International Journal of Forecasting 22 (2006) 751 – 770

www.elsevier.com/locate/ijforecast

Restricted forecasting with VAR models: An analysis of a test for

joint compatibility between restrictions and forecasts
Nicolás Gómez a,*, Vı́ctor M. Guerrero b,1
a
Departamento de Fı́sica y Matemáticas, Universidad Iberoamericana 01210 D.F. México, México
b
Departamento de Estadı́stica, Instituto Tecnológico Autónomo de Mexico (ITAM) 01000 D.F. México, México

Abstract

A restricted forecasting compatibility test for Vector Autoregressive Error Correction models is analyzed in this work. It is
shown that a variance–covariance matrix associated with the restrictions can be used to cancel out model dynamics and
interactions between restrictions. This allows us to interpret the joint compatibility test as a composition of the corresponding
single restriction compatibility tests. These tests are useful for appreciating the contribution of each and every restriction to the
joint compatibility between the whole set of restrictions and the unrestricted forecasts. An estimated process adjustment for the
test is derived and the resulting feasible joint compatibility test turns out to have better performance than the original one. An
empirical illustration of the usefulness of the proposed test makes use of Mexican macroeconomic data and the targets proposed
by the Mexican Government for the year 2003.
D 2005 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

Keywords: Cointegration; Compatibility test; Economic targets; Finite sample adjustment; Multiple time series

1. Introduction the variable at time T + h with information up to T

Decision making is mainly based on predictions
of the most relevant variables for the problem at
hand, and time series data are usually employed to
get forecasts. In the present case we let y t denote
the value of the variable at time t; then a forecast of
* Corresponding author. Tel.: +52 55 5950 4000x7660; fax: +52
55 5950 4284.
E-mail addresses: nicolas.gomez@uia.mx (N. Gómez),
guerrero@itam.mx (V.M. Guerrero).
1
Tel.: +52 55 5628 4000x3837; fax: +52 55 5628 4086.
0169-2070/$ - see front matter D 2005 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.ijforecast.2005.12.002
is given by
 
ŷy T þh ¼ f yT ; yT 1; . . . :
If available, additional information should also be
considered to get the forecast. For example, let us
consider a case where the Government announces an
economic target for next year. Since the Government
has the power to implement the economic or social
policies to approach the target, the future information
available (the target in this case) should be taken into
account to get the forecast. Thus, given some targets
for the variable under study, it is important to know
752 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770
the future path that will lead to achieving that target.
The path obtained as a combination of the forecasts
from a model and the additional information provided
by the targets (known as restricted forecasts) produce
such a scenario of future values. Let us suppose that
h T+i is the target for time T + i, thus the forecast will
be given by
 
ỹy Tþh ¼ f hT þi; yT ; yT 1; : : : :
When working with a single economic variable
whose data consist of a univariate time series, the
problem of incorporating the additional information
has already been treated in the literature. Guerrero
(1989) and Trabelsi and Hillmer (1989) obtained the
corresponding optimum forecast, in the minimum
Mean Square Error (MSE) sense. Such a forecast
allows an analyst to incorporate additional informa-
tion as binding or unbinding restrictions.
The literature on restricted forecasts for multiple
time series includes several papers. Doan, Litterman,
and Sims (1984), Greene, Howrey, and Hymans
(1986), van der Knoop (1987) and Pankratz (1989)
dealt with the combination of historical information
with additional information provided by way of
linear restrictions. Pandher (2002) attacked the same
problem within the state-space framework. Guerrero,
Pena, Senra, and Alegrı́a (2005) focussed on
implementing a Vector Error Correction (VEC)
model for monitoring Mexican economic targets
and verifying the compatibility between historical
and additional information. A joint compatibility
test (JCT) for the restriction was proposed to that
end.
When the targets impose linearly independent
restrictions on the forecasts we can always obtain
the corresponding restricted forecasts, no matter how
discordant the targets are from the historical
information. So, the JCT proves to be a useful tool
for deciding whether or not the targets are compat-
ible with the forecasts based only on the historical
record. Knowing whether these two sources of
information are compatible or not is useful in many
ways. On the one hand, when the two sources of
information are compatible with each other, incor-
porating the targets into the forecast as additional
information will reduce the forecast MSE. On the
other, when the test rejects compatibility, we are led
to analyze the reason why this occurred. Then we
would like to know which restrictions were the most
likely causes of rejection. However, the JCT does
not provide any clue about this. Therefore the need
to be able to appreciate the individual contribution
of each restriction to the joint compatibility arises.
Here we address this issue by means of a
decomposition of the JCT into tests for each
individual restriction involved. Since the JCT
introduced in Guerrero et al. (2005) is based on
asymptotic theory, a feasible version of this test that
works well with estimated parameters is derived
here.
This article is organized as follows. Section 2
presents the statistical methodology needed to get a
restricted forecast with vector autoregressive (VAR)
models. In Section 3 we show how a matrix associated
with the uncertainty when imposing unbinding
restrictions can be used to cancel out both model
dynamics and interactions between restrictions. This
fact allows us to obtain the JCT as the sum of several
single compatibility tests (SCT), one for each indi-
vidual restriction. Section 4 presents a finite sample
Monte Carlo study of the JCT in order to validate its
performance. This study led us to an adjustment for
the fact that the parameters are estimated. The
corresponding restricted forecast formulas for the
feasible JCT are also derived. In Section 5 we
illustrate the methodology with an empirical applica-
tion that uses a model for the Mexican Economy and
introduces the economic targets for the year 2003, as
announced by the Mexican Government at the end of
2002. Finally, Section 6 presents some conclusions.
2. Methodology
Many economic time series may tend to move up
or down over time in a non-stationary way, but groups
of variables may drift together. A multiple time series
model is useful for analyzing the relationships among
these variables, and cointegration analysis helps to
discover the linear relationships that hold over the
long run. If the variables are cointegrated, the basic
tool to use for analysis and forecasting is a VEC
model. The forecast restricted by extra-model infor-
mation is described, as well as a compatibility test of
unrestricted forecasts and extra-model information.
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770
2.1. The VEC model
Let y t = ( y 1t , . . ., y kt )V be a k  1 vector of random
variables at time t, and let us assume that y t follows a
finite pth-order Gaussian VAR model,
yt ¼ LDt þ C 1 yt1 þ . . . þ C p ytp þ et ð1Þ
where /i is a k  k coefficient matrix for i = 1, . . ., p,
D t = (D 1t , . . ., D nt )V is an n  1 vector that includes
deterministic variables to account for seasonality and
intervention effects, as well as exogenous variables
with respect to y t . e t = (e 1t , . . ., e kt )V is a k-dimensional
independent and identically distributed N(0, e )
random error vector with time-invariant positive-
definite covariance matrix E(e t eV) t = e . The effects
of D t on y t are captured by the parameter matrix  of
order k  n.
Model (1) can also be written in terms of the lag
operator B as
C ðBÞyt ¼ LDt þ et ð2Þ
where /(B) = I  /1B  /2B2  . . .  /p Bp is a ma-
trix polynomial of order p on the lag operator B such
that B n y t = y t n for naN. This model can be
reparameterized as
C4ðBÞDyt ¼ LDt  Cð1ÞByt þ et ; ð3Þ
where D is the first difference operator, /(1) = I 
/1  . . .  /p and /*(B) is a matrix polynomial of
order p  1 on the lag operator B such that
/(B) = /(1)B + /*(B)D. Of course, it is also possible
to obtain the VAR in levels from the VAR in
differences. Thus Eqs. (2) and (3) are equivalent
representations of the same stochastic process. See for
instance Clements and Hendry (2004, Section 8.2).
When the variables in the VAR model are integrated
of order d z 1, written as I(d), Ordinary Least Squares
(OLS) estimation of the parameters in model (2) is
subjected to hazards typical of regressions involving
nonstationary variables, see Park and Phillips (1988)
and Park and Phillips (1989). The variables of a k-
dimensional process y t are cointegrated of order (d,
b), briefly y t ~CI(d, b), if all components of y t are I(d)
and there exists a linear combination cy t , with
c = (c 1, . . . ,c k ) p 0, which is I(d  b). We assume
753
det(/(x)) p 0 for |x| b 1, that is, all the zeros are on
or outside the unit circle. If the determinantal
polynomial has unit roots with multiplicity one, then
the variables of y t are I(1). Moreover, we assume that
the rank of /(1) is r, implying that there are d = k  r
unit roots in the system. When r N 0 the variables are
cointegrated, in the sense that there exists a linear
combination bVy t , with b = (b 1, . . ., b k )V p 0, which is
I(0). In that case  /(1) = AbV where A and b are
k  r full rank matrices. The rows of bV are then
referred to as the cointegration vectors of the system.
A result due to Engle and Granger (1987), known
as Granger’s Representation Theorem, is relevant
here. It states that if the k variables of y t are CI(1, 1)
then there exists a VEC representation for the system,
that is, Eq. (3) can be written as
C4ðBÞDyt ¼ LDt þ azt1 þ et ;
where zt = bVy t is stationary, the term /*(B) Dy t in
this model captures the short-run relationships among
the variables and  /(1)By t = Azt1 captures the
long-run relationships.
To introduce the restricted forecasting methodolo-
gy with VEC models, let us assume first that the
model and its parameters are known. Therefore we
will not consider such issues as specification, estima-
tion or validation of the model, although they must be
considered in practical applications. In Section 4 we
will address the problem of forecasting with estimated
parameters. Let us start with the kT  1 vector
Y = ( y 1V, . . ., y TV)V that contains all the past information
of the multiple time series, and let the kH  1 vector
Y F = ( yVT +1, . . ., yVT +H )V contain the H z 1 future values
to be forecasted for each series. The optimum (in the
MSE sense) linear forecast of y T+h for h = 1, . . ., H, is
its conditional expectation
   
E yTþh jY ¼ LDT þh þ C 1 E yTþh1 jY þ . . .
 
þ C p E yT þhp jY ð4Þ
where E( y T+hi |Y) = y T+hi for i z h. Such a forecast
produces the forecast error vector
X
h1
yT þh  EðyT þh jY Þ ¼ Cj eT þhj ; for h ¼ 1; . . . ; H
j¼0
ð5Þ
754 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770

Pj
where Cj ¼ k¼1 Cjk C k with C0 = I and /k = 0 for pairs of zeros and then the pair 1, 0. This kind of
k N p. It should be stressed that, by virtue of the restriction arises when the government announces
equivalence between Eqs. (2) and (3), the forecasts a target for the first variable for the end of the
provided by Eq. (4) make use of all the information in year.
the VEC model. Another example considers the difference between
Expression (5) can be written as two future values of the first variable r 1 = y 1,T+H 
y 1,T +1 + u 1 with u 1~N(0, r 11,u ) while the second
Y F  EðY F jY Þ ¼ CeF ð6Þ
variable has a restriction on the average of its future
where e F = (eVT +1 , . . ., eVT +H )V ~ N (0, I H  e ) is a values r 2 = y 2,T+i / H + u 2 with u 2~N(0, r 22,u ). Fur-
kH  1 vector, with  the Kronecker product. The ther, if u 1 and u 2 are uncorrelated, these two
kH  kH matrix C is lower triangular with Ik in its restrictions can be written as
main diagonal, C1 in its first subdiagonal, C2 in the
    
second subdiagonal and so on. Thus the optimum r1 u1 u 0 r11;u 0
¼CY F þ ; 1 fN ; ;
unrestricted forecasts of Y F is its conditional expec- r2 u2 u2 0 0 r22;u
tation Y F,H = E(Y F |Y), whose MSE is given by
    where
RY F;H uE Y F  Y F;H Y F  Y F;H VjY !
¼ CðIH  Re ÞCV: 1 0 ... 0 0 1 0
C¼ 1 1 1 :
0 ... 0 0
The forecast MSE for an integrated process is H H H
generally unbounded as the horizon H goes to infinity, In the empirical example shown in Section 5 we
that is, the forecast uncertainty increases without limit, face the problem of imposing three restrictions on the
see for example Lütkepohl (1991, Section 11.3). future values of a six-variable system of the Mexican
economy. One of those restrictions is of the difference
2.2. Stochastic linear restrictions on future values type and the other two involve averages.

Suppose we have additional information on the 2.3. Restricted forecasts

future of y t . For example, when the government wants
to reach some economic targets, we assume the When forecasting Y F we should consider both
forecasts of the variables y t are restricted by those historical information, Y, and additional information,
targets, at least to produce a scenario. We also assume R. The optimum forecast that takes into account both
that the additional information is in the form of a sources of information is E(Y F |Y,R), which will be
stochastic linear restriction denoted as Y RF,H . It can be shown, see for instance
R ¼ CY F þ u: ð7Þ Nieto and Guerrero’s (1995) Theorem 1, that

Here C is an M  kH matrix of full row rank, Y RF;H ¼ Y F;H þ A RCY F;H Þ
R = (r 1, . . ., r M )V is the M  1 vector of values that the
linear combinations take on and u = (u 1, . . ., u M )V~ where A= Y F,H CV6 1 and 6 u CY F,H CV + u is an
N(0, u ) is an M  1 random vector, with the ijth M  M symmetric positive-semidefinite matrix. More-
element of u given by r ij,u = cov(u i , u j ), for i, over, its MSE is given by
j = 1, . . ., M.   h   i
To illustrate this situation let us consider a MSE Y RF;H uE Y F  Y RF;H Y F  Y RF;H VjY ; R
bivariate VAR system where the first variable must
satisfy an isolated stochastic linear restriction as ¼ ðI  ACÞRY F;H :
r = y 1,T+H + u with u~N(0, r u ). Then in terms of Eq. Rearranging terms we have
(7) we have R = r, u = u and Y F = ( y 1,T+1, y 2,T+1, . . .,
 
y 1 , T + H 1 , y 2 , T + H 1 , y 1 , T + H , y 2 , T + H ) V w i t h
RY F;H ¼ MSE Y RF;H þ ACRY F;H ;
C = (0, 0, . . ., 0, 0, 1, 0) where there are first H  1
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770
since MSE(Y RF,H ) and ACY F,H are positive semide-
finite matrices, we see that Y RF,H is at least as precise as
Y F,H .
2.4. Compatibility test
In spite of the optimality of Y RF,H we should test
whether the two sources of information are compat-
ible with each other, in which case the combination
makes sense. Guerrero et al. (2005) proposed to define
an M  1 distance vector
duR  CY F;H ¼ CCeF þ u;
whose elements are denoted by d i ; i = 1, . . ., M. Then
R and Y F,H are said to be compatible if the distance
vector is close to zero. From the normality assumption
of e F and u, we know that before observing the values
of Y F,H and R,
d fN ð0; XÞ:
Thus the following JCT statistic arises naturally
K udVX1 d fv2M :
Then, after observing the values of Y F,H we say
that R is in the compatibility region at level a if
   
K calc ¼ R  CY F;H VX1 R  CY F;H Vv2M ðaÞ;
with v 2M (a) being the (1  a)th quantile of the v 2M
distribution. This distribution is exact when the
matrices C, e and u are known. When they are
consistently estimated, the v 2 distribution will be
valid asymptotically. It should be noticed that even
though the y t variables are I(1), the statistic K
follows a standard v 2 distribution since it is derived
from the distance vector d, which has a normal
distribution because of Eqs. (6) and (7).
Let us note that the JCT is an omnibus test that does
not consider any specific alternative hypothesis. If the
JCT leads us to conclude incompatibility between
sources of information, one reason could be that the
additional information imposed a binding restriction
u = 0, in which case the possibility of having u p 0
arises as an alternative. It can also happen that only
some of the individual restrictions are incompatible
with their corresponding unrestricted forecasts. In such
a case we require individual compatibility tests that
enable us to appreciate the contribution of each
755
individual restriction to the JCT. These last two issues
will be studied in the following sections.
3. Additivity of the compatibility test
By what follows we shall refer to K as the JCT and
the test of a single restriction will be called an SCT. In
this section we analyze how the M SCTs are related to
K. Then we propose to modify the matrix u to
achieve additivity.
ð8Þ
3.1. Decomposition of the JCT into single tests
By analogy with Eq. (8), for the ith restriction,
with i = 1, . . ., M, we get
di uri  Ci Y F;H ¼ Ci CeF þ ui ;
where C i is the ith row of C (representing the linear
combinations of Y F involved in this single restriction)
and u i ~N(0, r ii,u ). From this we know that d i ~N
(0, x ii ), where x ii u Ci Y F,H CVi + r ii ,u is the iith
ð9Þ element of 6. As before, we define the SCT of the
ith restriction as
Kii ¼ di x1 1 2 f 2
ii di ¼ xii di v1 :
The following proposition states that, in general,
the JCT cannot be obtained simply as the sum of the
SCTs.
Proposition 1. Let Y F,H be a symmetric real positive-
definite matrix and C be of full row rank. Then, the
JCT can be written in terms of the SCTs as
X
M X
M
K¼ x ii xii Kii þ 2
x̄ x ij xij Kij ;
x̄ ð10Þ
i¼1 ibj
where x ij and x̄ij denote the (i, j)-th elements of 6
and 6 1 respectively. Besides, K ij u x ij 1d i d j if
x ij p 0 and K ij u 0 if x ij = 0, for i, j = 1, . . ., M.
Proof. Since rank (C) = M, CY F,H CV is a symmetric
positive-definite matrix. Now, u is the symmetric
and positive-semidefinite covariance matrix of u, thus
6 = CY F,H CV + u is also a symmetric positive-
definite matrix. This means that det(6) N 0, so that
6 1 exists. Hence, substituting the elements of d and
6 1 directly into Eq. (9), the JCT takes the indicated
form. 5
756 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770
When the matrix 6 involved in the distance vector
d is diagonal, we get
K ¼ x1 2 1 2
11 d1 þ . . . þ xM M dM ¼ K11 þ . . . þ KM M
that is, the SCTs add up to the JCT.
The converse is not true, that is, K = K 11 +
. . . + K MM does not imply that 6 is diagonal. To see
this let us consider a simple case with M = 2, so that
K = K 11 + K 22 = (x 22 d 12 + x 11 d 22 ) / (x 11 x 22 ). Now,
from Eq. (10) the general form of the JCT is K =
(x 22d 12 + x 11d 22  2x 21d 1d 2) / (x 11x 22  2x 21). Thus,
solving for the correlation term x 21 = x 12, the JCT
will be the sum of the SCTs only if x 21 = 0 or
x 21 = 2d 1d 2 / (K 11 + K 22). Therefore, this example
shows that there exists a non-diagonal matrix 6 that
yields K = K 11 + K 22.
3.2. Decomposition of K
It should be clear that many factors that affect the
economic system are not under the control of the
government and usually imply some uncertainty of the
targets. The problem in our case is how to assign this
uncertainty. One possibility is that u be given by the
same external source that provides the linear restric-
tions, for instance when it comes from a competing
forecasting method. Otherwise, choosing this matrix
should be done with care in order not to lose precision
in Y RF,H . For example, Guerrero and Peña (2003)
considered that this matrix has a diagonal form and
chose its elements to make the restrictions compatible
with the historical data. Here we propose another
alternative.
We suggest modifying the matrix u to cancel out
model dynamics and interactions between restrictions,
thus obtaining additivity of the SCTs with respect to
the JCT. To that end, let us define
   
Ru ðaÞu adiagðQÞ þ diag CRY F;H CV  CRY F;H CV ;
for some real parameter a N 0 and where Q is an
M  1 vector whose elements are denoted by q i N 0
for i = 1, . . ., M.2 These q i values will be interpreted as
weights associated with the uncertainty of the
2
If A= (a ij ) is an0 n  n matrix 1 and a = (a 0 i ) is an n1 1 vector, the
a11 . . . 0 a1
operators diagðAÞ ¼ @ . . . . . . . . . A and diagðaÞ ¼ @ . . . . . . . . . A produce n  n
0 . . . ann 0 . . . an
diagonal matrices.
restrictions involved, as will be seen in the following
section. More precisely, matrix (11) takes the form
0 1
aq1  x12 . . .  x1M
R u ð aÞ ¼ @ . . . ... ... ... A:
 xM 1  xM 2 . . . aqM
Let k 1(a) z k 2(a) z . . . z k M(a) denote the eigen-
values of this matrix and define the modified matrix as
R4u uRu ða4Þ s:t: kM ða4Þ ¼ 0 ð12Þ
which, by construction, is a symmetric positive
semidefinite matrix and defines a covariance matrix.
The function k M (a) may have more than one zero, in
which case the solution of Eq. (12) will not be unique.
Below we shall see that this situation is avoided by the
form of u (a). In fact, this follows from the linear
dependence of the eigenvalues on the parameter a.
Let v i be an eigenvector of u corresponding to k i ,
for i = 1, . . ., M, thus
Ru vi ¼ ki vi ; with vi Vvi ¼ 1:
The differential of this expression is
d ðRu vi Þ ¼ diagðQÞd ðaÞvi þ Ru d ðvi Þ
¼ d ðki Þvi þ ki d ðvi Þ ¼ d ðki vi Þ
see Magnus and Neudecker (2002). Premultiplying by
vVi we have
vi VdiagðQÞd ðaÞvi þ vi VRu d ðvi Þ
¼ vi Vd ðki Þvi þ vi Vki d ðvi Þ:
Hence, since u is symmetric, we get
vi VdiagðQÞd ðaÞvi ¼ vi Vd ðki Þvi ;
and
d ðki Þ ¼ ½vi VdiagðQÞvi d ðaÞ;
because the eigenvector v i is normalized by vVv i i =1
and a and k i are scalars.
ð11Þ The last expression implies that
dki ðaÞ
¼ c for i ¼ 1; . . . ; M ; ð13Þ
da
where c = vVi diag(Q) and v i N 0 is a constant. So the
eigenvalues are linear increasing functions of a, which
implies that a* is uniquely determined. This fact
allows us to find the zero of k M (a) by simply finding
... 0
the point where the straight line defined by Eq. (13)
crosses the a axis.
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770 757

Remark 1. It should be noticed that additivity is As is usual in Monte Carlo experiments, we lose in
achieved by introducing uncertainty in the restrictions. generality in order to gain in understanding the small
Although we introduce the minimum amount of sample properties of a particular case. Suppose that at
uncertainty in the sense of Eq. (12), it could be that the end of the current year the government announces
an incompatible binding restriction becomes a com- the economic targets for next year and that the data are
patible unbinding restriction. This fact may be quarterly (data for the current quarter are still
considered a drawback of this proposal; nevertheless unavailable). We consider three typical linear combi-
it may be useful in the practical applications as can be nations of Y F for the stochastic restriction:
seen in the empirical application of this paper.

0 0 ... 0 0 1 0
C1 ¼ ;
4. A feasible JCT for estimated processes 0 0 ... 0 0 0 1
0 1 1 1
We study the finite sample properties of the JCT by 0 0 0 ... 0
B 4 4 C
C2 ¼ @ A ð16Þ
way of a numerical simulation of a bivariate system. 1 1
Three typical restrictions and four sample sizes T = 20, 0 0 0 ... 0
4 4
50, 100 and 200 are used. This leads us to consider first
that the process is estimated, and second that the and
covariance matrix of the distance vector should take 
into account the sample size. Thus, a feasible JCT for 1 0 0 0 ... 0 0 1 0
C3 ¼ :
estimated processes is obtained and its finite sample 0 1 0 0 ... 0 0 0 1
properties are studied in a similar way. Another ð17Þ
simulation with a model used in the following section
The C1 matrix puts a restriction at the end of the
for the Mexican economy was carried out for the JCT
forecast horizon. This is useful when the government
and for the feasible JCT in order to validate the
announces the value of a variable for the end of the
empirical application of these tests. We did not carry
following year. C2 applies when the government
out a simulation study for the univariate situation
announces the average value for a variable in the
because Box and Tiao (1976) did that for a statistic
following year. If the 1 / 4 values are replaced by 1 the
similar in nature to K. In fact, they found that the
restriction is on the sum of the variable. This
distribution of their statistic was v 2 as in Eq. (9).
specification will be used in the empirical illustration
However, for an estimated AR( p) model with n
when dealing with the inflation rate (in logs). Finally,
observations, the bestimation errors inflate the mean
C3 restricts the difference of values for the end of the
value of v 2. . . by a factor approximating 1 + ( p / n).Q
year. This is used when dealing with the rate of
growth of a variable. We apply this restriction to GDP
4.1. A Monte Carlo study of the JCT
(in logs) in the empirical illustration.
The Monte Carlo experiment was done with the
We assume that the process follows the bivariate
following algorithm and program routines were
VAR(2) model given in Lütkepohl (1991, equations
written in Matlab 6.5-Release 13 (MathWorks, Inc.
3.2.25–3.2.26).
Software). Given C, u , T, H, and p:
  
yt ¼
0:02
þ
0:5 0:1
yt1 þ
0 0
y þe ; 1. Generate a series {y t }T+Ht=p+1 with Eqs. (14) and (15)
0:03 0:4 0:5 0:25 0 t2 t and initial values E(y t ) = (0.07027, 0.15135)V. Follow-
ð14Þ ing the partition used by Lütkepohl we have a time
series y 1, . . ., y T of length T as well as a presample
with error covariance matrix y p+1, . . ., y 0. The future values of the series Y F are
 given by y T, . . ., y T+H .
9 0 2. Estimate the VAR( p) for the process. Compute Y F,H
Re ¼  104 : ð15Þ
0 4 and Y F,H .
758 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770

3. Generate a random vector u~N(0, u ). For binding Table 2

restrictions u = 0 take u = 0. p a values for the compatibility test with unbinding restrictions
4. Compute R = CY F + u. Sample Combination matrix (H = 5)
size
5. Compute 6, d and K. C1 C2 C3
p .10 p .05 p .01 p .10 p .05 p .01 p .10 p .05 p .01
Steps 1–5 were replicated N = 1000 times for p = 2, 20 .27 .19 .08 .35 .28 .17 .23 .17 .08
T = 20, 50, 100 and 200, with three restrictions of the 50 .14 .08 .03 .15 .11 .04 .12 .06 .02
forms (16) and (17) and considering H = 5, with 100 .10 .05 .01 .12 .06 .01 .11 .06 .01
u = 0 fixed. Since K is distributed as v 22 we know 200 .10 .04 .01 .10 .06 .02 .10 .05 .01
that P(K z v 22(a)) = a. In practice we should expect the Proportion of K calc values exceeding the v 22 (a) quantile with u*
proportion of sample values of the K that exceed the obtained with N = 1000 Monte Carlo replications of Eqs. (14) and
(15).
v 22(a) quantile to be close to a. We denote this
proportion by p a . Table 1 reports these values for
a = .10, .05 and .01. The p a values for T = 20, 50 and tainty CY F (see Eq. (7)), and the linear combinations
100 are higher than a. Thus the v 22 distribution leads of the unrestricted forecast CY F,H . Table 2 reports the
to over-rejection of the compatibility hypothesis. For p a proportions for this experiment.
example, when T = 20, C = C1 and a = 0.10, u = 0 In summary, the p a proportions are higher than a for
and H =5. The reported value says that 32% of the trials small samples. This result does not depend on whether
produced K statistics above v 22(.10). Notice that for the restriction is binding or unbinding, but assigning
T =20 and 50, the sampling distribution of K depends on uncertainty in this way leads the p a values closer to a
the restriction imposed. For example, the pa values with than those reported for the binding case. In the following
C2 are higher than those for C1 and C3. This may be due section we propose an adjustment for the JCT that also
to the fact that using the C2 matrix involves restricting takes into account the fact that the model parameters are
more values of Y F than using C1 or C3 does. estimated. As we will see in the next subsection, this
This experiment was repeated with the matrix u* yields an additional improvement for small samples.
in place of u . Since this matrix is computed once the
VAR has been estimated, the algorithm described 4.2. The VAR forecast for estimated processes
above must be changed. Given C, T, H and p, Step 3
should be replaced by 3V. Compute Q as a diagonal The optimal (h)-step forecast of Eq. (1) is given
matrix, with the squared values of m = C(Y F  Y F,H ) in by Eq. (4); however, if the true parameters H =
its main diagonal. Compute u* as defined in Eq. (12). (K, P 1, . . ., P p ) are replaced by their estimators
Generate a random vector u ~ N(0, u*). Ĥ= (K̂, P̂1, . . ., P̂p ) we get the forecast
The Q matrix computed this way assigns uncer-    
tainty to each restriction in accordance with the E yT þh jY ¼ K̂
Ê KDT þh þ P̂ E yT þh1 jY þ . . .
P 1 Ê
 
distance between the linear restriction without uncer- þ P̂ E yT þhp jY for h ¼ 1; . . . ; H;
P p Ê

Table 1 where Ê(y T+hi |Y) = y T+hi for i z h. By calling Ŷ F,H

p a values for the compatibility test with binding restrictions the Ê(Y F |Y) vector we have
Sample Combination matrix (H = 5)    
size Y F;H ¼ Y F  Y F;H þ Y F;H  Ŷ
Y F  Ŷ Y F;H
C1 C2 C3  
p .10 p .05 p .01 p .10 p .05 p .01 p .10 p .05 p .01
¼ CeF þ Y F;H  Ŷ Y F;H : ð18Þ
20 .32 .25 .14 .48 .40 .26 .30 .21 .11 Under general conditions for the process y t , Dufour
50 .18 .12 .05 .22 .14 .05 .17 .10 .04
100 .14 .09 .02 .14 .09 .03 .14 .09 .02
(1985) proved that the forecast errors have zero
200 .10 .06 .02 .10 .05 .01 .10 .06 .01 conditional mean, E(Y F  Ŷ F,H |Y) = 0, so the forecast
Proportion of K calc values exceeding the v 22 (a) quantile with u = 0
is conditionally unbiased even if the parameters are
obtained with N = 1000 Monte Carlo replications of Eqs. (14) and estimated. On the right side of Eq. (18), all the e t
(15). contained in e F correspond to periods t N T, whereas
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770 759

all the y t contained in the second term correspond to 4.4. The feasible JCT with an estimated process
t V T, and therefore the two terms are uncorrelated.
Thus the MSE of Ŷ F,H becomes A compatibility test must consider the distribution
   of the distance vector of an estimated process as well,
MSE Ŷ Y F;H uRŶY F;H ¼ RY F;H þ MSE YF;H Ŷ
Y F;H Þ:
that is
ð19Þ
d̂duR  CŶ
Y F;H
To evaluate the second term on the right hand side of hpffiffiffiffi i pffiffiffiffi
this equation we need the distribution of H. Since small ¼ CCeF þ C T Y F;H  Ŷ
Y F;H = T þ u: ð22Þ
sample distributions of VAR estimators are not avail-
able, we cannot hope to get more than an asymptotic From the normality assumption of e F, u and Eq.
distribution for MSE(Y F,H  Ŷ F,H ). To obtain this we (20) we know that, approximately
 
proceed as in Lütkepohl (1991, Section 3.5). So, let d̂d fN 0; CRŶY F;H CV þ Ru :
b u vec(H) and b u vec(H) be its OLS estimator, 
whose asymptotic covariance matrix is bˆ, and Thus, for estimated processes the statistic d̂dV CRŶY F;H
pffiffiffiffi  d   CV þ Ru Þ1 d̂d is plausible. Consistent estimators of the
T b̂ b  b Y N 0; Rb̂b : MSE matrices are obtained by replacing the unknown
parameters by their estimators. The resulting estimator
Then, under quite general conditions we know that, ˆ Ŷ .
of Ŷ F,H will be denoted by 
conditional on Y, F,H

 In addition to the estimated process adjustment we

pffiffiffiffi  d BY F;H BY VF;H should adjust the statistic for using an estimated rather
T Ŷ Y F;H  Y F;H jY Y N 0; Rb̂
b :
BbV Bb than a known covariance matrix. This leads us to
ð20Þ divide the statistic by its degrees of freedom and then
use an F distribution to make inferences. Finally, the
From here, an approximation to the MSE(Y F,H  Ŷ F,H ) feasible JCT gets defined as
is T 1˜ , where
 1
BY F;H BY VF;H X d̂d=M fFM ;T M p1
K ¼ d̂d VX̄
K̄
R¼E
R̃ RFb̂
b : ð21Þ
BbV Bb where 6̄ u Cˆ ŶF,H CV + u .
Therefore Eq. (19) becomes Thus, R is not in the compatibility region at level
of significance a if
RŶY F;H cRY F;H þ T 1 R̃
R;
K calc zFM ;T M p1 ðaÞ
K̄
˜ is derived in Appendix
the explicit expression for 
A. with F M,TMp1(a) being the (1  a)th quantile of the
F M,TMp1 distribution. Since MK̄ has the same
4.3. The restricted forecast with an estimated process quadratic form as K, the results obtained in Section
3 remain valid. For example, the modified matrix (12)
The optimum restricted forecast with an estimated makes use of
process is
  h   i
Y RF;H ¼ Ŷ
Ŷ Y F;H þ Â
A R  CŶY F;H Ru ðaÞu adiagðQÞ þ diag CR̂
R ŶY F;H CV  CR̂
R ŶY F;H CV
where  X 1 and X̂
A ¼ RŶY F;H CVX̂ XuCRŶY F;H CV þ Ru . Its and the additivity property of the feasible JCT yields
MSE is given by
  h   i K ¼ K̄
M K̄ K 11 þ . . . þ K̄
K MM :
MSE ŶY RF;H uE Y F  Ŷ Y RF;H Y F  ŶY RF;H VjY ; R
  4.5. A Monte Carlo study for the feasible JCT
¼ I  Â
AC RŶY F;H :
As before, we know that Ŷ RF,H is at least as precise Table 3 presents the proportion, p a , of sample
as Ŷ F,H . These expressions are derived in Appendix B. values of K̄ exceeding the quantile F M,TMp1(a).
760 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770

Table 3 Table 5
p a values for the feasible compatibility test with binding restrictions p a values for the estimated Mexican system
Sample Combination matrix (H = 5) Covariance matrix (H = 8) u = 0 u*
size C1 C2 C3 Statistic p .10 p .05 p .01 p .10 p .05 p .01
p .10 p .05 p .01 p .10 p .05 p .01 p .10 p .05 p .01 K .41 .32 .18 .26 .18 .07
20 .23 .14 .05 .30 .24 .13 .18 .11 .04 K̄ .10 .04 .01 .04 .01 .00
50 .13 .07 .02 .17 .11 .03 .12 .07 .01 Proportions of K and K̄ for u = 0 and u * values exceeding their
100 .12 .06 .02 .15 .09 .03 .11 .06 .01 theoretical quantiles for the estimated VEC of the Mexican
200 .11 .07 .01 .12 .06 .01 .11 .05 .01 economy with restrictions given by Eq. (23) and forecast horizon
of H = 8. The Monte Carlo experiment was done for N = 1000
Proportion of K̄ calc values exceeding the F M,TMp1(a) quantile
replications.
with u = 0 obtained with N = 1000 Monte Carlo replications of
Eqs. (14) and (15).

should be taken into account when drawing conclu-

The p a values are closer to a than those reported for sions from a restricted forecasting analysis.
the JCT in Table 1. For sample sizes T z 50 it can
be seen that F M,TMp1 approximates the distribu-
4.6. A simulation study for the Mexican data example
tion of K̄ well. As before, it is also clear that the
sampling distribution of K̄ depends on the restric-
An empirical application for the Mexican economy
tion used.
is presented in the following section to illustrate the
The Monte Carlo experiment with K̄ was also
methodology. The VAR(3) model employed includes
done for the unbinding case with u*. Table 4 six variables plus a constant and some dummy
reports the corresponding proportions for this case.
variables. Since the system turned out to be CI(1, 1),
The magnitudes of the p a values are similar to those
the estimation was done with a VEC representation.
in Table 3. So, the F M,TMp  1 distribution can also
After adjusting by the order of the process the sample
be considered a good approximation for the finite
size is equal to T = 52.
distribution of K̄ for sample sizes T z 50, but its
A numerical algorithm similar to that described
sampling distribution depends on the type of
above was used for the VAR process whose estimated
restriction employed.
results appear in Subsection 5.2 with N = 1000, T = 52,
Remark 2. The simulations provide some evidence p = 3, the C matrix given in Eq. (23), M = 3, and H = 8.
that the feasible JCT statistic produces p a values Table 5 shows the p a proportions of K and K̄ for
closer to the nominal ones. However, the improve- u = 0 and u*.
ment may not be enough for some applications. This It is clear that K̄ has better performance than K,
even though the p a values for K̄ with u* are much
smaller than a. Apparently the uncertainty distorts
the distribution of the JCTs for small samples.
Table 4 Nevertheless, the p a values for K̄ with u* are
p a values for the feasible compatibility test with unbinding closer to a than those for K. So, the feasible JCT
restrictions
performs better than the JCT for the VEC model of
Sample Combination matrix (H = 5) the Mexican economic system.
size
C1 C2 C3
p .10 p .05 p .01 p .10 p .05 p .01 p .10 p .05 p .01
5. Empirical illustration
20 .16 .10 .04 .21 .15 .06 .13 .08 .01
50 .10 .06 .02 .12 .06 .02 .10 .06 .01
At the end of 2002 the Mexican Government
100 .09 .05 .01 .09 .05 .01 .09 .05 .01
200 .09 .05 .01 .10 .05 .01 .10 .05 .01 published the economic targets for 2003, see SHCP
(2002). There, it was foreseen that the rate of growth
Proportion of K̄ calc values exceeding the F M,TMp1(a) quantile
with u* obtained with N = 1000 Monte Carlo replications of Eqs. of GDP would move from 1.7% in 2002 to 3.0% in
(14) and (15). 2003. The annual inflation rate was targeted to be
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770 761

reduced from 4.9% in 2002 to 3.0% in 2003 and the 5.1. Order of integration
trade balance deficit was supposed to move from
 15,234.6 ( 2.4% of GDP in 2002) to  18,035.5 The order of integration of the series was decided
( 2.8% of GDP in 2003). by Augmented Dickey–Fuller (ADF) tests. The
The data set consists of 55 quarterly observations. augmented regression model included a constant,
When the Mexican Government announced the centered dummies for PMEXt , LUNMPt , TRDBt
targets the data available ran up to 2002:III. So, the and PUSAt , and a deterministic trend for LGDPt
estimation period covers data from 1989:I to 2002:III. and LMONBt . The general equation was
The data are available from the web site at ITAM
X
3 X
p
with the address http://allman.rhon.itam.mx/%7E Dyt ¼ a þ c0 t þ ci Dit þ d0 yt1 þ dj Dytj
guerrero/Series_VEC.pdf. A description of the vari- i¼1 j¼1
ables employed follows.
Gross domestic product (LGDP). Measured in þ error
thousands of Mexican pesos at constant prices of 1993. In order to account for the Mexican crisis of 1995
Source: Instituto Nacional de Estadı́stica Geografı́a e we included two dummy variables for the first and
Informática, Sistema Nacional de Cuentas Nacionales. second quarters of that year. Table 6 shows the results
The series is log transformed, LGDPt =ln(GDPt ). of the ADF tests with and without dummy variables,
Mexican inflation rate (PMEX). First difference of that is, with and without accounting for structural
log Consumer Price Index (ipcmex) with base change (SC).
1994 = 100. Source: Bank of Mexico. The Consumer The order of the autoregression p was selected to
Price Index is monthly and the quarterly series are the guarantee no autocorrelation in the residuals. The s
values at the end of the quarters. PMEXt = ln(ipc- statistic allows one to test H 0: d 0 = 0. Critical values
mext )  ln (ipcmext1). do not consider intervention dummy variables to
Unemployment rate (LUNMP). Source: Instituto account for the 1995 crisis. Except for domestic
Nacional de Estadı́stica Geografı́a e Informática, inflation (PMEX) in levels, the order of integration of
National Urban Employment Survey. The data are log the variables did not depend on the inclusion of the
transformed, LUNMPt = ln(UNMPt ). dummy variables, but there still was some doubt
Real demand of money (LMONB). Currency held whether or not inflation is stationary. However, since
by the public plus domestic currency and checking the test result can be distorted by the inclusion of
accounts in resident banks. This is a monthly series dummy variables, we assumed that PMEX is I(1).
given in nominal terms in thousand of Mexican
pesos (basemon). The quarterly series is obtained 5.2. VEC estimation
by averaging the monthly values and is deflated by
ipcmex. Source: Bank of Mexico. The series is log The VAR model included the six economic varia-
transformed, LMONBt = ln(basemont / ipcmext ). bles previously described, y t = (PMEX t , LGDP t ,
Trade balance deficit (TRDB). Defined as income
minus expenditure of the foreign sector. This is a Table 6
quarterly series given in millions of dollars (DEF). ADF unit root test results
Source: Bank of Mexico. The series is transformed by Variable H 0: I(1) H 0: I(2)
dividing it by 10,000 to homogenize the data, Without SC With SC Without SC With SC
TRDBt = DEFt / 10,000.
p s p s p s p s
US inflation rate (PUSA). First difference of the
log US Consumer Price Index (ipcusa) with base LGDP 0 2.16 0 3.06 0 6.50* 0 10.21*
PMEX 3 2.43 0 5.70* 1 7.76* 0 9.34*
1982–84 = 100. Source: US Department of Labor, LUNMP 1 1.43 1 2.42 0 5.52* 0 7.12*
Bureau of Labor Statistics. The US Consumer Price LMONB 1 1.68 2 0.25 0 4.75* 2 6.61*
Index is monthly and the quarterly series are the TRDB 0 2.18 4 1.66 0 6.56* 1 7.43*
values at the end of the quarters, PUSA t = PUSA 3 2.81 – – 1 9.40* – –
ln(ipcusat )  ln(ipcusat  1). Note: * indicates rejection of H 0 at the 5% significance level.
762 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770

Table 7 The estimation results are as follows (t-values in

Johansen cointegration analysis parentheses)
NULL Trace Crit Crit Eigen Crit Crit 
Statistic 95% 99% Statistic 95% 99%  0:279 0:642  0:467  1:607  2:659 0:049
a¼
â ;
ð  1:71Þ ð3:48Þ ð  1:21Þ ð  1:28Þ ð  1:82Þ ð0:79Þ
r V0 124.23** 95.75 104.96 46.58** 40.08 45.87 b ¼0
b̂ ð1  0:038 0:058  0:036 0:017  0:221Þ; 1
   0:028  0:113 0:102
r V1 77.65* 69.82 77.82 38.34* 33.88 39.37 B ð  2:13Þ ð9:02Þ ð5:30Þ C
B C
r V2 39.31 47.86 54.68 20.12 27.59 32.72 B   0:071  0:057  0:084  0:060  0:084 C
B C
B ð  3:63Þ ð  3:91Þ ð  7:27Þ ð  4:23Þ ð  3:84Þ C
r V3 19.19 29.80 35.46 14.63 21.13 25.87 B
B 0:018
C
B  0:112  0:162  0:143  0:144  0:135 C C
r V4 4.56 15.49 19.94 4.41 14.26 18.52 B ð 2:38Þ ð  2:74 Þ ð  5:30Þ ð  5:95 Þ ð  4:84 Þ ð  2:96Þ C
Lˆ ¼ B
B 
C;
C
r V5 0.15 3.84 6.64 0.15 3.84 6.64 B 0:348 0:213  0:637 0:372 C
B ð2:59Þ ð2:13Þ ð6:57Þ ð2:50Þ C
B C
Note: * indicates rejection at the 5% level and ** indicates rejection B   0:235  0:351  C
B C
B ð2:02Þ ð3:12Þ C
at the 1% level. B C
@  0:017  0:012   A
0 ð2:60Þ ð3:00Þ 1
 0:350     
B ð  2:16Þ C
B C
B  0:588  0:525  0:265  0:046   1:193 C
LMONBt , TRDBt , LUNMPt , PUSAt )V, a constant, B
B ð  3:20Þ ð  3:56Þ
B ð3:13Þ ð2:03Þ ð2:32Þ
C
C
C
B C
centered dummy variables to account for seasonal C1 ¼ B
Ĉ 4
B






 
 0:548 0:341 8:749 C
 C;
B C
effects, and two dummy variables to account for the B
B ð  3:57Þ ð2:27Þ ð2:50Þ C C
B 3:795      C
1995 crisis, i.e. D t = (const, S 1,t , S 2,t , S 3,t , I 95:I,t , I 95:II,t )V. B C
@       0:335 A
Although the US inflation rate is basically exoge- 0
 0:225     0:804
ð
1  1:97Þ

nous to the Mexican economy, it entered the model B ð  2:22Þ

B ð2:09Þ CC
B  0:224      C
as an endogenous variable because in the estimation B C
B ð  1:95Þ C
B C
stage it was found that no other variable in the C 42 ¼ B
Ĉ B       C C;
B     0:313  8:230 C
B C
system affects PUSA significantly. The system B
B ð  2:42 Þ ð2:76Þ C
C
@       A
became CI(1, 1) and an integrated VAR(3) model      
provided a reasonable representation for the system
in levels. Therefore the system was estimated with with R 2 equal to 0.89, 0.96, 0.92, 0.78, 0.70, and 0.81
the following VEC model for DPMEX, DLGDP, DLMONB, DTRDB, DLUNMP
and DPUSA, respectively. The matrices  ˆ , /̂*1, and
/̂ 2
* show only the numerical values of those elements
Dyt ¼ LDt þ abVyt1 þ C41 Dyt1 þ C42 Dyt2 þ et found significant at the 5% level.

Fig. 1. Cointegration relationship.

 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770 763

Fig. 2. Observed and estimated series in levels and corresponding residuals of first differences, with a two standard error band.

The following matrix shows the contemporaneous There is negative correlation between the Mex-
residual correlations. By symmetry, only the lower ican inflation rate and the real demand for mon-
diagonal part is shown. ey, as would be expected. There is also positive
correlation between the US inflation rate and the
2 PMEX LGDP LMONB TRDB LUNMP PUSA 3 trade balance deficit. In Table 7 we report the re-
PMEX 1:00
LGDP 66 0:15 1:00 7
7 sults ofJohansen tests (see for instance, Johansen,
LMONB 6
6  0:39 0:09 1:00 7:
TRDB 6  0:08 0:01 0:23 1:00
7
7 1988).
6 7
LUNMP 4  0:32  0:11  0:06  0:17 1:00 5 At the 5% significance level there are two coin-
PUSA 0:24 0:09  0:22 0:27  0:19 1:00 tegrating relationships and only one at the 1% level.
764 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770

Table 8 where 0 = (0, . . ., 0) is the 1  6 zero vector and

Feasible compatibility tests for Mexican economic targets e i = (0, . . ., 1, . . ., 0) is a 1  6 vector with 1 in the ith
Covariance matrix u = 0 u* position and zeros elsewhere. The restricted forecasts
Statistic Value p-value Value p-value are obtained for both the binding case (u = 0) and the
K̄ 1 0.12 0.73 0.10 0.76 unbinding case with u*.
K̄ 2 0.30 0.59 0.19 0.66 Table 8 reports the feasible JCT for u = 0 and
K̄ 3 0.08 0.77 0.07 0.79 u* as well as their SCTs. The p-values were obtained
K̄ 0.15 0.93 0.12 0.95
from the F 3,42 distribution. Neither the binding nor the
Single and joint feasible compatibility tests for the Mexican unbinding restrictions were rejected by the compati-
economic targets in 2003, with u = 0 and *.u Targets in 2003:
bility tests at the usual significance levels. In the
3.0% rate of growth for GDP, trade balance deficit of 15, 234.6
(2.4% of GDP) and 3.0% annual inflation rate. unbinding restriction case it can be verified that
3K̄ = K̄ 1 + K̄ 2 + K̄ 3. Even though none of the single
We decided to use only one cointegration relationship. restrictions is incompatible, it is interesting to see that
The resulting equation was the restriction on inflation produces an SCT statistic
(K̄ 2 = 0.19) that contributes 53% to the feasible JCT
value (3K̄ = 0.36).
et ¼ PMEX t  0:04LGDPt þ 0:06LMONBt
The estimated matrix employed for getting addi-
 0:04TRDBt þ 0:02LUNMPt  0:22PUSAt : tivity and its eigenvalues are

the graph of which is shown in Fig. 1. The observed 0 1

0:0006  0:0002 0:0024
and fitted series of the VEC model, as well as their B C
residuals, are shown in Fig. 2. Ru* ¼ @  0:0002 0:0001  0:0013 A and
0:0024  0:0013 0:3728
0 1
5.3. Forecasts restricted to fulfill the targets 0:3728
B C
L ¼ @ 0:0006 A:
The economic targets described above are written
as 0:0000

LGDP2003:IV  LGDP2002:IV ¼ lnð1:03Þ

PMEX 2003:I þ PMEX 2003:II þ PMEX 2003:III
þ PMEX 2003:IV ¼ lnð1:03Þ
TRDB2003:I þ TRDB2003:II þ TRDB2003:III
þ TRDB2003:IV ¼  18; 035:5=10; 000:
Thus, the linear stochastic restriction (7) considers
an H = 8 period-ahead forecast and is specified by
0 1
lnð1:03Þ
B C
R ¼ @ lnð1:03Þ A and
 18; 035:5=10; 000
0 1
 e2 0 0 0 e2 0 0 0
B C
C ¼ @0 e1 e1 e1 e1 0 0 0 A
0 e4 e4 e4 e4 0 0 0
Fig. 3 shows the unrestricted and binding restricted
(u = 0 ) forecast paths with their 90% probability
intervals for each of the six economic variables. The
corresponding paths for the unbinding restricted (u*)
forecasts are shown in Fig. 4.
Table 9 summarizes the 2003 Mexican Govern-
ment economic targets, the unrestricted forecast, the
binding restricted forecast (u = 0 ) and the unbind-
ing restricted forecast (u*). Here, we can appreciate
that the Government is taking into account the
effects of its future monetary policies. For instance,
the unrestricted forecast leads to economic growth
and inflation rates of 4.01% and 5.28% respectively.
Thus, the monetary policy that causes inflation to
stay at around 3.00% yields a decrease in economic
ð23Þ
growth. Notice that the binding restricted forecasts
attain the targets exactly. These values are relaxed
by the unbinding restricted forecasts if uncertainty is
allowed.
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770
Fig. 3. Binding restricted forecasts with 90% probability intervals and unrestricted forecasts (with origin at 2002:III).

765
 766
N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770
Fig. 4. Unbinding restricted forecasts with 90% probability intervals and unrestricted forecasts (with origin at 2002:III).
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770 767

Table 9 ing compatibility between targets and unrestricted

Forecasts of the Mexican system
Variables Targets Forecasts
Unrestricted Restricted Restricted
with u = 0 with u*
GDP 3.00% 4.01% 3.00% 3.14%
PMEX 3.00% 5.28% 3.00% 3.53%
UNMP – 2.63% 2.68% 2.66%
MONB – 9.47% 9.92% 9.67%
DEF  18,035.5 16,701.4 18,035.5 17,730.7
VEC forecasts results for the Mexican system in 2003.
5.4. Forecasts restricted by unrealistic targets
Since neither the JCT nor the SCTs provided
evidence of incompatibility, it is difficult to appreciate
the relevance of using u* to detect the contributions of
the single restrictions to the JCT. We therefore consider
an unrealistic case in which the targets are: 10.5% rate of
growth for GDP in 2003, trade balance deficit of  32,
151 ( 5.0% of GDP) and 3.0% annual inflation rate.
Table 10 reports the feasible JCT and the corresponding
SCTs for this situation. Considering a significance level
of 10% we should reject the joint compatibility
hypothesis for the binding case (u = 0 ). As we can
see, the sources of incompatibility are GDP and TRDB,
but at this point we should be cautious when interpreting
these results because of the lack of additivity of the SCTs
with respect to the JCT. On the other hand, additivity can
be obtained at the cost of introducing uncertainty in the
sense of (12), yielding a JCT value lower than that
obtained with u = 0 . Thus, some binding restrictions
which are incompatible may become compatible when
introducing uncertainty. However, introducing uncer-
tainty by u* does not always make the restrictions
compatible. This is the case reported in Table 10, where
the restriction on GDP becomes compatible, while the
restriction on TRDB remains incompatible. The joint
compatibility test also rejects compatibility. Of course,
these results depend on the significance level employed.
Finally, from the additivity of the JCT with u* we can
see that the main source of incompatibility is the TRDB
restriction, which amounts to 58% of the JCT value.
6. Conclusions
This paper presents a complement to the multi-
variate restricted forecasting methodology for study-
forecasts from a VEC model. The compatibility test
can be used to detect structural breaks during the
forecast horizon and to monitor the attainment of
economic targets. The restrictions can be binding or
unbinding. In the latter case the uncertainty may
come from subject matter knowledge or, as is
proposed here, can be deduced from the data. We
propose to compute the covariance matrix associated
with the restrictions, so that it cancels out model
dynamics and interactions between restrictions, and
thus obtain additivity of the SCTs with respect to
the JCT, perhaps at the cost of inducing compati-
bility. This fact may be considered a drawback of
our proposal.
Since the compatibility test for multivariate
series relies on asymptotic theory, a numerical si-
mulation was performed to check the small sample
performance of the JCT. Then, a feasible JCT that
takes into account estimated parameters was
obtained and its finite sample properties were also
studied. In general, for finite sample sizes the
feasible JCT proved to perform better than the
original JCT.
The proposed methodology was illustrated with a
six-variable Mexican economic system with quarter-
ly data. Implementation of the model focussed on
the economic targets for GDP, inflation rate and
trade balance deficit for 2003. A numerical simula-
tion of this system was carried out to validate the
use of the feasible JCT in this situation. It turned
out that the economic targets were compatible with
the unrestricted forecasts. Finally, some unrealistic
targets were considered to illustrate an incompati-
bility situation.
Table 10
Feasible compatibility tests for unrealistic economic targets
Covariance matrix u = 0 u*
Statistic Value p-value Value p-value
K̄ 1 4.65 0.04 2.69 0.11
K̄ 2 0.30 0.59 0.29 0.60
K̄ 3 11.27 0.00 4.08 0.05
K̄ 4.71 0.01 2.35 0.09
Single and joint feasible compatibility tests for unrealistic
economic targets in 2003, with u = 0 and u = u*. Unrealistic
economic targets in 2003: 10.5% rate of growth for GDP, trade
balance deficit of 32, 151 (3.0% annual inflation rate).
768 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770

Acknowledgments After some algebraic calculations, it can be shown

that
We thank three anonymous referees, an Associate 
BvecðBÞ 0ðH1ÞnðnþkpÞ
Editor and Prof. Jan De Gooijer, for their insightful ¼ ĨI  JV and ĨI ¼
comments and suggestions that helped improve the BbV IðnþkpÞðnþkpÞ
presentation of this paper. We are also thankful for the
are k  k(n + p) and (Hn + kp)  (n + kp) matrices.
comments provided by Prof. Michael Creel on a
Now, we can write
previous version of this article. Vı́ctor M. Guerrero
acknowledges the support provided by Asociación  
BE yT þh jY X
h1
Mexicana de Cultura, A.C. for carrying out this work. ¼ ZT VðBVÞh1i ĨI  JBi JV:
Nicolás Gómez acknowledges research support from BbV i¼0
Fomento de Investigación y Cultura Superior, A. C.
(FICSAC). A preliminary version of this paper was From this we know that
presented at the 23rd International Symposium on 2 3
Forecasting held in Mérida, Yucatán (México). ZT VĨI  JJV
6X 1 7
6 ZT VðBVÞ1i ĨI  JBi JV 7
BYF;H 6 7
Appendix A. Approximation of MSE(Y F,H ) for an 6 7
¼ 6 i¼0 7:
estimated process BbV 6v 7
6 H1 7
4X H1i i 5
ZT VðBVÞ ĨI  JB JV
To derive an expression for ˜ , the derivatives of i¼0
BY F,H / BbV are needed. They can be obtained by noticing

that expression (4) can be written as E yT þh jY ¼ Let Y t = (1, yV,t . . ., y tp+1V)V be a (kp + 1)  1 vector
JBh ZT , where and Z = (Y 0, . . ., Y T1) a (kp + 1)  T
  ˆ matrix. Thus,
ZT ¼ DVT þH DVT þH1 . . . DV the asymptotic covariance matrix of b can be written as
T þ1 yV T1 . . . yV
T yV T pþ1 V
Tpþ2 yV
b̂ =& 1  e , where & u plimZZV/T and e is the
is an (nH + kp)  1 matrix of variables, error covariance matrix. Thus expression (21) becomes
2 3
0n
6 I n 0n
6
7
7 BY F;H  1  BY F;H V
6 7 R¼E
R̃ &  Re
6 O O 7 BbV Bb
6 I n 0n 7 
B¼6 6
7 R m;n for m; n ¼ 1; . . . ; H;
¼ R̃
6 K P1 . . . Pp1 Pp 7
7
6 Ik . . . 0k 0k 7
6 7
4 O v v 5 where
Ik 0k
Rm; n ¼
R̃
is an (nH + kp)  (nH + kp) matrix of coefficients and X
m1 X
n1 
tr ðBVÞm1i ĨIC1 ĨIVBn1j
J ¼ ð0kn 0kn . . . 0kn Ik 0k . . . 0k Þ i¼0 j¼0

EðZT ZT VÞCi Re Cj V
is a k  (nH + kp) matrix.
Since for m;n¼1; . . .; H; are k  k matrices and Ci ¼ JBi JV.
  " #
BE yT þh jY X
h1
h1i i BvecðBÞ
¼ ðZT V JÞ ðBVÞ B Appendix B. Restricted forecasts for an estimated
BbV BbV
"
i¼0
# process
X
h1
BvecðBÞ
¼ ZT VðBVÞh1i  JBi : In this appendix we obtain the restricted forecast
i¼0
BbV
expression for the VAR model with estimated
 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770 769

coefficients Ĥ. From the linear projection theory we Now, since

know that
  X 1 CRŶY F;H
ACRŶY F;H ¼ RŶY F;H CVX̂
Ŷ RF;H ¼ Ŷ
Y Y F;H þ Ê
E Y F  Ŷ
Y F;H jR  CŶ
Y F;H ;

is a symmetric semidefinite positive matrix and ŶF,H

where Ŷ RF,H u Ê(Y F |Y, R) is the restricted forecast for
is the MSE(Ŷ RF,H ), we know that Ŷ RF,H is at least as
an estimated process and Y F  Ê(Y F |Y) is orthogonal
precise as Ŷ F,H .
to Y conditional on R  Ê(R|Y). Now
   
E Y F  Ŷ
Ê Y F;H jR  CŶ
Y F;H ¼ Â
A R  CŶ
Y F;H References

Box, G. E. P., & Tiao, G. C. (1976). Comparison of forecasts and

for some values of  which is an H  M matrix. By actuality. Applied Statistics, 25, 195 – 200.
substituting Eqs. (18) and (22) into the following Clements, M. P., & Hendry, D. F. (2004). A companion to economic
orthogonality condition we get forecasting. Blackwell Publishing.
Doan, T., Litterman, R., & Sims, C. (1984). Forecasting and
    conditional projection using realistic prior distributions. Econo-
Ef Y F  Ŷ
Ê E Y F  Ŷ
Y F;H  Ê Y F;H jR  CŶ
Y F;H
  metric Reviews, 3, 1 – 100.
 R  CŶ Y F;H g ¼ 0 Dufour, J. M. (1985). Unbiasedness of predictions from estimated
vector autoregressions. Econometric Theory, 1, 387 – 402.
Engle, R. F., & Granger, C. W. J. (1987). Co-integration and error
so that correction representation, estimation and testing. Econometrica,
    55, 251 – 276.
RY F;H þ T 1 R̃ AC RY F;H þ T 1 R̃
R CV  Â R CV  Â
ARu Greene, M. N., Howrey, E. P., & Hymans, S. W. (1986). The use of
outside information in econometric forecasting. In E. Huh, &
¼ 0; D. A. Belsley (Eds.), Model reliability (pp. 90 – 116). Cam-
bridge, MA7 MIT Press.
and
Guerrero, V. M. (1989). Optimal conditional ARIMA forecasts.
Journal of Forecasting, 8, 215 – 229.
 X 1
A ¼ RŶY F;H CVX̂ X ¼ CRŶY F;H CV þ Ru :
where X̂
Guerrero, V. M., & Peña, D. (2003). Combining multiple time series
predictors: A useful inferential procedure. Journal of Statistical
Therefore, the optimal forecast of Y F conditional
Planning and Inference, 116, 249 – 276.
on Y and R for an estimated VAR process is Guerrero, V. M., Pena, B., Senra, E., & Alegrı́a, A. (2005).
  Restricted forecasting with a VEC model: Validating the
Ŷ RF;H ¼ Ŷ
Y Y F;H þ Â
A R  CŶ
Y F;H : feasibility of economic targets. Available at Eva Senra’s website
http://www2.uah.es/.
Its forecast error vector is Johansen, S. (1988). Likelihood-based inference in cointegrated
vector autorregressive models. Oxford University Press.
  Lütkepohl, H. (1991). Introduction to multiple time series analysis.
Y RF;H ¼ Y F  Ŷ
Y F  Ŷ Y F;H  Â
A R  CŶ
Y F;H Berlin7 Springer-Verlag.
 n hpffiffiffiffi i pffiffiffiffio Magnus, J. R., & Neudecker, H. (2002). Matrix differential
¼ I  ÂAC CeF þ T Y F;H  Ŷ
Y F;H = T calculus. Wiley.
Nieto, F. H., & Guerrero, V. M. (1995). Kalman filter for singular
and conditional state-space models when the system state and
 Â
Au;
the observational error are correlated. Statistics and Probability
Letters, 22, 303 – 310.
and hence Pandher, G. S. (2002). Forecasting multivariate time series with
  h   i linear restrictions using constrained structural state-space
MSE Ŷ Y RF;H uE Y F  Ŷ
Y RF;H Y F  Ŷ
Y RF;H VjY ; R models. Journal of Forecasting, 21, 281 – 300.
Pankratz, A. (1989). Time series forecasts and extra-model
    information. Journal of Forecasting, 8, 75 – 83.
AC RY F;H þ T 1 R̃
¼ I  Â R I  CVÂ
AV þ Â
ARu Â
AV Park, J. P., & Phillips, P. C. B. (1988). Statistical inference in
  regression with integrated process: Part 1. Econometric Theory,
¼ I  Â
AC RŶY F;H : 4, 468 – 497.
770 N. Gómez, V.M. Guerrero / International Journal of Forecasting 22 (2006) 751–770

Park, J. P., & Phillips, P. C. B. (1989). Statistical inference in
regression with integrated process: Part 2. Econometric Theory,
5, 95 – 131.
SHCP (2002). Criterios Generales de Polı́tica Económica 2003.
México7 Secretarı́a de Hacienda y Crédito Público.
Trabelsi, A., & Hillmer, S. (1989). A benchmarking approach to
forecast combination. Journal of Business and Economic
Statistics, 7, 353 – 362.
van der Knoop, H. S. (1987). Conditional forecasting with a
multivariate time series model. Economics Letters, 22, 233 – 236.