which implies that
the common cyclical part is reduced to white noise by taking a linear combination of
5
the variables in the …rst di¤erences adjusted for longrun e¤ects. Imposing WF re
strictions is convenient because it allows studying of both cointegration and common
cyclical feature without the constraint
3
: ÷ : _ :.
We can rewrite the VECM with WF restrictions as a model of reducedrank
structure. In (2) let A
t1
= [^¸
0
t1
. .....^¸
0
tp+1
[
0
and 4 = [I
1
. ..... I
p1
[. Therefore,
we obtain:
^¸
t
= c,´¸
t1
÷ 4A
t1
÷
t
(3)
If assumption (1) holds, then matrices I
i
. i = 1. .... j are all of rank (: ÷ :) and we
can write 4 =
,
?
d =
,
?
[d
1
. ..... d
p1
[, where,
,
?
is : (: ÷ :) full column rank
matrix, d has dimension (: ÷ :) :(j ÷ 1). and the matrices d
i
. i = 1. .... j ÷ 1 all
have rank (: ÷ :) :. Hence, given assumption (1), there exists
, of : : such
that
,
0
,
?
= 0. That is,
,
?
: (: ÷ :) is a full column rank orthogonal to the
complement of
, with :c:/(
,.
,
?
) = :. Rewriting model (3) we have:
^¸
t
= c,´¸
t1
÷
,
?
(d
1
. d
2
. .... d
p1
) A
t1
÷
t
(4)
= c,´¸
t1
÷
,
?
dA
t1
÷
t
(5)
Estimation of (5) is carried out via the switching algorithms (see Centoni et al.,
2007; Hecq, 2006) that use the procedure in estimating reducedrank regression
models as suggested by Anderson (1951). There is a formal connection between a
reducedrank regression and the canonical analysis, as noted by Izenman (1975),
Box and Tiao (1977), Tso (1981) and Velu Reinsel and Wichern (1986). When all
the matrix coe¢cients of the multivariate regression are full rank, it can be esti
mated by the usual least squares or maximum likelihood procedures. But when
the matrix coe¢cients are of reducedrank they have to be estimated using the
reducedrank regression models of Anderson (1951). The use of canonical analysis
may be regarded as a special case of reducedrank regression. More speci…cally, the
maximumlikelihood estimation of the parameters of the reducedrank regression
3
Since the SCCF also imposes constraints on the longrun matrix
0
= ÷(1), which has
dimension n, the cointegration restrictions, r, and SCCF restrictions, s, must satisfy r + s _ n.
6
model may result in solving a problem of canonical analysis
4
. Therefore, we can use
the expression Cc:Co::¦A
t
. 2
t
[A
t1
¦ which denotes the partial canonical correla
tions between A
t
and 2
t
: both sets concentrate out the e¤ect of A
t1
allowing us
to obtain canonical correlation (see Johansen, 1995), represented by the eigenvalues
´
`
1
´
`
2
´
`
3
.......
´
`
n
. The Johansen test statistic is based on canonical cor
relation. In model (2) we can use the expression Cc:Co::¦^¸
t
. ¸
t1
[A
t1
¦ where
A
t1
= [^¸
0
t1
. .....^¸
0
tp+1
[
0
which summarizes the reducedrank regression proce
dure used in the Johansen approach. This means that we extracts the canonical
correlations between ^¸
t
and ¸
t1
: both sets concentrate out the e¤ect of lags of
A
t1
. In order to test for the signi…cance of the : largest eigenvalues, we can rely
on Johansen’s trace statistic (6):
¸
r
= ÷1
n
i=r+1
1:(1 ÷
´
`
2
i
) i = 1. .... : (6)
where the eigenvalues 0 <
´
`
n
< ... <
´
`
1
are the solution of : [`:
11
÷:
1
10
:
00
:
01
[ =
0, where :
ij
. i. , = 0.1. are the second moment matrices: :
00
=
1
T
T
t=1
n
0t
n
0
0t
,
:
10
=
1
T
T
t=1
n
1t
n
0
0t
, :
01
=
1
T
T
t=1
n
0t
n
0
1t
, :
11
=
1
T
T
t=1
n
1t
n
0
1t
of the residuals n
0t
and n
1t
obtained in the multivariate least squares regressions ^¸
t
= (^¸
t1
,...,^¸
tp+1
)÷
n
0t
and ¸
t1
= (^¸
t1
. ...^¸
tp+1
)÷n
1t
respectively (see, Hecq et al., 2006; Johansen,
1995). The result of the Johansen test is a superconsistent estimation of ,. More
over, we could also use a canonical correlation approach to determine the rank of
the common features space due to WF restrictions. This is a test for the existence of
cofeatures in the form of linear combinations of the variables in the …rst di¤erences,
corrected for longrun e¤ects which are white noise (i.e.,
,
0
(^¸
t
÷ c,´¸
t1
) =
,
0
t
where
,
0
t
is a white noise). We use canonical analysis is this work to estimate, and
select the lagrank of VAR models, as shown in next sections.
4
This estimation is referred to as full information maximum likelihood  FIML. (see Johansen,
1995).
7
3 Model Selection Criteria
In model selection we use two procedures to identify the VAR model order: the
standard selection criteria, IC(j), and the modi…ed informational criteria, IC(j. :),
a novelty in the literature, which consists of identifying j and : simultaneously.
The model estimation following the standard selection criteria, IC(j), originally
used by Vahid and Engle (1993) entails the following steps:
1. Estimate j using standard informational criteria: Akaike (AIC),
Schwarz (SC) and HannaQuinn (HQ). We chose the lag length of the
VAR in levels that minimizes the information criteria.
2. Using the lag length chosen in the previous step, …nd the number of
cointegration vectors, :. using the Johansen cointegration test
5
.
3. Conditional on the results of the cointegration analysis, estimate a
…nal VECM then calculate the multistep ahead forecast.
The above procedure is followed when there is evidence of cointegration restric
tions. We check the performance of IC(j) when WF restrictions are imposed on the
true model. Additionally we check the performance of IC(j. :) alternative selection
criteria. Vahid and Issler (2002) analyzed a covariancestationary VAR model with
SCCF restrictions. They showed that the use of IC(j. :) performs better than IC(j)
in VAR model lag order selection. In the present work we analyze the cointegrated
VAR model with WF restrictions in order to analyze the performance of IC(j) and
IC(j. :) for model selection. The question investigated is: Does IC(j. :) perform
better than IC(j)? This is an important question we aim to answer in this work.
The procedure to choose the lag order and the rank of the structure of shortrun
restrictions is carried out by minimizing the following modi…ed information criteria
(see; Vahid and Issler, 2002; Hecq, 2006).
¹1C (j. :) =
T
i=ns+1
ln(1 ÷ `
2
i
(j)) ÷
2
1
N (7)
5
Cointegration rank and vectors are estimated using the FIML, as shown in Johansen (1991).
8
HQ(j. :) =
T
i=ns+1
ln(1 ÷ `
2
i
(j)) ÷
2 ln(ln 1)
1
N (8)
oC(j. :) =
T
i=ns+1
ln(1 ÷ `
2
i
(j)) ÷
ln 1
1
N (9)
N = [: (: (j ÷ 1)) ÷ : :[ ÷ [: (: (j ÷ 1) ÷ (: ÷ :))[
where : is the number of variables in model (2) and ` is a number of parameters.
` is obtained by subtracting the total number of mean parameters in the VECM
(i.e., :
2
(j÷1)÷::), for given : and j, from the number of restrictions the common
dynamics imposes from :(:(j÷1))÷:(:÷:). The eigenvalues `
i
are calculated
for each j. In order to calculate the pair (j. :) we assume that no restriction exists,
that is, : = : (see Hecq, 2006). We …x j in model (3) and then …nd `
i
i = 1. 2...: by
computing the cc:co::(^¸
t
. A
t1
[ ¸
t1
). This procedure is followed for every j and
in the end we choose the j and : that minimize the IC(j. :). After selecting the pair
(j. :) we can test the cointegration relation using the procedure of Johansen. Finally
we estimate the model using the switching algorithms as shown in the next section.
Notice that in this simultaneous selection, testing the cointegration relation is the
last procedure followed, so we are inverting the hierarchical procedure followed by
Vahid and Engle (1993) where the …rst step is to select the number of cointegration
relations. This may be an advantage, especially when : is overestimated. Few works
have analyzed the order of VAR models considering modi…ed IC(j. :). As mentioned,
Vahid and Issler (2002) suggested the use of IC(j. :) to simultaneously choose the
order j and a number of reduced rank structure : in a covariance stationary VAR
model subject to SCCF restrictions. However, no work has analyzed the order of
the VAR model with cointegration and WF restrictions using a modi…ed criterion,
which is exactly the contribution of this paper.
To estimate the VAR model considering cointegration and WF restrictions we
use the switching algorithms model as considered by Hecq (2006). Consider the
9
VECM given by:
^¸
t
= c,
0
¸
t1
÷
,
?
dA
t1
÷
t
(10)
A full description of switching algorithms is presented below in four steps:
otcj1 : Estimation of the cointegration vectors ,.
Using the optimal pair (¯ j. ¯ :) chosen by the information criteria (7), (8)
or (9), we estimate , (and so its rank, : = ¯ :) using the Johansen coin
tegration test.
otcj2 : Estimation of
,
?
and d.
Taking the estimate of , in step one, we proceed to estimate
,
?
and d.
Hence, we run a regression of ^¸
t
and of A
t1
on
´
,
0
¸
t1
. We label the
residuals n
0
and n
1
, respectively. Therefore, we obtain a reduced rank
regression:
n
0t
=
,
?
dn
1t
÷
t
(11)
where d can be written as d =
_
C
1
. .... C
( p1)
_
of (: ÷ ¯ :) :(¯ j ÷1) and
,
?
of : (: ÷ ¯ :). We estimate (11) by FIML. Thus, we can obtain
,
?
and
´
d.
otcj3 : Estimation of the maximum likelihood (ML) function.
Given the parameters estimated in steps 1 and 2 we use a recursive
algorithm to estimate the maximum likelihood (ML) function. We cal
culate the eigenvalues associated with
´
d,
´
`
2
i
i = 1. .... ¯ : and the matrix
of residuals
max
r; s= s
. Hence, we compute the ML function:
1
0
max; r<n; s= s
= ÷
1
2
_
ln
¸
¸
¸
¸
¸
max
r<n; s= s
¸
¸
¸
¸
¸
÷
s
i=1
ln
_
1 ÷
´
`
2
i
_
_
(12)
If ¯ : = :, instead of (12) we use the derived loglikelihood: 1
max; r=n; s= s
=
÷
T
2
ln
¸
¸
¸
max
r=n; s= s
¸
¸
¸. The determinant of the covariance matrix for ¯ : = :
cointegration vector is calculated by
ln
¸
¸
¸
¸
¸
max
r=n; s= s
¸
¸
¸
¸
¸
= ln
¸
¸
:
00
÷ :
01
:
1
11
:
10
¸
¸
÷
s
i=1
ln
_
1 ÷
´
`
2
i
_
(13)
10
where :
ij
refers to cross moment matrices obtained in multivariate least
square regressions from ^¸
t
and A
t1
on ¸
t1
. In this case, estimation
does not entail using an iterative algorithm yet because the cointegrating
space spans 1
n
.
otcj4 : Reestimation of ,.
We reestimate , to obtain a more appropriate value for the parameters.
In order to reestimate , we compute the Cc:Co::
_
^¸
t
. ¸
t1
[
´
dA
t1
_
and thus using the new
´
, we can repeat step 2 to reestimate
,
?
and
d. Then, we calculate the new value of the ML function in the step
3. Hence, we obtain 1
1
max; r= r; s= s
to calculate ^1 = (1
1
max; r= r; s= s

1
0
max; r= r; s= s
).
We repeat steps 1 to 4 to choose
,
?
and d until convergence is reached ( i.e.,
^1 < 10
7
). In the end, the optimal parameters ¯ j, ¯ : and ¯ : are obtained and they
can be used to estimate and forecast of a VECM with WF restrictions.
4 MonteCarlo Design
The simple real business cycle models and also the simplest closed economy monetary
dynamic stochastic general equilibriummodels are threedimensional. Consumption,
saving and output and prices, output and money are notable examples. Motivated
by these applications and according the previous work of Vahid and Issler (2002) we
construct a MonteCarlo experiment in a threedimensional environment. Therefore,
the data generating processes considering a VAR model with three variables, one
cointegration vector, and two cofeatures vectors (i.e., : = 3, : = 1 and : = 2,
11
respectively). , and
, satisfy:
, =
_
_
1.0
0.2
÷1.0
_
_
.
, =
_
_
1.0 0.1
0.0 1.0
0.ò ÷0.ò
_
_
_
_
1t
2t
3t
_
_
s `
_
_
_
_
0
0
0
_
_
.
_
_
1.0 0.0 0.0
0.0 1.0 0.0
0.0 0.0 1.0
_
_
_
_
Consider the VAR(3) model: ¸
t
= ¹
1
¸
t1
÷ ¹
2
¸
t2
÷ ¹
3
¸
t3
÷
t
. The VECM
representation as a function of the VAR level parameters can be written as:
^¸
t
= (¹
1
÷ ¹
2
÷ ¹
3
÷ 1
3
)¸
t1
÷ (¹
2
÷ ¹
3
)^¸
t1
÷ ¹
3
^¸
t2
÷
t
(14)
The VAR coe¢cients must simultaneously obey the restrictions: a) The cointe
gration restrictions: c,
0
= (¹
1
÷ ¹
2
÷ ¹
3
÷ 1
3
) ; b) WF restrictions:
,
0
¹
3
= 0
(iii)
,
0
(¹
2
÷ ¹
3
) = 0 and c) the covariancestationary condition. Considering the
cointegration restrictions, we can rewrite (14) as the following VAR(1):
¸
t
= 1 ¸
t1
÷ ·
t
(15)
¸
t
=
_
_
´¸
t
´¸
t1
,
0
¸
t
_
_
. 1 =
_
_
÷(¹
2
÷ ¹
3
) ÷¹
3
c
1
3
0 0
÷,(¹
2
÷ ¹
3
) ÷,
0
¹
3
,
0
c ÷ 1
_
_
and ·
t
=
_
_
t
0
,
0
t
_
_
Thus, equation (15) will be covariancestationary if all eigenvalues of matrix
1 lie inside the unit circle. An initial idea to design the MonteCarlo experiment
can consist of constructing the companion matrix (1) and verifying whether the
eigenvalues of the companion matrix all lie inside the unit circle. This can be
carried out by selecting their values from a uniform distribution, and then verifying
whether or not the eigenvalues of the companion matrix all lie inside the unit circle.
However, this strategy could lead to a wide spectrum to search for adequate values
for the companion matrix. Hence, we follow an alternative procedure. We propose
an analytical solution to generate a covariancestationary VAR, based on the choice
of the eigenvalues, and then on the generation of the respective companion matrix.
In the appendix we present a detailed discussion of the …nal choice of these free
parameters, including analytical solutions. In our simulation, we constructed 100
12
data generating processes and for each of these we generate 1000 samples containing
1000 observations. To reduce the impact of the initial values, we considered only the
last 100 and 200 observations. All the experiments were conducted in the MatLab
environment.
5 Results
Figure 1 shows one example of the threedimensional VAR model with cointegration
and WF restrictions for 100 and 200 observations.
0 1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 0 9 0 1 0 0
 1 0 0
 9 0
 8 0
 7 0
 6 0
 5 0
 4 0
 3 0
 2 0
T im e
0 2 0 4 0 6 0 8 0 1 0 0 1 2 0 1 4 0 1 6 0 1 8 0 2 0 0
 1 0 0
 8 0
 6 0
 4 0
 2 0
0
T im e
Figure 1. One example of a VAR(3) model with n=3,r=1 and s=2 for 100 and 200 observations
The values in Table 1 represent the percentage of time that the model selection
criterion, IC(j), chooses the cell corresponding to the lag and number of cointegra
tion vectors in 100,000 runs. The true lagcointegrating vectors are identi…ed by
boldface numbers and the selected lagcointegration vectors chosen the most times
by the criterion are underlined. In Table 1, the results show that, in general, the
AIC most often chooses the correct lag length for 100 and 200 observations. For
13
example, for 100 observations, the AIC, HQ and SC criteria chose the true lag, j,
54.08%, 35.62% and 17.48% of the times respectively. Note that all three criteria
chose the correct rank of cointegration (: = 1). When 200 observations was consid
ered, the correct lag length was chosen 74.72%, 57.75% and 35.28% of the times for
AIC, HQ and SC respectively. Again, all three criteria selected the true cointegrated
rank : = 1. Table 2 contains the percentage that the alternative model selection
criterion, IC(j. :), chooses that cell, corresponding to the lagrank and number of
cointegrating vectors in 100,000 runs. The true lagrankcointegration vectors are
identi…ed by boldface numbers and the best lagrank combination chosen the most
times by each criterion are underlined. In Table 2, the results show that, in general,
the AIC(j. :) criterion more frequently chooses the lagrank for 100 and 200 obser
vations. For instance, for 100 observations, the AIC(j. :), HQ(j. :) and SC(j. :)
criteria more often choose the true pair (j. :) = (3. 1), 56.34%, 40.85% and 25.2% of
the times, respectively. For 200 observations, the AIC(j. :), HQ(j. :) and SC(j. :)
criteria more frequently choose the true pair (j. :) = (3. 1), 77.06%, 62.58% and
45.03% of the times respectively. Note that all three criteria more often choose
the correct rank of cointegration (: = 1) in both samples. What happened when
the weak form common cyclical restrictions are ignored? Tables 1 and 2 also show
the relative performance of IC(j. :) visàvis IC(j). For instance, for 1 = 100 the
SC(j. :) has a success rate of 25.2% in selecting the true j = 3, while the SC(j)
only has a success rate of 17.48%. This represents a gain of more than 44%. For
1 = 200, the gains are more than 27%. For 1 = 100 the HQ(j. :) selects the true
j =3 with 40.85% accuracy while the HQ(j) only has a success rate of 35.62%. This
represents a gain of 14%. For 1 = 200, the gains are more than 8%. For 1 = 100
the AIC(j. :) has a success rate of 56.34% in choosing the true j = 3, in comparison
with a rate of 54.08% for the AIC(j), a gain of more than 4%. For T=200, the gains
are more than 3%. Thus, it appears that when using the AIC(p,s) criteria the cost
of ignoring the weak form common cyclical restriction is low.
The most relevant results can be summarized as follows:
14
÷ All criteria (AIC, HQ and SC) choose the correct parameters more
often when using IC(j. :) visàvis IC(j).
÷ There is a cost of ignoring additional weak form common cyclical re
strictions in the model especially when the SC(j) criterion is used. In
general, the standard Schwarz, SC(j), or HannanQuinn, HQ(j), selec
tion criteria should not be used for this purpose in small samples due to
the tendency to identify an underparameterized model.
÷ The AIC performs better in selecting the true model more frequently
for both the IC(j. :) and the IC(j) criteria.
6 Conclusions
In this work, we considered an additional weak form restriction of common cycli
cal features in a cointegrated VAR model in order to analyze the appropriate way
to choose the correct lag order. These additional WF restrictions are de…ned in
the same way as cointegration restrictions. While cointegration refers to relations
among variables in the longrun, the common cyclical restrictions refer to relations
in the shortrun. Two methodologies have been used for selecting lag length; the
traditional information criterion, IC(j), and an alternative criterion (IC(j. :)) that
selects the lag order j and the rank structure : due to the weak form common
cyclical restrictions.
The results indicate that the information criterion that selects the lag length
and the rank order performs better than the model chosen by conventional criteria.
When the WF restrictions are ignored there is a non trivial cost in selecting the
true model with standard information criteria. In general, the standard Schwarz
or HannanQuinn selection criteria should not be used for this purpose in small
samples, due to the tendency to identify an underparameterized model.
In applied work, when the VAR model contains WF and cointegration restric
tions, we suggest the use of AIC(j. :) criteria to choose the lagrank, since it provides
15
considerable gains in selecting the correct VAR model. Since no work in the litera
ture has analyzed a VAR model with WF common cyclical restrictions, the results of
this work provide new insights and incentives to proceed with this kind of empirical
work.
References
Anderson, T. W, (1951) Estimating linear restrictions on regression coe¢cients
for multivariate normal distributions. Ann. Math. Statist. 22 327351. [Correction
(1980) Ann. Statist. 8 1400].
Box GE, Tiao GC. (1977). A canonical analysis of multiple time series. Bio
metrika, 64:355365.
Braun PA, Mittnik S. (1993). Misspeci…cations in Vector Autoregressions and
Their E¤ects on Impulse Responses and Variance Decompositions. Journal of Econo
metrics, 59, 31941.
Caporale GM. (1997). Common Features and Output Fluctuations in the United
Kingdom. EconomicModelling, 14: 19.
Centoni M, Cubadda G, Hecq. (2007) Common Shocks, Commo Dynamics, and
the International Business Cycle. EconomicModelling, 24, 149166.
Carrasco CG, Gomes F. (2009). Evidence on Common Features and Business
Cycle Synchronization in Mercosur. Brazilian Review of Econometrics. Forthcom
ing.
Engle R, Granger C. (1987). Cointegration and error correction: representation,
estimation and testing. Econometrica, , 55, 25176.
Engle R, Kozicki S.(1993). Testing for common features. Journal of Business
and Economic Statistics, 11: 369395.
16
Engle R, Issler JV. (1995). Estimating Common Sectoral Cycles. Journal of
Monetary Economics, 35:83113.
Engle R, Issler JV.(1993). Common Trends and Common Cycles in Latin Amer
ica. Revista Brasileira de Economia, vol. 47, n
o
2, pp. 149176, 1993.
Gonzalo, J. (1994). Five Alternative Methods of Estimating LongRun Equilib
rium Relationships. Journal of Econometrics, 60 issue 12, 203233.
Guillén, Issler, Athanasopoulos. (2005). Forecasting Accuracy and Estimation
Uncertainty using VAR Models with Short and LongTerm Economic Restrictions:
A MonteCarlo Study. Ensaios Econômicos EPGE, 589.
Hecq A. (2006). Cointegration and Common Cyclical Features in VAR Models:
Comparing Small Sample Performances of the 2Step and Iterative Approaches.
Mimeo.
Hecq A, Palm FC, Urbain JP. (2006). Testing for Common Cyclical Features in
VAR Models with Cointegration, Journal of Econometrics, 132: 117141.
Hecq, A. (2002), Common Cycles and Common Trends in Latin America. Medium
Econometrische Toepassingen, 10, 2025.
Izenman AJ. (1975). Reduced rank regression for the multivariate linear model,
Journal of. multivariate Analysis, 5: 248264.
Johansen, S. (1991). Estimation and Hypothesis Testing of Cointegration Vectors
in Gaussian Vector Autoregressive Models. Econometrica, 59, 15511580.
Johansen S. (1995). Likelihoodbased Inference in Cointegrated Vector Autore
gressive Models. Oxford Univ. Press. Z.
Kilian L. (2001). Impulse response analysis in vector autoregressions with un
known lag order. Journal of Forecasting, 20: 161179.
17
Lütkepohl H. (1993). Introduction to Multiple Time Series Analysis. Springer,
New York.
Mamingi N, Iyare SO. (2003). Convergence and Common Features in Interna
tional Output: A Case Study of the Economic Community of West African States,
19751997. AsianAfricanJournalofEconomicsandEconometrics, 3: 116.
Tso MK. (1981). Reduced rank regression and Canonical analysis. Journal of
the Royal Statistical Society, 43: 183189.
Vahid F, Engle R. (1993). Common trends and common cycles. Journal of
Applied Econometrics, 8: 341360.
Vahid F, Issler JV. (2002). The Importance of Common Cyclical Features in
VAR Analysis: A Monte Carlo Study. Journal of Econometrics, 109: 341363.
Velu RP, Reinsel GC, Wichern DW. (1986). Reduced rank models for multiple
times series. Biometrika, 73: 105118.
18
Appendix A: Tables
0 1 2 3 0 1 2 3
Selected Lag
1 0.000 0.996 0.359 0.031 0.000 0.095 0.016 0.003
2 0.002 32.146 1.136 0.048 0.000 17.073 0.686 0.033
3 2.792 54.082 0.902 0.041 0.012 74.721 1.488 0.108
AIC(p) 4 0.737 4.068 0.091 0.003 0.005 4.177 0.081 0.006
5 0.392 0.987 0.031 0.000 0.013 0.828 0.020 0.000
6 0.219 0.333 0.014 0.000 0.023 0.257 0.005 0.000
7 0.166 0.173 0.006 0.000 0.039 0.133 0.002 0.000
8 0.133 0.107 0.005 0.000 0.060 0.115 0.001 0.000
1 0.000 3.884 1.915 0.165 0.000 1.098 0.243 0.021
2 0.002 52.593 1.907 0.080 0.000 37.390 1.614 0.098
3 2.600 35.617 0.612 0.027 0.012 57.749 1.146 0.082
HQ(p) 4 0.065 0.189 0.007 0.000 0.001 0.158 0.004 0.000
5 0.059 0.037 0.000 0.000 0.009 0.082 0.001 0.000
6 0.073 0.025 0.000 0.000 0.016 0.076 0.000 0.000
7 0.059 0.019 0.001 0.000 0.030 0.070 0.000 0.000
8 0.053 0.011 0.000 0.000 0.044 0.055 0.001 0.000
1 0.000 8.344 6.609 0.511 0.000 3.964 1.385 0.093
2 0.003 61.966 2.279 0.105 0.000 55.156 2.776 0.169
3 2.042 17.485 0.313 0.015 0.012 35.283 0.728 0.044
SC(p) 4 0.049 0.045 0.000 0.000 0.001 0.083 0.002 0.000
5 0.071 0.025 0.000 0.000 0.007 0.076 0.001 0.000
6 0.057 0.016 0.000 0.000 0.013 0.063 0.000 0.000
7 0.036 0.009 0.000 0.000 0.025 0.056 0.000 0.000
8 0.017 0.003 0.000 0.000 0.027 0.035 0.001 0.000
Numbers represent the percentage times that the model selection criterion choice that cell corresponding to the lag and
number of cointegration vectors in 100,000 realizations. The true lagcointegrating vectors are indentified by bold numbers.
Table I Performance of information criterion, IC(p ), in selecting the lag order p
Number of observations = 100
Selected cointegrated vectors
Number of observations = 200
Selected cointegrated vectors
Frequency of lag(p) and cointegrating vectors (r) choice by different criteria for trivariate VAR model in
levels when the true model have parameters: p = 3 and r = 1.
19
Johansen Tested coint. Vectors (r) 0 1 2 3
Selected rank (s) 1 2 3 1 2 3 1 2 3 1 2 3
Selected lag (p)
Sample size = 100
1            
2 0.002 0.000 0.000 39.049 0.001 0.000 1.218 0.000 0.000 0.056 0.000 0.000
3 0.301 0.000 0.000 56.341 0.003 0.000 1.559 0.000 0.000 0.053 0.000 0.000
AIC(p ,s ) 4 0.004 0.000 0.000 1.186 0.001 0.000 0.070 0.000 0.000 0.001 0.000 0.000
5 0.000 0.000 0.000 0.114 0.001 0.000 0.012 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.020 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.006 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1            
2 0.002 0.000 0.000 55.563 0.000 0.000 1.888 0.000 0.000 0.081 0.000 0.000
3 0.267 0.000 0.000 40.855 0.000 0.000 1.207 0.000 0.000 0.043 0.000 0.000
HQ(p ,s ) 4 0.000 0.000 0.000 0.088 0.000 0.000 0.005 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1            
2 0.004 0.000 0.000 70.971 0.000 0.000 2.574 0.000 0.000 0.113 0.000 0.000
3 0.221 0.000 0.000 25.204 0.000 0.000 0.887 0.000 0.000 0.025 0.000 0.000
SC(p ,s ) 4 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Sample size = 200
1            
2 0.000 0.000 0.000 18.797 0.000 0.000 0.681 0.000 0.000 0.038 0.000 0.000
3 0.000 0.000 0.000 77.065 0.002 0.000 2.260 0.000 0.000 0.145 0.000 0.000
AIC(p ,s ) 4 0.000 0.000 0.000 0.908 0.000 0.000 0.035 0.000 0.000 0.001 0.000 0.000
5 0.000 0.000 0.000 0.063 0.000 0.000 0.002 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.003 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1            
2 0.000 0.000 0.000 33.952 0.000 0.000 1.370 0.000 0.000 0.086 0.000 0.000
3 0.000 0.000 0.000 62.576 0.000 0.000 1.877 0.000 0.000 0.111 0.000 0.000
HQ(p ,s ) 4 0.000 0.000 0.000 0.027 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1            
2 0.000 0.000 0.000 50.983 0.000 0.000 2.351 0.000 0.000 0.146 0.000 0.000
3 0.000 0.000 0.000 45.028 0.000 0.000 1.416 0.000 0.000 0.076 0.000 0.000
SC(p ,s ) 4 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
5 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
6 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
7 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
8 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
Table II. Performance of information criterion, IC(p,s), in selecting p and s simultaneously.
Numbers represent the percentage times that the simultaneous model selection criterion IC(p,s) choice that cell, corresponding to the lagrank and number of cointegrating vectors in
100,000 realizations. The true lagrankcointegration vectors are identified by bold numbers and the best lagrankcointegration vectors chosen by criterias are identified by underline
lines.
20
Appendix B: VAR Restrictions for the DGPs
Consider the vector autoregressive, VAR(3), model :
¸
t
= ¹
1
¸
t1
÷ ¹
2
¸
t2
÷ ¹
3
¸
t3
÷
t
(16)
with parameters: ¹
1
=
_
_
c
1
11
c
1
12
c
1
12
c
1
21
c
1
22
c
1
22
c
1
31
c
1
32
c
1
32
_
_
, ¹
2
=
_
_
c
2
11
c
2
12
c
2
12
c
2
21
c
2
22
c
2
22
c
2
31
c
2
32
c
2
32
_
_
and ¹
3
=
_
_
c
3
11
c
3
12
c
3
12
c
3
21
c
3
22
c
3
22
c
3
31
c
3
32
c
3
32
_
_
. We consider the cointegration vectors , =
_
_
,
11
,
21
,
31
_
_
, the cofea
tures vectors
, =
_
_
,
11
,
12
,
21
,
22
,
31
,
32
_
_
and the adjustament matrix c =
_
_
c
11
c
21
c
31
_
_
. The
longrun relation is de…ned by c,
0
= (¹
1
÷¹
2
÷¹
3
÷1
3
). Thus, the VECM respre
sentation is:
^¸
t
= c,
0
¸
t1
÷ (¹
2
÷ ¹
3
)^¸
t1
÷ ¹
3
^¸
t2
÷
t
(17)
We can rewrite equation (17) as a VAR(1):
¸
t
= 1 ¸
t1
÷ ·
t
(18)
where ¸
t
=
_
_
´¸
t
´¸
t1
,
0
¸
t
_
_
. 1 =
_
_
÷(¹
2
÷ ¹
3
) ÷¹
3
c
1
3
0 0
÷,(¹
2
÷ ¹
3
) ÷,
0
¹
3
,
0
c ÷ 1
_
_
and ·
t
=
_
_
t
0
,
0
t
_
_
1) Shortrun restrictions (WF)
We now impose the common cyclical restrictions (i) and (ii) on model (16).
Let, G = ÷[1
21
1 ÷ 1
31
[, 1 = [(1
32
÷ 1
31
),(1
21
÷ 1
22
)[, 1
j1
=
,
j1
,
,
11
, 1
j2
=
,
j2
,
,
12
(, = 2. 3) and o = ,
11
G ÷ ,
21
1 ÷ ,
31
(i)
,
0
¹
3
= 0 == ¹
3
=
_
_
÷Gc
3
31
÷Gc
3
32
÷Gc
3
33
÷1c
3
31
÷1c
3
32
÷1c
3
33
÷c
3
31
÷c
3
32
÷c
3
33
_
_
(ii)
,
0
(¹
2
÷ ¹
3
) = 0 ==
,
0
¹
2
= 0 == ¹
2
=
_
_
÷Gc
2
31
÷Gc
2
32
÷Gc
2
33
÷1c
2
31
÷1c
2
32
÷1c
2
33
÷c
2
31
÷c
2
32
÷c
2
33
_
_
2) Longrun restrictions (cointegration)
21
The cointegration restrictions are speci…ed by (iv) and (v):
(i·) ,
0
(¹
2
÷ ¹
3
) = [÷(c
2
31
÷ c
3
31
)o ÷ (c
2
32
÷ c
3
32
)o ÷ (c
2
33
÷ c
3
33
)o[ and
,
0
¹
3
= [÷c
3
31
o ÷ c
3
32
o ÷ c
3
33
o[
(·) ,
0
c÷1 = , =
_
,
11
,
21
,
31
¸
_
_
c
11
c
21
c
31
_
_
÷1 = ,
11
c
11
÷,
21
c
21
÷,
31
c
31
÷1
Taking into account the short and longrun restrictions, the companion matrix
1 can be represented as:
1 =
_
_
÷(¹
2
÷ ¹
3
) ÷¹
3
c
1
3
0 0
÷,(¹
2
÷ ¹
3
) ÷,
0
¹
3
,
0
c ÷ 1
_
_
=
_
¸
¸
¸
¸
¸
¸
¸
¸
_
÷G(c
2
31
÷ c
3
31
) ÷G(c
2
32
÷ c
3
32
) ÷G(c
2
33
÷ c
3
33
) ÷Gc
3
31
÷Gc
3
32
÷Gc
3
33
c
11
÷1(c
2
31
÷ c
3
31
) ÷G(c
2
32
÷ c
3
32
) ÷G(c
2
33
÷ c
3
33
) ÷1c
3
31
÷1c
3
32
÷1c
3
33
c
21
÷(c
2
31
÷ c
3
31
) ÷G(c
2
32
÷ c
3
32
) ÷(c
2
33
÷ c
3
33
) ÷c
3
31
÷c
3
32
÷c
3
33
c
31
1 0 0 0 0 0 0
0 1 0 0 0 0 0
0 0 1 0 0 0 0
÷(c
2
31
÷ c
3
31
)o ÷(c
2
32
÷ c
3
32
)o ÷(c
2
33
÷ c
3
33
)o ÷c
3
31
o ÷c
3
32
o ÷c
3
33
o /
_
¸
¸
¸
¸
¸
¸
¸
¸
_
with / = ,
0
c ÷ 1 = ,
11
c
11
÷ ,
21
c
21
÷ ,
31
c
31
÷ 1
3) Covariancestationary restrictions
Equation (18) will be covariancestationary if all eigenvalues of matrix 1 lie
inside the unit circle. That is, eigenvalue of matrix 1 is a number ` such that:
[1 ÷ `1
7
[ = 0 (19)
The solution of (19) is:
`
7
÷ !`
6
÷ O`
5
÷ d`
4
= 0 (20)
where the parameters !, O, and d are: ! = G(c
2
31
÷c
3
31
)÷1(c
2
32
÷c
3
32
)÷c
2
33
÷c
3
33
÷/,
O = Gc
3
31
÷1c
3
32
÷(c
2
33
÷c
3
33
)/ ÷G/(c
2
31
÷c
3
31
) ÷1/(c
2
32
÷c
3
32
) ÷c
31
o(c
2
33
÷c
3
33
) ÷
oc
21
(c
2
32
÷c
3
32
) ÷oc
11
(c
2
31
÷c
3
31
) ÷c
3
33
and d = ÷c
3
33
/ ÷Gc
3
31
/ ÷1c
3
32
/ ÷c
31
c
3
33
o ÷
c
3
32
oc
21
÷c
3
31
oc
11
, and the …rst four roots are `
1
= `
2
= `
3
= `
4
= 0.We calculated
22
the parameters of matrices ¹
1
, ¹
2
and ¹
3
as functions of roots (`
5
. `
6
and `
7
) and
free parameters. Hence we have three roots satisfying equation (20)
`
3
÷ !`
2
÷ O` ÷ d = 0 (21)
for `
5
, we have: `
3
5
÷ !`
2
5
÷ O`
5
÷ d = 0 ..................................1¡1
for `
6
, we have: `
3
6
÷ !`
2
6
÷ O`
6
÷ d = 0 ..................................1¡2
for `
7
, we have: `
3
7
÷ !`
2
7
÷ O`
7
÷ d = 0 ..................................1¡3
Solving equations 1, 2 and 3 yields: ! = ÷`
7
÷`
6
÷`
5
, O = `
6
`
7
÷`
6
`
5
÷`
5
`
7
and d = ÷`
5
`
6
`
7
. Equaling these parameters with the relations above we have:
c
2
31
= ÷(÷1c
2
32
÷1c
2
32
/ ÷c
31
oc
2
33
÷`
6
`
7
÷`
6
÷`
7
÷c
2
33
/ ÷`
5
`
6
`
7
÷/ ÷`
5
`
7
÷
`
5
`
6
÷ c
2
33
÷ oc
2
32
c
21
÷ `
5
),(oc
11
÷ G ÷ G/)
c
3
32
= (÷o
2
`
7
c
11
c
31
÷/
2
`
7
G÷`
6
G/
2
÷/`
7
oc
11
÷`
6
oc
11
/÷c
3
31
oc
11
G÷c
3
31
o
2
c
2
11
÷
Gc
3
31
/oc
11
÷`
5
G/
2
÷`
5
oc
11
/÷`
7
`
6
c
31
oG÷`
7
`
5
c
31
oG÷o
2
c
11
`
5
c
31
÷o
2
c
11
`
6
c
31
÷
o`
5
G/c
31
÷oc
31
`
6
G/÷`
5
`
7
`
6
G÷`
6
`
7
G/÷`
5
`
7
G/÷`
5
`
6
G/÷oG/
2
c
31
÷o
2
c
11
/c
31
÷
o
2
c
11
c
31
c
2
33
÷o
2
c
2
31
c
2
33
G÷oG
2
c
3
31
c
31
÷oc
11
c
2
33
/ ÷G/
3
÷oc
11
/
2
÷o
2
c
11
1c
2
32
c
31
÷
o
2
c
11
c
31
Gc
3
31
÷o
2
c
2
32
c
21
Gc
31
÷oc
2
32
c
21
G/÷oc
31
G
2
c
3
31
/÷oc
31
c
2
33
G/÷oc
11
1c
2
32
/÷
o`
7
G/c
31
÷`
5
`
6
c
31
oG÷`
5
`
7
`
6
c
31
oG÷`
5
`
7
`
6
oc
11
),(oc
11
1c
31
÷1Gc
31
÷/Gc
21
÷
1c
31
G/ ÷ oc
11
c
21
÷ Gc
21
),o
c
3
33
= ÷(1/
3
G÷`
5
G/
2
1÷oc
11
`
6
1`
7
`
5
÷1/`
7
oc
11
÷1/
2
`
7
G÷o
2
c
21
`
7
c
11
÷
`
6
G/oc
21
÷oc
21
`
7
G/÷`
6
G/
2
1÷`
6
oc
11
1/÷`
6
o
2
c
11
c
21
÷`
5
G/oc
21
÷`
5
oc
11
1/÷
`
5
o
2
c
11
c
21
÷`
7
`
6
oc
21
G÷1/`
7
`
6
G÷1/`
7
`
5
G÷1/`
5
`
6
G÷`
7
`
6
1G`
5
÷o
2
c
11
c
21
1c
2
32
÷
o
2
c
11
c
21
/÷o
2
c
11
c
21
c
2
33
÷o
2
c
2
21
c
2
32
G÷oc
11
1/
2
÷oc
21
G
2
c
3
31
÷oc
21
G/
2
÷o
2
c
3
31
1c
2
11
÷
o
2
c
11
c
21
Gc
3
31
÷o
2
c
21
c
2
33
Gc
31
÷oc
11
1´2/c
2
32
÷oc
11
1/c
2
33
÷oc
11
c
3
31
1G÷oc
11
1/Gc
3
31
÷
o1/c
2
33
Gc
31
÷oc
21
G
2
c
3
31
/÷oc
21
`
5
`
6
G÷oc
21
`
5
`
7
`
6
G÷oc
21
1c
2
32
G/÷oc
21
`
7
`
5
G),(oc
11
1c
31
÷
1Gc
31
÷ /Gc
21
÷ 1c
31
G/ ÷ oc
11
c
21
÷ Gc
21
),o
We can calculate c
2
31
, c
3
32
and c
3
33
…xing the set `
1
= `
2
= `
3
= `
4
= 0 and sort
independently from uniform distributions (÷0.9: 0.9) the values of c
3
31
. c
2
32
. c
2
33
. `
5
.
23
`
6
and `
7
. Hence, each parameter of the matrices ¹
1
, ¹
2
and ¹
3
are de…ned and we
can generate the DGPs of VAR(3) model with cointegration and WF restrictions.
24
1
Introduction
In the modeling of economic and …nancial time series, the vector autoregressive (VAR) model has become the standard linear model used in empirical works. An important aspect of empirical research on the speci…cation of VAR models is determination of the lag order of the autoregressive lag polynomial, since all inferences in the VAR model depend on the correct model speci…cation. Several works have demonstrated the e¤ect of lag length selection. Lutkepohl (1993) indicated that selecting a higher order lag length than the true one causes an increase in the mean square forecast errors of the VAR and that under…tting the lag length often generates autocorrelated errors. Braun and Mittnik (1993) showed that impulse response functions and variance decompositions are inconsistently derived from the estimated VAR when the lag length di¤ers from the true length. When cointegration restrictions are considered in the model, the e¤ect of lag length selection on the cointegration tests has been demonstrated. For example, Johansen (1991) and Gonzalo (1994) pointed out that VAR order selection can a¤ect proper inference about cointegrating vectors and rank. Recently empirical works have considered other kinds of restrictions in the VAR model (e.g., Engle and Issler, 1995; Caporale, 1997; Mamingi and Sunday, 2003). Engle and Kozicki (1993) showed that VAR models can have other types of restrictions, called common cyclical features, which are restrictions on the shortrun dynamics. These restrictions are de…ned in the same way as cointegration restrictions, but while cointegration refers to relations among variables in the longrun, common cyclical restrictions refer to relations in the shortrun. Vahid and Engle (1993) proposed the serial correlation common feature (SCCF) as a measure of common cyclical features. SCCF restrictions might be imposed in a covariance stationary VAR model or in a cointegrated VAR model. The concept of serial correlation common features appears to be useful. It means that stationary time series move together in a way such that there are linear combinations of these variables which yield white noise processes and that their impulse response functions are collinear. In several practical applica2
all matrices except the longrun matrix have rank less than the number of variables being analyzed. Palm and Urbain (2006) de…ned a weak version of SCCF restrictions. 1 3 . with WF restrictions. Hence. between …rst di¤erenced cointegrated I(1)) has been analyzed. which they denominated weakform (WF) common cyclical restrictions. Engle and Issler (1995) found common cycle comovement in U.g. When shortrun restrictions are imposed in cointegrated VAR models. This representation is also known as Granger Representation Theorem (Engle and Granger 1987). Among the classical procedures are information criteria. The literature has shown how to select an adequate lag order of a covariance stationary VAR model and an adequate lag order of a VAR model subject to cointegration restrictions. When SCCF are imposed. the rank of common cyclical features is not limited by the choice of cointegrating rank. Hecq. For instance. SC and HQ criteria of a covariance stationary VAR model.S. Schwarz (SC) and HannanQuinn (HQ) (Lutkepohl. such as Akaike (AIC). Vahid and Issler (2002) analyzed the standard information criteria in a covariance stationary VAR model subject to SCCF restriction and more recently Guillén. Some advantages emerge when WF restrictions are considered.tions the existence of shortrun comovements between stationary time series (e. 1993). Kilian (2001) studied the performance of traditional AIC.. Hecq (2002) and Engle and Issler (1993) found common cycles in Latin American countries. all matrices of a VECM have rank less than the number of variables analyzed. On the other hand. sectoral output data. WF restrictions impose less constraint on VECM parameters. and Carrasco and Gomes (2009) found common international cycles in GNP data for Mercosur countries. A fundamental di¤erence between SCCF and WF restrictions is the form in which each one imposes restrictions on the representation of the vector error correction model (VECM)1 . First. due to the fact that the weak form common cyclical method does not impose constraints on the cointegration space. Issler and Athanasopoulos (2005) studied the standard information criteria in VAR models with cointegration and SCCF restricWhen a VAR model has cointegration restriction it can be represented as a VECM. it is possible to de…ne a weak version of SCCF restrictions.
However. such that: yt = p X i=1 Ai y t i + "t (1) where. s. I(1). s) criterion. Section 2 presents the econometric model. : : : . ) . called VAR(p). consisting of simultaneously selecting the lag order p and the number of weak forms of common cyclical features. Hecq et al. 2006). Ai . The objective of this paper is to investigate the performance of information criteria in selecting the lag order of a VAR model when the data are generated from a true VAR with cointegration and WF restrictions. which is referred to as IC(p. We carry out the following two procedures: a) the use of standard criteria. if This is quite recent in the literature (see. 2006). particularly when the SC(p) criterion is used. Section 4 shows a Monte Carlo simulation and Section 5 presents the results. s) performs better than the model chosen by conventional criteria. The information criterion that selects the pair (p. 2 The Econometric Model We show the VAR model with shortrun and longrun restrictions.. referred to as the correct model. The cost of ignoring additional WF restrictions in vector autoregressive modeling can be high. we consider a Gaussian vector autoregression of …nite order p. referred to here as IC(p). Section 6 contains our conclusions. Palm and Urbain. yt is a vector of n …rst order integrated series. when cointegrated VAR models contain an additional weak form of common cyclical features. and b) the use of an alternative model selection criterion (see Vahid and Issler. s)2 . 4 . especially the AIC(p. Finally. 2002 and Hecq. Section 3 discusses the information criteria. The rest of this work is organized as follows. p are matrices of dimension n 2 n. i = 1. as proposed by Vahid and Engle (1993). The most relevant results can be summarized as follows.tions. E ("t ) = 0 and E ("t "0 ) ={ . there are no reported works on how to appropriately determine the VAR model order. "t N ormal (0. First.
1). De…nition 1 The weak form (WF) holds in (2) if. if t 6= . where ~ "t is an sdimensional vector that constitutes an innovation process with respect to information prior to period t. we consider WF restrictions in the VECM if there exists a cofeature matrix ~ that satis…es the following assumption: 0 Assumption 1 : ~ j = 0s n for j = 1. i=1 Ai . ::::. where (1 L) is the di¤erence operator. p 1. (L) yt = "t where L represents the lag operator and (L) = Pp Pp i In (1) = In i=1 Ai L such that when L = 1. r. ::::. such that are (n (1) can be expressed as . We consider an interesting WF restriction (as de…ned by Hecq. 0 < r < n. The k=j+1 Ak for j = 1. y1 g : Equivalent to de…nition 1. Palm and Urbain. :::. where is nonsingular}. whose columns span the 0 0 yt 1 ) = ~ "t . such that ~ ( yt Pp p 1 X i=1 i yt i + "t 1 and 0 (2) = In . yt 2 . a VECM is obtained: yt = where: 0 0 yt 1 + = (1). 2006) that does not impose constraints on longrun relations. and b) the characteristic equation r roots equal to 1 and all other are outside the unit circle. If cointegration is considered in (1) the (n n) matrix ( ) satis…es two conditions: a) rank ( (1)) = (1) = 0 r. where and r) matrices with full column rank. The model (1) can be written equivalently as. there exists an (n 0 cofeature space. We can represent a VAR model as a Vector Error Correction Model (VECM). p s) matrix ~ of rank s. Therefore this is a naturally weaker alternative assumption which implies that the common cyclical part is reduced to white noise by taking a linear combination of 5 . in addition to cointegration restriction. These are the j (L)j = 0 has n assumptions imply that yt is cointegrated of order (1. By decomposing the polynomial matrix (L) = (1) L + (L) . given by fyt 1 . j = VAR(p) model can include additional shorthorizon restrictions as shown by Vahid and Engle (1993). The elements of adjustment coe¢ cients and the columns of span the space of cointegration vectors.t= and 0n n.
and SCCF restrictions. i has dimension (n s) n(p 1). i = 1. given assumption (1). Therefore.the variables in the …rst di¤erences adjusted for longrun e¤ects. Hecq. the maximumlikelihood estimation of the parameters of the reducedrank regression 0 Since the SCCF also imposes constraints on the longrun matrix = (1).. :::. which has dimension n. The use of canonical analysis may be regarded as a special case of reducedrank regression. ::::. Hence. = ~? s) = ~ ?[ 1 . ::::: yt 0 p+1 ] and =[ 1 . More speci…cally. as noted by Izenman (1975). 1 2 . and the matrices = 1. 2007. p 1 all s such have rank (n n. it can be estimated by the usual least squares or maximum likelihood procedures. That is. there exists ~ of n (n 0 that ~ ~ ? = 0. i. :::. r. There is a formal connection between a reducedrank regression and the canonical analysis. ~ ? n s) is a full column rank orthogonal to the complement of ~ with rank( ~ . p 1 ]. the cointegration restrictions. then matrices can write matrix. 3 6 . p are all of rank (n (n where. In (2) let Xt we obtain: yt = yt 1 1 0 0 = [ yt 1 . 2006) that use the procedure in estimating reducedrank regression models as suggested by Anderson (1951). ::::. When all the matrix coe¢ cients of the multivariate regression are full rank. Tso (1981) and Velu Reinsel and Wichern (1986). ~ ? ) = n. :::. Rewriting model (3) we have: yt = = yt yt 1 1 + ~? ( 1. must satisfy r + s n. Imposing WF restrictions is convenient because it allows studying of both cointegration and common cyclical feature without the constraint3 r + s n. s. Box and Tiao (1977). We can rewrite the VECM with WF restrictions as a model of reducedrank structure. + Xt 1 + "t (3) s) and we If assumption (1) holds. ~ ? is n s) full column rank i. But when the matrix coe¢ cients are of reducedrank they have to be estimated using the reducedrank regression models of Anderson (1951). p 1 ) Xt 1 + "t (4) (5) + ~ ? Xt + "t Estimation of (5) is carried out via the switching algorithms (see Centoni et al. p 1 ].
1995). m01 = T T u0t u01t . represented by the eigenvalues ^ 1 > ^ 2 > ^ 3 ::::::: > ^ n . yt = ( yt 1 . Therefore. ::: yt p+1 )+u1t respectively (see. This means that we extracts the canonical correlations between yt and yt 1 : both sets concentrate out the e¤ect of lags of Xt 1 . t=1 ~ ~ p+1 )+ i=r+1 where the eigenvalues 0 < ^ n < ::: < ^ 1 are the solution of : j m11 0.. The Johansen test statistic is based on canonical correlation. ::::: yt p+1 ] 0 which summarizes the reducedrank regression proce dure used in the Johansen approach. In order to test for the signi…cance of the r largest eigenvalues. 0 corrected for longrun e¤ects which are white noise (i. 2006.e. where mij . Zt jXt 1 g which denotes the partial canonical correlations between Xt and Zt : both sets concentrate out the e¤ect of Xt 1 allowing us to obtain canonical correlation (see Johansen. m11 = T T u1t u01t of the residuals u0t ~ ~ ~ ~ ~ t=1 t=1 t=1 ~ ~ and u1t obtained in the multivariate least squares regressions ~ u0t and yt 1 yt = ( yt 1 . Hecq et al. i. 4 7 . 1995). are the second moment matrices: m00 P P P 1 1 1 m10 = T T u1t u00t . n (6) m101 m00 m01 j = P 1 = T T u0t u00t . This estimation is referred to as full information maximum likelihood . (see Johansen. yt 1 jXt 1 g where Xt 1 0 0 = [ yt 1 . j = 0:1. Moreover. In model (2) we can use the expression CanCorrf yt . 1995). The result of the Johansen test is a superconsistent estimation of .:::. we could also use a canonical correlation approach to determine the rank of the common features space due to WF restrictions. :::.FIML. Johansen. we can rely on Johansen’ trace statistic (6): s = T n X r Ln (1 ^2) i i = 1. and select the lagrank of VAR models.model may result in solving a problem of canonical analysis4 . We use canonical analysis is this work to estimate. we can use the expression CanCorrfXt .. as shown in next sections. ~ ( yt 0 y t 1 ) = ~ "t 0 where ~ "t is a white noise). This is a test for the existence of cofeatures in the form of linear combinations of the variables in the …rst di¤erences.
which consists of identifying p and s simultaneously. The above procedure is followed when there is evidence of cointegration restrictions. The question investigated is: Does IC(p. and the modi…ed informational criteria. Vahid and Issler (2002) analyzed a covariancestationary VAR model with SCCF restrictions. AIC (p. IC(p). s) performs better than IC(p) in VAR model lag order selection. using the Johansen cointegration test5 . r. Conditional on the results of the cointegration analysis. 8 . s) perform better than IC(p)? This is an important question we aim to answer in this work. Vahid and Issler. 3. Estimate p using standard informational criteria: Akaike (AIC). estimate a …nal VECM then calculate the multistep ahead forecast. s) = 5 i=n s+1 T X ln(1 2 i (p)) + 2 T N (7) Cointegration rank and vectors are estimated using the FIML. 2. a novelty in the literature. Additionally we check the performance of IC(p. Using the lag length chosen in the previous step. The model estimation following the standard selection criteria. as shown in Johansen (1991). We chose the lag length of the VAR in levels that minimizes the information criteria. In the present work we analyze the cointegrated VAR model with WF restrictions in order to analyze the performance of IC(p) and IC(p. …nd the number of cointegration vectors. originally used by Vahid and Engle (1993) entails the following steps: 1. 2006). IC(p).3 Model Selection Criteria In model selection we use two procedures to identify the VAR model order: the standard selection criteria. s) for model selection. IC(p. We check the performance of IC(p) when WF restrictions are imposed on the true model. The procedure to choose the lag order and the rank of the structure of shortrun restrictions is carried out by minimizing the following modi…ed information criteria (see. They showed that the use of IC(p. s). Schwarz (SC) and HannaQuinn (HQ). 2002. s) alternative selection criteria. Hecq.
The eigenvalues i are calculated for each p.e. Consider the 9 . no work has analyzed the order of the VAR model with cointegration and WF restrictions using a modi…ed criterion. Few works have analyzed the order of VAR models considering modi…ed IC(p. After selecting the pair (p. 2006). so we are inverting the hierarchical procedure followed by Vahid and Engle (1993) where the …rst step is to select the number of cointegration relations. which is exactly the contribution of this paper. especially when r is overestimated. from the number of restrictions the common dynamics imposes from s (n (p 1)) s (n s). However. Notice that in this simultaneous selection. N is obtained by subtracting the total number of mean parameters in the VECM (i. s) to simultaneously choose the order p and a number of reduced rank structure s in a covariance stationary VAR model subject to SCCF restrictions. testing the cointegration relation is the last procedure followed. n2 (p 1)+nr).HQ(p. s) we can test the cointegration relation using the procedure of Johansen. Vahid and Issler (2002) suggested the use of IC(p. In order to calculate the pair (p. This may be an advantage. that is. 2:::n by j yt 1 ). Finally we estimate the model using the switching algorithms as shown in the next section. s) = N = [n (n (p i=n s+1 T X ln(1 r] 2 i (p)) + ln T T (p N 1) + (n s))] (9) 1)) + n [s (n where n is the number of variables in model (2) and N is a number of parameters. s). This procedure is followed for every p and in the end we choose the p and s that minimize the IC(p. s). To estimate the VAR model considering cointegration and WF restrictions we use the switching algorithms model as considered by Hecq (2006). for given r and p. s) = i=n s+1 T X ln(1 2 i (p)) + 2 ln(ln T ) T N (8) SC(p. We …x p in model (3) and then …nd computing the cancorr( yt . r = n (see Hecq.. As mentioned. s) we assume that no restriction exists. Xt 1 i i = 1.
2 r<n. We estimate (11) by FIML. We label the residuals u0 and u1 . s=s . (8) or (9). Using the optimal pair (p. The determinant of the covariance matrix for r = n 2 cointegration vector is calculated by ln max X = ln m00 m01 m111 m10 10 r=n. ^ i i = 1. r = r) using the Johansen coin in step one. We cal2 culate the eigenvalues associated with ^ . s=s = Pmax T ln r=n. s=s s X i=1 ln 1 ^2 i (13) . (and so its rank. Step2 : Estimation of ~ ? and Taking the estimate of Hence. r=n. we estimate tegration test. s) chosen by the information criteria (7). Therefore. . we proceed to estimate ~ ? and 0 yt and of Xt 1 on ^ yt 1 . ~ ? of n can be written as (n = C1 . :::. s=s " # max s X X 2 T L0 r<n. instead of (12) we use the derived loglikelihood: Lmax. respectively. we can obtain ~ ? Step3 : Estimation of the maximum likelihood (ML) function. Hence. Given the parameters estimated in steps 1 and 2 we use a recursive algorithm to estimate the maximum likelihood (ML) function.VECM given by: yt = 0 yt 1 + ~ ? Xt 1 + "t (10) A full description of switching algorithms is presented below in four steps: Step1 : Estimation of the cointegration vectors . we compute the ML function: r. Thus. s=s i=1 If r = n. we obtain a reduced rank regression: u0t = ~ ? u1t + "t where and ^ . we run a regression of . s and the matrix P of residuals max . s=s = ln ln 1 ^ i (12) max. C(p 1) (11) of (n s) n(p 1) and s). :::.
s=s ): r=r. we calculate the new value of the ML function in the step 3. s=s  We repeat steps 1 to 4 to choose ~ ? and until convergence is reached ( i. r=r. the optimal parameters p. Motivated by these applications and according the previous work of Vahid and Issler (2002) we construct a MonteCarlo experiment in a threedimensional environment. n = 3.. r=r. h i ^ Xt 1 In order to reestimate we compute the CanCorr yt . Hence. output and money are notable examples. 4 MonteCarlo Design The simple real business cycle models and also the simplest closed economy monetary dynamic stochastic general equilibrium models are threedimensional. we obtain L1 max. Then. Therefore. 11 . one cointegration vector.e. yt 1 j and thus using the new ^ we can repeat step 2 to reestimate ~ ? and . L0 max. In this case. the data generating processes considering a VAR model with three variables. Consumption. saving and output and prices. s=s to calculate L = (L1 max. r = 1 and s = 2. estimation does not entail using an iterative algorithm yet because the cointegrating space spans Rn : Step4 : Reestimation of : We reestimate to obtain a more appropriate value for the parameters.e. L < 10 7 ): In the end. and two cofeatures vectors (i..where mij refers to cross moment matrices obtained in multivariate least square regressions from yt and Xt 1 on yt 1 . r and s are obtained and they can be used to estimate and forecast of a VECM with WF restrictions.
An initial idea to design the MonteCarlo experiment can consist of constructing the companion matrix (F ) and verifying whether the eigenvalues of the companion matrix all lie inside the unit circle. based on the choice of the eigenvalues. this strategy could lead to a wide spectrum to search for adequate values for the companion matrix. we follow an alternative procedure. F = 4 0 yt 2 (A2 + A3 ) I3 (A2 + A3 ) 0 "t 5 and vt = 4 0 5 0 0 +1 "t (15) 3 . This can be carried out by selecting their values from a uniform distribution. and ~ satisfy: 2 2 "1t 4 "2t "3t Consider the VAR(3) model: yt = A1 yt 3 2 1:0 = 4 0:2 5 . and then on the generation of the respective companion matrix. equation (15) will be covariancestationary if all eigenvalues of matrix F lie inside the unit circle. In the appendix we present a detailed discussion of the …nal choice of these free parameters. In our simulation. including analytical solutions. we constructed 100 12 3 2 4yt = 4 4yt 1 5 .respectively). We propose an analytical solution to generate a covariancestationary VAR. b) WF restrictions: ~ A3 = 0 0 (iii) ~ (A2 + A3 ) = 0 and c) the covariancestationary condition. Hence. we can rewrite (14) as the following VAR(1): t =F t 1 + vt A3 0 0 A3 3 2 t Thus. However. Considering the cointegration restrictions. and then verifying whether or not the eigenvalues of the companion matrix all lie inside the unit circle. The VECM representation as a function of the VAR level parameters can be written as: yt = (A1 + A2 + A3 I3 )yt 1 (A2 + A3 ) yt 1 A3 y t 2 + "t (14) The VAR coe¢ cients must simultaneously obey the restrictions: a) The cointegration restrictions: 0 = (A1 + A2 + A3 0 I3 ) . 4 0:6 0 0:6 1 1:0 0:0 0:5 0:6 0:6 1:0 0:6 0:6 1:0 2 3 0:1 1:0 5 0:5 31 5A 3 + A2 y t + A3 y t + "t . ~ = 4 1:0 3 02 3 2 0 1:0 5 s N @4 0 5 .
The true lagcointegrating vectors are identi…ed by boldface numbers and the selected lagcointegration vectors chosen the most times by the criterion are underlined. All the experiments were conducted in the MatLab environment. the AIC most often chooses the correct lag length for 100 and 200 observations.data generating processes and for each of these we generate 1000 samples containing 1000 observations.r=1 and s=2 for 100 and 200 observations The values in Table 1 represent the percentage of time that the model selection criterion.000 runs. One example of a VAR(3) model with n=3. we considered only the last 100 and 200 observations. 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 T im e 60 70 80 90 100 0 20 40 60 80 100 0 20 40 60 80 100 T im e 120 140 160 180 200 Figure 1. IC(p). the results show that. To reduce the impact of the initial values. in general. chooses the cell corresponding to the lag and number of cointegration vectors in 100. 5 Results Figure 1 shows one example of the threedimensional VAR model with cointegration and WF restrictions for 100 and 200 observations. In Table 1. For 13 .
75% and 35. HQ(p. the AIC(p. The most relevant results can be summarized as follows: 14 .34%. 62.34% in choosing the true p = 3.72%.85% accuracy while the HQ(p) only has a success rate of 35. all three criteria selected the true cointegrated rank r = 1. What happened when the weak form common cyclical restrictions are ignored? Tables 1 and 2 also show the relative performance of IC(p. s) visàvis IC(p). For T = 200. Note that all three criteria chose the correct rank of cointegration (r = 1). s). 56. HQ and SC criteria chose the true lag. 40. s) criteria more frequently choose the true pair (p. in comparison with a rate of 54.06%.000 runs. 1). the AIC(p. while the SC(p) only has a success rate of 17. s) and SC(p.example. s) criterion more frequently chooses the lagrank for 100 and 200 observations. The true lagrankcointegration vectors are identi…ed by boldface numbers and the best lagrank combination chosen the most times by each criterion are underlined. For T = 100 the AIC(p.58% and 45. For instance.08%.2% in selecting the true p = 3. Thus. respectively. When 200 observations was considered. s) = (3. s) = (3. s) and SC(p. chooses that cell. p.48% of the times respectively. 54.28% of the times for AIC. corresponding to the lagrank and number of cointegrating vectors in 100. For T=200. for 100 observations. the results show that. the gains are more than 3%. 1). s) has a success rate of 25. In Table 2. This represents a gain of more than 44%. for T = 100 the SC(p. Table 2 contains the percentage that the alternative model selection criterion.62%. the gains are more than 8%. IC(p. For T = 100 the HQ(p. s) criteria more often choose the true pair (p. For 200 observations.48%.s) criteria the cost of ignoring the weak form common cyclical restriction is low. the AIC. the gains are more than 27%. the correct lag length was chosen 74. it appears that when using the AIC(p. HQ and SC respectively. For instance. the AIC(p. Again. s) selects the true p =3 with 40.62% and 17. 35. a gain of more than 4%.2% of the times. 57. Note that all three criteria more often choose the correct rank of cointegration (r = 1) in both samples. for 100 observations. s) has a success rate of 56. 77.85% and 25. For T = 200.08% for the AIC(p). s). This represents a gain of 14%.03% of the times respectively. in general. s). HQ(p.
HQ and SC) choose the correct parameters more often when using IC(p. the standard Schwarz. s) visàvis IC(p). In general. and an alternative criterion (IC(p.All criteria (AIC. since it provides 15 . when the VAR model contains WF and cointegration restrictions. the traditional information criterion. s) criteria to choose the lagrank. The results indicate that the information criterion that selects the lag length and the rank order performs better than the model chosen by conventional criteria. While cointegration refers to relations among variables in the longrun. or HannanQuinn. These additional WF restrictions are de…ned in the same way as cointegration restrictions. we suggest the use of AIC(p. In applied work. The AIC performs better in selecting the true model more frequently for both the IC(p. we considered an additional weak form restriction of common cyclical features in a cointegrated VAR model in order to analyze the appropriate way to choose the correct lag order. When the WF restrictions are ignored there is a non trivial cost in selecting the true model with standard information criteria. s) and the IC(p) criteria. the standard Schwarz or HannanQuinn selection criteria should not be used for this purpose in small samples. selection criteria should not be used for this purpose in small samples due to the tendency to identify an underparameterized model. due to the tendency to identify an underparameterized model. SC(p). 6 Conclusions In this work. Two methodologies have been used for selecting lag length. There is a cost of ignoring additional weak form common cyclical restrictions in the model especially when the SC(p) criterion is used. the common cyclical restrictions refer to relations in the shortrun. In general. HQ(p). IC(p). s)) that selects the lag order p and the rank structure s due to the weak form common cyclical restrictions.
Misspeci…cations in Vector Autoregressions and Their E¤ects on Impulse Responses and Variance Decompositions. Tiao GC. Biometrika. Cointegration and error correction: representation. Ann. Braun PA. EconomicModelling. Cubadda G. Since no work in the literature has analyzed a VAR model with WF common cyclical restrictions. Evidence on Common Features and Business Cycle Synchronization in Mercosur. 55. Engle R. Carrasco CG. Econometrica. Caporale GM. Box GE.considerable gains in selecting the correct VAR model. Statist. 8 1400]. . W. EconomicModelling. [Correction (1980) Ann. Mittnik S. A canonical analysis of multiple time series. Granger C. Statist. 16 . T. Gomes F. Brazilian Review of Econometrics. 31941. the results of this work provide new insights and incentives to proceed with this kind of empirical work. (1987). 14: 19.(1993). (2009). References Anderson. (1993). Journal of Business and Economic Statistics. estimation and testing. Engle R. 64:355365. (2007) Common Shocks. Kozicki S. Testing for common features. Journal of Econometrics. Common Features and Output Fluctuations in the United Kingdom. Commo Dynamics. (1997). 25176. 59. and the International Business Cycle. (1977). 24. 22 327351. Forthcoming. Hecq. 149166. (1951) Estimating linear restrictions on regression coe¢ cients for multivariate normal distributions. Math. Centoni M. 11: 369395.
multivariate Analysis. 47. Five Alternative Methods of Estimating LongRun Equilibrium Relationships. 10. Reduced rank regression for the multivariate linear model. Engle R. Johansen. 60 issue 12. (1995). Z. Testing for Common Cyclical Features in VAR Models with Cointegration. Mimeo. 17 . Oxford Univ. Urbain JP. Journal of Forecasting. Common Trends and Common Cycles in Latin America. Press. Issler JV. 35:83113. 59. Issler JV. Common Cycles and Common Trends in Latin America. 203233.and LongTerm Economic Restrictions: A MonteCarlo Study. Izenman AJ. (2006). Guillén. Revista Brasileira de Economia. Estimation and Hypothesis Testing of Cointegration Vectors in Gaussian Vector Autoregressive Models. Impulse response analysis in vector autoregressions with unknown lag order. Cointegration and Common Cyclical Features in VAR Models: Comparing Small Sample Performances of the 2Step and Iterative Approaches. (2001). S. Hecq A. Palm FC. (2005). Kilian L. 132: 117141. J. Journal of. (2002). vol. no 2.(1993). A. Journal of Econometrics. (1991). Gonzalo. Hecq. Medium Econometrische Toepassingen. Johansen S. 1993.Engle R. (1995). (2006). Estimating Common Sectoral Cycles. 20: 161179. pp. Likelihoodbased Inference in Cointegrated Vector Autoregressive Models. Forecasting Accuracy and Estimation Uncertainty using VAR Models with Short. (1994). Econometrica. Journal of Econometrics. 589. 5: 248264. (1975). 15511580. 2025. 149176. Issler. Ensaios Econômicos EPGE. Hecq A. Athanasopoulos. Journal of Monetary Economics.
Iyare SO. Engle R. Issler JV. Vahid F. Reduced rank regression and Canonical analysis. (1981). 109: 341363. Wichern DW. The Importance of Common Cyclical Features in VAR Analysis: A Monte Carlo Study. Vahid F. Reduced rank models for multiple times series. New York. 8: 341360. Velu RP. Mamingi N. Journal of Econometrics. (1993). Introduction to Multiple Time Series Analysis. Convergence and Common Features in International Output: A Case Study of the Economic Community of West African States. 19751997. 18 . Biometrika. Springer. Reinsel GC. Journal of Applied Econometrics. Journal of the Royal Statistical Society. 43: 183189. (2003). (1986).Lütkepohl H. 73: 105118. Common trends and common cycles. (1993). (2002). Tso MK. 3: 116. AsianAfricanJournalofEconomicsandEconometrics.
027 0.146 0.027 0.776 0.257 0. in selecting the lag order p Frequency of lag(p) and cointegrating vectors (r) choice by different criteria for trivariate VAR model in levels when the true model have parameters: p = 3 and r = 1.000 0.002 52.749 1.612 0.902 0.000 0.025 0.000 17.000 0.000 0.000 0.000 0.073 0.016 0.070 0.000 37.000 0.907 2.105 0.593 1.059 0.006 0.158 0.000 0.004 0.000 0.996 0.042 17.003 0.189 0.002 0.884 1.048 0.012 35.000 0.136 2.009 0.828 0.313 0.108 0.005 4.037 0.156 2.006 0.000 0.000 0.000 0.333 0.025 0.009 0.001 0.081 0.065 0.000 0.076 0.Appendix A: Tables Table I Performance of information criterion.000 0.609 0.041 0.098 0.133 0.068 0.045 0.000 1.071 0.095 0. IC(p ).614 0.279 2.023 0. The true lagcointegrating vectors are indentified by bold numbers.011 0.012 74.003 61.044 0.390 1.001 0.017 0.016 0.107 0.098 0.966 2.007 0.003 0.000 0.082 0.014 0.036 0.073 0.000 0.057 0.093 0.001 0.083 0.053 0.000 0.000 0.686 0.025 0.000 0.001 0.000 0.005 0.080 0.915 0.177 0.344 6.059 0.987 0.169 0.385 0.000 3.049 0.000 3.013 0.001 0.091 0.511 0.000 0.283 0.000 0.359 0.000 0.485 0.488 0.728 0. 19 .035 0.002 0.165 0.060 0.600 35.063 0.003 0.000 0.000 0.964 1.000 0.000 0.031 0.007 0.056 0.392 0.020 0.019 0.792 54.243 0.012 57.617 0.000 0.000 8.000 55.016 0.015 0.000 0.000 0.033 0.000 0.001 0.001 0.115 0. Number of observations = 100 Selected cointegrated vectors 0 1 2 3 Number of observations = 200 Selected cointegrated vectors 0 1 2 3 Selected Lag 1 2 3 AIC(p) 4 5 6 7 8 1 2 3 0.133 0.721 1.146 1.219 0.000 0.000 realizations.055 0.000 0.013 0.044 0.000 0.039 0.000 0.030 0.082 0.000 0.000 0.173 0.001 0.737 4.005 0.000 0.002 32.031 0.166 0.000 0.076 0.000 0.082 0.000 HQ(p) 4 5 6 7 8 1 2 3 SC(p) 4 5 6 7 8 Numbers represent the percentage times that the model selection criterion choice that cell corresponding to the lag and number of cointegration vectors in 100.021 0.
000 0.000 0.000 0.000 0.000 0.000 0.000 2.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.563 40.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.002 0.065 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.002 0.000 0.000 0.888 1.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.088 0.000 0.000 0.000 55.370 1.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.s ) SC(p .000 0.000 0.000 0.000 0.000 0.000 0.020 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0. in selecting p and s simultaneously.000 0.000 0.000 0.000 1.000 0.000 0.000 0.s).000 0.000 0.000 0.000 0.000 70.000 0.000 0.855 0.000 0.000 0.025 0.000 0.000 0.s ) SC(p .000 0.351 1.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.301 0.003 0.000 0.000 0.000 0.000 0. 20 .000 0.000 0.000 0.000 0.000 0.000 0.000 39.000 0.000 0.000 0.000 0.000 0.000 0.000 0.s ) Numbers represent the percentage times that the simultaneous model selection criterion IC(p.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.887 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.003 0.877 0.221 0.000 0.086 0.000 0.000 0.000 0.000 33.000 0.000 0.000 0.416 0.000 0.146 0.049 56.000 0.000 0.000 0.000 0.000 0.908 0.000 0.s ) 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 1 0 2 3 1 1 2 3 1 2 2 3 1 3 2 3 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0. Vectors (r) Selected rank (s) Selected lag (p) Sample size = 100 1 2 3 AIC(p .000 HQ(p .043 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 18.207 0. corresponding to the lagrank and number of cointegrating vectors in 100.056 0.000 0.000 realizations.s) choice that cell.000 0.000 0. IC(p.000 0.004 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.038 0.000 0.000 0.260 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.012 0.000 0.000 0.000 0.000 0.001 0.000 0.000 2.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.000 0.000 0.000 0.797 77.000 0.267 0.028 0.000 0.000 0.000 0.000 0.952 62.000 0.000 0.000 0.000 0.000 0.000 0.000 0.081 0.000 0.000 0.000 0.000 0.000 0.218 1.076 0.000 0.000 0.000 0.063 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.681 2.000 0.576 0.000 0.000 0.000 0.000 0.070 0.000 0.000 0.000 0.000 0.000 0.000 0.000 AIC(p .000 0.000 0.000 0.000 0.000 0.983 45.000 0.000 0.035 0.000 0.s ) HQ(p .000 0.000 0.002 0.000 0.204 0.000 0.000 0.000 0.559 0.114 0.000 0.111 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0. Johansen Tested coint.000 0. Performance of information criterion.000 0.000 0.574 0.000 0.000 0.027 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.001 0.000 0.000 0.004 0.000 0.000 0.000 0.000 0.000 0.000 0. The true lagrankcointegration vectors are identified by bold numbers and the best lagrankcointegration vectors chosen by criterias are identified by underline lines.145 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.s ) Sample size = 200 0.000 0.000 0.006 0.000 50.000 0.000 0.Table II.000 0.000 0.341 1.000 1.000 0.000 0.000 0.000 0.000 0.000 0.186 0.000 0.000 0.002 0.000 0.113 0.000 0.000 0.000 0.000 0.000 0.971 25.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.005 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.053 0.000 0.000 0.
Rj1 = ~ j1 = ~ 11 . Let.Appendix B: VAR Restrictions for the DGPs Consider the vector autoregressive. VAR(3). model : y t = A 1 y t 1 + A2 y t 2 + A3 y t 3 + " t 3 2 2 2 1 a11 a2 a11 a1 a1 12 12 12 with parameters: A1 = 4 a1 a1 a1 5. 3) and S = j2 12 3 Ga3 33 0 Ka3 5 (i) ~ A3 = 0 ==> A3 = 4 33 a3 33 2 Ga2 31 0 0 (ii) ~ (A2 + A3 ) = 0 ==> ~ A2 = 0 ==> A2 = 4 Ka2 31 a2 31 2) Longrun restrictions (cointegration) 21 + 21 K + 31 Ga3 Ga3 31 32 3 Ka31 Ka3 32 a3 a3 31 32 Ga2 32 Ka2 32 a2 32 3 Ga2 33 Ka2 5 33 a2 33 . K = [(R32 11 G 2 R31 )=(R21 R22 )]. the VECM respre =4 2 5. the cofea11 21 31 3 5 : The yt 1 (A2 + A3 ) yt 1 A3 y t 2 + "t (17) We can rewrite equation (17) as a VAR(1): t =F t 1 + vt A3 0 0 A3 where t 3 2 4yt = 4 4yt 1 5 . Rj2 = ~ = ~ (j = 2. G = [R21 K + R31 ]. A2 = 4 a2 a2 22 21 22 22 21 a2 a2 2 a1 a1 a1 31 32 31 32 32 2 3 3 a11 a3 a3 12 12 4 a3 a3 a3 5 : We consider the cointegration vectors = 4 21 22 22 a3 a3 a3 31 32 32 2 ~ ~ 12 3 11 tures vectors ~ = 4 ~ 21 ~ 22 5 and the adjustament matrix ~ ~ 31 32 (16) 3 a2 12 a2 5 and A3 = 22 a2 3 32 11 21 31 longrun relation is de…ned by sentation is: yt = 0 0 = (A1 + A2 + A3 I3 ): Thus. F = 4 0 yt 2 (A2 + A3 ) I3 (A2 + A3 ) 0 (18) 3 2 3 "t 5 and vt = 4 0 5 0 0 +1 "t 1) Shortrun restrictions (WF) We now impose the common cyclical restrictions (i) and (ii) on model (16).
+ a3 ) + a3 and 31 33 Ga3 b 31 = 3 Ka3 b + 32 4 + a3 S 32 3 21 + a31 S and the …rst four roots are 22 = 2 = = 0:We calculated . the companion matrix F can be represented as: F =4 2 (A2 + A3 ) I3 (A2 + A3 ) A3 0 0 A3 5= 0 +1 3 5+1 = 11 11 + 21 21 + 31 31 + 1 0 2 6 6 6 6 6 6 6 6 4 G(a2 + a3 ) 31 31 K(a2 + a3 ) 31 31 (a2 + a3 ) 31 31 1 0 0 (a2 + a3 )S 31 31 with b = 0 G(a2 + a3 ) 32 32 G(a2 + a3 ) 32 32 G(a2 + a3 ) 32 32 0 1 0 (a2 + a3 )S 32 32 + G(a2 + a3 ) 33 33 G(a2 + a3 ) 33 33 (a2 + a3 ) 33 33 0 0 1 (a2 + a3 )S 33 33 + +1 Ga3 31 Ka3 31 a3 31 0 0 0 a3 S 31 Ga3 32 Ka3 32 a3 32 0 0 0 a3 S 32 Ga3 33 Ka3 33 a3 33 0 0 0 a3 S 33 7 7 7 31 7 0 7 7 0 7 7 0 5 b 21 11 3 +1= 11 11 21 21 31 31 3) Covariancestationary restrictions Equation (18) will be covariancestationary if all eigenvalues of matrix F lie inside the unit circle. = Ga3 + Ka3 31 32 S 2 21 (a32 .The cointegration restrictions are speci…ed by (iv) and (v): (iv) 0 0 (A2 + A3 ) = [ (a2 + a3 )S 31 31 a3 S 32 = 11 (a2 + a3 )S 32 32 2 4 11 21 31 (a2 + a3 )S] and 33 33 A3 = [ a 3 S 31 (v) 0 a3 S] 33 21 31 +1 = 3 Taking into account the short.and longrun restrictions. That is. and are: = G(a2 +a3 )+K(a2 +a3 )+a2 +a3 b. 31 31 32 32 33 33 Kb(a2 + a3 ) + 32 32 a3 b 33 1 2 31 S(a33 (a2 + a3 )b 33 33 2 11 (a31 Gb(a2 + a3 ) 31 31 = + a3 ) + 33 3 31 a33 S + a3 ) + S 32 11 . eigenvalue of matrix F is a number jF The solution of (19) is: 7 such that: (19) I7 j = 0 + 6 + 5 + 4 =0 (20) where the parameters .
a2 . 7. = 6 7 + 6 5 + 5 7 Equaling these parameters with the relations above we have: Ka2 b + 32 21 5 2 31 Sa33 11 6 7 6 6 7 6 7 a2 = 31 5 6 ( Ka2 32 a2 b 33 5 6 7 +b 5 7 a2 + Sa2 33 32 7 )=(S 7 G Gb) 11 + 6 a3 = ( S 2 32 Ga3 bS 31 S S2 S S K 2 7 5 11 31 +S 5 11 31 5 b2 11 b G Gb2 +b 7 S 31 SG S 11 b a3 S 31 11 5 3 2 11 G+a31 S 31 2 11 31 + Gb2 + 6 S 7 5 5 7 31 SG 5 6 S2 S2 2 11 6 Gb 11 11 31 Gb 5 7 6 G+ 6 7 Gb+ Gb+ Gb SGb2 S S 11 b 2 31 +S 11 b 31 31 2 2 31 a33 + S 2 2 2 3 31 a33 G + SG a31 21 G 31 5 7 6 21 )=S 31 + S 2 3 11 a33 b + Gb 2 3 31 G a31 b S2 2 11 Ka32 2 2 3 31 Ga31 +S a32 31 5 6 Sa2 32 21 Gb+S 5 7 6 2 31 a33 Gb+S 31 2 11 Ka32 b+ Gb 31 SG 21 31 SG+ S 11 )=(S 11 K KG 31 +bG 21 31 Gb S 11 +G 5 a3 = 33 6 5 (Kb3 G 21 7 6 7 Gb2 K + S 6 11 6 6 K 7 5 + Kb 7 S 6 11 21 + Kb2 5 7 G S2 5 21 7 11 + GbS S2 11 11 11 21 +S 21 Gb Gb2 K + 7 6 S 11 Kb 7 5 S2 11 GbS 21 + 5 2 S 11 Kb 11 2 21 Ka32 + 2 11 S 21 G+Kb G+Kb S G+Kb 2 5 6 G 7 6 KG S2 S2 S2 21 b S2 11 2 2 21 a33 +S 2 2 21 a32 G 11 Kb +S 21 G 2 3 a31 S 21 Gb +S 2 a3 K 31 S S 21 7 5 3 2 21 Ga31 +S 31 +S 21 2 21 a33 G 31 +S 2 3 a31 b 31 Gb 2 11 K^2ba32 +S 5 6 2 11 Kba33 S 3 11 a31 KG 3 11 KbGa31 SKba2 G 33 KG 31 21 G S 21 G S +G 21 5 7 6 G S 2 21 Ka32 Gb G)=(S 11 K 31 + bG K S 11 21 21 )=S We can calculate a2 . Hence we have three roots satisfying equation (20) 3 6 and 7) and + 5 6 7 2 + =0 =0 =0 = + =0 ::::::::::::::::::::::::::::::::::Eq1 ::::::::::::::::::::::::::::::::::Eq2 ::::::::::::::::::::::::::::::::::Eq3 7 6 5. 0:9) the values of a3 .the parameters of matrices A1 . 2 and 3 yields: and = 5 6 7. a2 . we have: we have: we have: 3 5 3 6 3 7 + + + 2 5 2 6 2 7 + + + + + + Solving equations 1. (21) for for for 5. 6. 31 32 33 23 . free parameters. a3 and a3 …xing the set 31 32 33 1 = 2 = 3 = 4 = 0 and sort 5. independently from uniform distributions ( 0:9. A2 and A3 as functions of roots ( 5 .
Hence.6 and 7. 24 . A2 and A3 are de…ned and we can generate the DGPs of VAR(3) model with cointegration and WF restrictions. each parameter of the matrices A1 .