Christoffersen, P. and Diebold, F.X.

(1998),
"Cointegration and Long-Horizon Forecasting,"
Journal of Business and Economic Statistics, 16, 450-458.

Cointegration and Long-Horizon Forecasting

Peter F. Christoffersen
Research Department, International Monetary Fund, Washington, DC 20431
([email protected])

Francis X. Diebold
Department of Economics, University of Pennsylvania, Philadelphia, PA 19104
and NBER ([email protected])

Abstract: We consider the forecasting of cointegrated variables, and we show that

at long horizons nothing is lost by ignoring cointegration when forecasts are
evaluated using standard multivariate forecast accuracy measures. In fact, simple
univariate Box-Jenkins forecasts are just as accurate. Our results highlight a
potentially important deficiency of standard forecast accuracy measures—they fail
to value the maintenance of cointegrating relationships among variables— and we
suggest alternatives that explicitly do so.

KEY WORDS: Prediction, Loss Function, Integration, Unit Root

 -2-

1. INTRODUCTION

Cointegration implies restrictions on the low-frequency dynamic behavior of

multivariate time series. Thus, imposition of cointegrating restrictions has immediate
implications for the behavior of long-horizon forecasts, and it is widely believed that
imposition of cointegrating restrictions, when they are in fact true, will produce superior
long-horizon forecasts. Stock (1995, p. 1), for example, provides a nice distillation of the
consensus belief when he asserts that “If the variables are cointegrated, their values are
linked over the long run, and imposing this information can produce substantial
improvements in forecasts over long horizons.” The consensus belief stems from the
theoretical result that long-horizon forecasts from cointegrated systems satisfy the
cointegrating relationships exactly, and the related result that only the cointegrating
combinations of the variables can be forecast with finite long-horizon error variance.
Moreover, it appears to be supported by a number of independent Monte Carlo analyses
(e.g., Engle and Yoo 1987; Reinsel and Ahn 1992; Clements and Hendry 1993, Lin and
Tsay 1996).

This paper grew out of an attempt to reconcile the popular intuition sketched
above, which seems sensible, with a competing conjecture, which also seems sensible.
Forecast enhancement from exploiting cointegration comes from using information in the
current deviations from the cointegrating relationships. That is, knowing whether and by
how much the cointegrating relations are violated today is valuable in assessing where the
variables will go tomorrow, because deviations from cointegrating relations tend to be
eliminated. However, although the current value of the error-correction term clearly
provides information about the likely near-horizon evolution of the system, it seems
unlikely that it provides information about thelong-horizon evolution of the system,
because the long-horizon forecast of the error-correction term is always zero. (The
error-correction term, by construction, is covariance stationary with a zero mean.) From
this perspective, it seems unlikely that cointegration could be exploited to improve long-
horizon forecasts.

Motivated by this apparent paradox, we provide a precise characterization of the

implications of cointegration for long-horizon forecasting. Our work is closely related to
important earlier contributions of Clements and Hendry (1993, 1994, 1995) and Banerjee,
Dolado, Galbraith and Hendry (1993, pp.278-285), who compare forecasts from a true
VAR to forecasts from a misspecified VAR in differences, whereas we compare the true
forecasts to exact forecasts from correctly-specified but univariate representations. We
focus explicitly and exclusively on forecasting, and we obtain a number of new theoretical
results which sharpen the interpretation of existing Monte Carlo results. Moreover, our
motivation is often very different. Rather than focusing, for example, on loss functions
invariant to certain linear transformations of the data, we take the opposite view that loss
functions—like preferences—are sovereign, and explore in detail how the effects of
imposing cointegration on long-horizon forecasts vary fundamentally with the loss
function adopted. In short, our results and theirs are highly complementary.
 -3-

We proceed as follows. In Section 2 we show that, contrary to popular belief,

nothing is lost by ignoring cointegration when long-horizon forecasts are evaluated using
standard accuracy measures; in fact, even univariate Box-Jenkins forecasts are equally
accurate. In Section 3 we illustrate our results with a simple bivariate cointegrated system.
In Section 4, we address a potentially important deficiency of standard forecast accuracy
measures highlighted by our analysis—they fail to value the maintenance of cointegrating
relationships among variables—and we suggest alternative accuracy measures that
explicitly do so. In Section 5, we consider forecasting from models with estimated
parameters, and we use our results to clarify the interpretation of a number of well-known
Monte Carlo studies. We conclude in Section 6.

2. MULTIVARIATE AND UNIVARIATE FORECASTS

OF COINTEGRATED VARIABLES

In this section we establish notation, recall standard results on multivariate

forecasts of cointegrated variables, add new results on univariate forecasts of cointegrated
variables, and compare the two. First, let us establish some notation.

Assume that the Nx1 vector process of interest is generated by

(1 L)xt µ C(L) t,

where µ is a constant drift term, C(L) is an NxN matrix lag operator polynomial of
possibly infinite order, and t is a vector of i.i.d. innovations. Then, under regularity
conditions, the existence of r linearly independent cointegrating vectors is equivalent to
rank(C(1)) = N-r, and the cointegrating vectors are given by the rows of the rxN matrix ,
where C(1) = µ = 0. That is, z= t xt is an r-dimensional stationary zero-mean time
series. We will assume that the system is in fact cointegrated, with 0<rank(C(1))<N. For
future reference, note that following Stock and Watson (1988) we can use the

decomposition C(L) C(1) (1 L)C (L), where Cj Ci, to write the system in
i j 1
“common-trends” form,

xt µt C(1) t C (L) t,
t
where t i.
i 1

We will compare the accuracy of two forecasts of a multivariate cointegrated

system that are polar extremes in terms of cointegrating restrictions imposed—first,
forecasts from the multivariate model, and second, forecasts from the implied univariate
models. Both forecasting models are correctly specified from aunivariate perspective, but
 -4-

one imposes the cointegrating restrictions and allows for correlated error terms across
equations, and the other does not.

We will make heavy use of a ubiquitous forecast accuracy measure, mean squared
error, the multivariate version of which is
MSE E(et hKet h),

where K is an NxN positive definite symmetric matrix andet h is the vector of h-step-ahead
forecast errors. MSE of course depends on the weighting matrix K. It is standard to set
K=I, in which case
MSE E(et het h) trace( h),

where h var(et h) . We call this the “trace MSE” accuracy measure. To compare the
accuracy of two forecasts, say 1 to 2, it is standard to examine the ratio
1 2
trace( h)/ trace( h) which we call the “trace MSE ratio.”

2.1 Forecasts From the Multivariate Cointegrated System

Now we review standard results (required by our subsequent analysis) on

multivariate forecasting in cointegrated systems. For expanded treatments, see Engle and
Yoo (1987) and Lin and Tsay (1996).

From the moving average representation, we can unravel the process recursively
from time t+h to time 1 and write
t t h i h h i
xt h (t h)µ Cj i Cj t i,
i 1 j 0 i 1 j 0

from which the h-step-ahead forecasts are easily calculated as

t t h i
x̂t h (t h)µ Cj i.
i 1 j 0

From the fact that

t h i
lim Cj C(1),
h j 0

we get that
lim x̂t h 0,
h

so that the cointegrating relationship is satisfied exactly by the long-horizon system

forecasts. This is the sense in which long-horizon forecasts from cointegrated systems
preserve the long-run multivariate relationships exactly.

We define the h-step-ahead forecast error from the multivariate system as

 -5-

êt h xt h x̂t h.

The forecast errors from the multivariate system satisfy

h h i
êt h
Cj ,
t i
i 1 j 0

so the variance of the h-step-ahead forecast error is

h h i h i
var[êt h] Cj Cj ,
i 1 j 0 j 0

where is the variance of t.

From the definition of êt h we can also see that the system forecast errors satisfy
h
êt h
êt h 1
Ch i t i
C(L) t h
,
i 1

where the last equality holds if we take j=0 for all j<t. That is, when we view the system
forecast error process as a function of the forecast horizon, h, it has the same stochastic
structure as the original process, x,t and therefore is integrated and cointegrated.
Consequently, the variance of the h-step ahead forecast errors from the cointegrated
system is of order h, that is increasing at the rate h,

var[êt h] O(h).

In contrast, the cointegrating combinations of the system forecast errors, just as the error-
correction process z,t will have finite variance for large h,

lim var[ êt h] Q < ,

h

where the matrix Q is a constant function of the stationary component of the forecast
error. Although individual series can only be forecast with increasingly wide confidence
intervals, the cointegrating combination has a confidence interval of finite width, even as
the forecast horizon goes to infinity.

2.2 Forecasts from the Implied Univariate Representations

Now consider ignoring the multivariate features of the system, forecasting instead
using the implied univariate representations. We can use Wold’s decomposition theorem
and write for any series (the n-th, say),

(1 L)xn,t µn n,jun,t j,
j 0
 -6-

where n,0 = 1 and un,t is white noise. It follows from this expression that the univariate
time-t forecast for period t+h is,
h h 1
x̃n,t h hµn xn,t n,i un,t n,i un,t 1 ...
i 1 i 2

Using obvious notation we can write

x̃n,t h hµn xn,t n(L)un,t,

and stacking the N series we have

x̃t h hµ xt (L)ut,

where (L) is a diagonal matrix polynomial with the individual n(L)’s on the diagonal.

Now let us consider the errors from the univariate forecasts. We will rely on the
following convenient orthogonal decomposition
ẽt h xt h x̃t h (xt h x̂t h) (x̂t h x̃t h) êt h (x̂t h x̃t h).

Recall that the system forecast is

t t h i
x̂t h µ(t h) Cj i µ(t h) C(1) t,
i 1 j 0

where the approximation holds as h gets large. Using univariate forecasts, the
decomposition for ẽt h , and the approximate long-horizon system forecast, we get

ẽt h êt h (µ(t h) C(1) t) (xt µh (L)ut).

Now insert the common trends representation for xt to get

ẽt h êt h µ(t h) C(1) t (µt C(1) t C (L) t µh (L)ut),

and finally cancel terms to get

ẽt h êt h (C (L) t (L)ut).

Notice that the t’s are serially uncorrelated and the u’st only depend on current
and past t’s; thus, êt h is orthogonal to the terms in the parenthesis. Notice also that the
term inside the parenthesis term is just a sum of stationary series and is therefore
stationary; furthermore, its variance is constant as the forecast horizon h changes. We can
therefore write the long-horizon variance of the univariate forecasts as

var(ẽt h) Var(êt h) O(1) O(h) O(1) O(h),

which is of the same order of magnitude as the variance of thesystem forecast errors.
Furthermore, since the dominating terms in the numerator and denominator are identical,
the trace MSE ratio goes to one as formalized in the following proposition:
 -7-

Proposition 1
trace(var(ẽt h))
lim 1.
h trace(var(êt h))
When comparing accuracy using the trace MSE ratio, the univariate forecasts
perform as well as the cointegrated system forecasts as the horizon gets large. This is the
opposite of the folk wisdom—it turns out that imposition of cointegrating restrictions
helps at short, but not long, horizons. Quite simply, when accuracy is evaluated with the
trace MSE ratio, there is no long-horizon benefit from imposing cointegration; all that
matters is getting the level of integration right.

Proposition 1 provides the theoretical foundation for the results of Hoffman and
Rasche (1996), who find in an extensive empirical application that imposing cointegration
does little to enhance long-horizon forecast accuracy, and Brandner and Kunst (1990),
who suggest that when in doubt about how many unit roots to impose in a multivariate
long-horizon forecasting model, it is less harmful to impose too many than to impose too
few. A similar result can be obtained by taking the ratio of Clements and Hendry’s (1995)
formulas for the MSE at horizon h from the system forecasts and the MSE of forecasts
that they construct that correspond approximately to those from a misspecified VAR in
differences.

Now let us consider the variance of cointegrating combinations of univariate

forecast errors. Above we recounted the Engle-Yoo (1987) result that the cointegrating
combinations of the system forecast errors have finite variance as the forecast horizon gets
large. Now we want to look at the same cointegrating combinations of theunivariate
forecast errors. From our earlier derivations it follows that

ẽt h êt h ( C (L) t (L)ut).

Again we can rely on the orthogonality of êt h to the terms in the parenthesis. The first
term, êt h, has finite variance, as discussed above. So too do the terms in the parenthesis,
because they are linear combinations of stationary processes. Thus we have

Proposition 2
var( ẽt h) Q var(C (L) t (L)ut) O(1).

The cointegrating combinations of the long-horizon errors from the univariate forecasts,
which completely ignore cointegration, also have finite variance. Thus, it is in fact not
imposition of cointegration on the forecasting system that yields the finite variance of the
cointegrating combination of the errors; rather it is the cointegration property inherent in
the system itself, which is partly inherited by the correctly specified univariate forecasts.
 -8-

3. A SIMPLE EXAMPLE

In this section, we illustrate the results from Section 2 in a simple multivariate

system. Consider the bivariate cointegrated system,
xt µ xt 1 t

yt xt vt,

where the disturbances are orthogonal at all leads and lags. The moving average
representation is
xt µ 1 0 t µ t
(1 L) C(L) ,
yt µ 1 L vt µ vt

and the error-correction representation is

xt µ 0 xt 1 t
(1 L) 1 .
yt µ 1 yt 1 t vt

The system’s simplicity allows us to compute exact formulae that correspond to the
qualitative results derived in the previous section.

3.1 Univariate Representations

Let us first derive the implied univariate representations for x and y. The univariate
representation for x is of course a random walk with drift, exactly as given in the first
equation of the system,
xt µ xt 1 t.

Derivation of the univariate representation for y is a bit more involved. From the moving-
average representation of the system, rewrite the process for yt as a univariate two-shock
process,
yt µ yt 1 (1 L)vt t

µ yt 1
zt,

where zt (1 L)vt t. The autocovariance structure for zt is

2 2 2
z(0) 2 v
2
z
(1) z
( 1) v

z( ) 0, 2.
 -9-

The only non-zero positive autocorrelation is therefore the first,

2
v 1
z(1) 2
,
2 2 2 2
q
2 v

2 2
where q / v is the signal to noise ratio. This is exactly the autocorrelation structure
of an MA(1) process, so we write zt ut 1 ut. To find the value for , we match
autocorrelations at lag 1, yielding
1
.
2
1 2 2q

This gives a second-order polynomial in , with invertible solution

4 2
(1/2)[ q 4 2q 2 2
q].
2 2
Although suppressed in the notation, will be a function of q / v and throughout.
Finally, we find the variance of the univariate innovation by matching the variances,
yielding

2 2 2 2 2 2 2
(1 ) u 2 v v(2 q),

or
2 2
2 v(2 q)
u .
2
(1 )
3.2 Forecasts From the Multivariate Cointegrated System

First consider forecasting from the multivariate cointegrated system. Write the time
t+h values in terms of time t values and future innovations as
h
xt h µh xt t i
i 1
h
yt h (µh xt) t i vt h .
i 1

The h-step-ahead forecasts are

x̂t h µh xt

ŷt h
µh xt,

and the h-step-ahead forecast errors are

 - 10 -

h
êx,t h t i
i 1
h
êy,t h t i vt h .
i 1

Note that the forecast errors follow the same stochastic process as the original system
(aside from the drift term),
t h 1 0 t h t h
(1 L)êt h C(L) .
t h (1 L)vt h 1 L vt h vt h

Finally, the corresponding forecast error variances are

2
var(êx,t h) h

2 2 2
var(êy,t h) h v.

Both forecast error variances are O(h). As for the variance of the cointegrating
combination, we have
h h
2
var[êy,t h êx,t h] var t i vt h t i v,
i 1 i 1

for all h, because there are no short run dynamics. Similarly, the forecasts satisfy the
cointegrating relationship at all horizons, not just in the limit. That is,
ŷt h
x̂t h
0, h 1, 2, ...

3.3 Forecasts From the Implied Univariate Representations

Now consider forecasting from the implied univariate models. Immediately, the
univariate forecast for x is the same as the system forecast,

x̃t h µh xt.

Thus,
h
ẽx,t h êx,t h t i,
i 1

so that
2
var(ẽx,t h) var(êx,t h) h O(h).

To form the univariate forecast for y, write

 - 11 -

h h
yt h µh yt zt i µh yt ut ut 1 zt i.
i 1 i 2

The time t forecast for period t+h is

ỹt h µh yt ut,
and the corresponding forecast error is
h h 1
ẽy,t h ut 1 zt i (1 ) ut i ut h,
i 2 i 1

yielding the forecast error variance

2
var(ẽy,t h) [(1 )2(h 1) 1] u.

Notice in particular that the univariate forecast error variance is O(h), as is the system
forecast error variance.

Now let us compute the variance of the cointegrating combination of univariate

forecast errors. We have

2
var[ẽy,t h ẽx,t h] var(ẽy,t h) var(ẽx,t h) 2 cov ẽy,t h, ẽx,t h .

2
The second variance term is simply h . To evaluate the first variance term we write
2 2
var(ẽy,t h) [(1 )2(h 1) 1] u [(1 )2h (2 )] u.
2
Substituting for u, and using the fact that
1 (1 )2 2
q
,
2 2 2 2
1 2 q 1 2 q
we get
(1 )2 2 2 2 2
var(ẽy,t h) h v(2 q) [ 2qh 2 ] v.
2 2
1 1
To evaluate the covariance term, use the fact that

ŷt h ỹt h ( µh xt) ( µh yt ut) vt ut,

and the decomposition result from Section 2 to write

h
ẽy,t h
êy,t h
(ŷt h
ỹt h) t i
vt h
vt ut.
i 1

Now recall the formula for the forecast error of x and the fact that future values of are
uncorrelated with future and current values of v, and with current values of u, so that
 - 12 -

h h
2
cov ẽy,t h, ẽx,t h E t i vt h vt ut t i h .
i 1 i 1

Armed with these results we have that

2
var[ẽy,t h ẽx,t h] (2 ) v < h,

which of course accords with our general result derived earlier that the variance of the
cointegrating combination of univariate forecast errors is finite.

3.4 Forecast Accuracy Comparison

Finally, compare the forecast error variances from the multivariate and univariate
representations. Of course x has the same representation in both, so the comparison hinges
on y. We must compare
2 2 2 2 2
var(êy,t h) h v v[ qh 1]

to
2
var(ẽy,t h) [ 2qh 2 ] v.

Thus,
2
var(ẽy,t h) var(êy,t h) (1 ) v.

The error variance of the univariate forecast is greater than that of the system forecast, but
it grows at the same rate.

Assembling all of the results, we have immediately that

trace(var(ẽt h)) var(ẽx,t h) var(ẽy,t h) qh 2
qh 2
.
trace(var(êt h)) var(êx,t h) var(êy,t h) qh 1 2
qh

In Figure 1 we show the values of this ratio as h gets large, for q = = 1. Note in
particular the speed with which the limiting result,
trace(var(ẽt h))
lim 1,
h trace(var(êt h))
obtains.

In closing this section, we note that in spite of the fact that the trace MSE ratio
approaches 1, the ratio of the variances of the cointegrating combinations of the forecast
errors does not approach 1 in this simple model; rather,
var[ẽy,t h ẽx,t h]
(2 ) > 1, h, q.
var[êy,t h êx,t h]
 - 13 -

This observation turns out to hold quite generally, and it forms the basis for an alternative
class of accuracy measures, to which we now turn.

4. ACCURACY MEASURES AND COINTEGRATION

4.1 Accuracy Measures I: Trace MSE

We have seen that long-horizon univariate forecasts of cointegrated variables

(which ignore cointegrating restrictions) are just as accurate as their system counterparts
(which explicitly impose cointegrating restrictions), when accuracy is evaluated using the
standard trace MSE criterion. So on traditional grounds there is no reason to prefer long-
horizon forecasts from the cointegrated system.

One might argue, however, that the system forecasts are nevertheless more
appealing because “... the forecasts of levels of co-integrated variables will ‘hang
together’ in a way likely to be viewed as sensible by an economist, whereas forecasts
produced in some other way, such as by a group of individual, univariate Box-Jenkins
models, may well not do so” (Granger and Newbold 1986, p. 226). But as we have seen,
univariate Box-Jenkins forecasts do hang together if the variables are cointegrated—the
cointegrating combinations, and only the cointegrating combinations, of univariate
forecast errors have finite variance.

4.2 Accuracy Measures II: Trace MSE in Forecasting the Cointegrating

Combinations of Variables

The long-horizon system forecasts, however, do a better job of satisfying the

cointegrating restrictions than do the univariate forecasts—the long-horizon system
forecasts always satisfy the cointegrating restrictions, whereas the long-horizon univariate
forecasts do so only on average. That is what is responsible for our earlier result in our
bivariate system that, although the cointegrating combinations of both the univariate and
system forecast errors have finite variance, the variance of the cointegrating combination
of the univariate errors is larger.

Such effects are lost on standard accuracy measures like trace MSE, however,
because the loss functions that underlie them do not explicitly value maintaining the
multivariate long-run relationships of long-horizon forecasts. The solution is obvious—if
we value maintenance of the cointegrating relationship, then so too should the loss
functions underlying our forecast accuracy measures. One approach, in the spirit of
Granger (1996), is to focus on forecasting the cointegrating combinations of the variables,
and to evaluate forecasts in terms of the variability of the cointegrating combinations of
the errors, et+h.
 - 14 -

Accuracy measures based on cointegrating combinations of the forecast errors

require that the cointegrating vector be known. Fortunately, such is often the case.
Horvath and Watson (1995, pp. 984-985) [see also Watson (1994) and Zivot (1996)], for
example note that

“Economic models often imply that variables are cointegrated with simple
and known cointegrating vectors. Examples include the neoclassical growth
model, which implies that income, consumption, investment, and the capital
stock will grow in a balanced way, so that any stochastic growth in one of
the series must be matched by corresponding growth in the others. Asset
pricing models with stable risk premia imply corresponding stable
differences in spot and forward prices, long- and short-term interest rates,
and the logarithms of stock prices and dividends. Most theories of
international trade imply long-run purchasing power parity, so that long-run
movements in nominal exchange rates are matched by countries’ relative
price levels. Certain monetarist propositions are centered around the
stability of velocity, implying cointegration among the logarithms of
money, prices and income. Each of these theories has distinct implications
for the properties of economic time series under study: First, the series are
cointegrated, and second, the cointegrating vector takes on a specific value.
For example, balanced growth implies that the logarithms of income and
consumption are cointegrated and that the cointegrating vector takes on the
value of (1, -1).”

Thus, although the assumption of a known cointegrating vector certainly involves a loss of
generality, it is nevertheless legitimate in a variety of empirically- and economically-
relevant cases. This is fortunate because of problems associated with identification of
cointegrating vectors in estimated systems, as stressed in Wickens (1996). We will
maintain the assumption of a known cointegration vector throughout this paper, reserving
for subsequent work an exploration of the possibility of analysis using consistent estimates
of cointegrating vectors.

Interestingly, evaluation of accuracy in terms of the trace MSE of the cointegrating

combinations of forecast errors is a special case of the general mean squared error
measure. To see this, consider the general N-variate case with r cointegrating
relationships, and consider again the mean squared error,

E(et hKet h) = Etrace(et hKet h) = Etrace(Ket het h) = trace(K h) ,

where h is the variance of et h . Evaluating accuracy in terms of trace MSE of the

cointegrating combinations of the forecast errors amounts to evaluating

E ( et h) ( et h) trace E ( et h) ( et h) trace(K h),

 - 15 -

where K = . Thus the trace MSE of the cointegrating combinations of the forecast
errors is in fact a particular variant of MSE formulated on the raw forecast errors, E(eKe)
= trace(K h) , where the weighting matrix K = is of (deficient) rank r (< N), the
cointegrating rank of the system.

4.3 Accuracy Measures III: Trace MSE from the Triangular Representation

The problem with the traditional E(eKe) approach with K = I is that, although it
values small MSE, it fails to value the long-run forecasts’ hanging together correctly.
Conversely, a problem with the E(eKe) approach with K = is that it values only the
long-run forecasts’ hanging together correctly, whereas both pieces seem clearly relevant.
The challenge is to incorporate both pieces into an overall accuracy measure in a natural
way, and an attractive approach for doing so follows from the triangular representation of
cointegrated systems exploited by Campbell and Shiller (1987) and Phillips (1991).
Clements and Hendry (1995) provide a numerical example that illustrates the appeal of the
triangular representation for forecasting. Below we provide a theoretical result that
establishes the general validity of the triangular approach for distinguishing between naive
univariate and fully specified system forecasts.

From the fact that has rank r, it is possible to rewrite the system so that the N
left-hand-side variables are the r error-correction terms followed by the differences of N-r
integrated but not cointegrated variables. That is, we rewrite the system in terms of

x1t x2t
,
(1 L)x2t

where the variables have been rearranged and partitioned into xt (x1t, x2t) , where
( ) and the variables in x2t are integrated but not cointegrated. We then evaluate
accuracy in terms of the trace MSE of forecastsfrom the triangular system,

e1,t h e2,t h e1,t h e2,t h Ir Ir

E E et h
et h
,
(1 L)e2,t h (1 L)e2,t h 0 (1 L) 0 (1 L)
which we denote trace MSEtri. Notice that the trace MSEtri accuracy measure is also of
E(e Ke) form, with
Ir Ir
K K(L) .
0 (1 L) 0 (1 L)

Recall Proposition 1, which says that under trace MSE, long-horizon forecast
accuracy from the cointegrated system is no better than that from univariate models. We
 - 16 -

now show that under trace MSEtri, long-horizon forecast accuracy from the cointegrated
system is always better than that from univariate models.

Proposition 3
˜
trace MSEtri
lim > 1.
h ˆ
trace MSEtri

Proof: Consider a cointegrated system in triangular form, that is, a system such that = [
I - ]. We need to show that for large h,
r N r N
var[ iêt h] var[(1 L)êj,t h] < var[ iẽt h] var[(1 L)ẽj,t h]
i 1 j r 1 i 1 j r 1

and
r N
var[ iêt h] var[(1 L)êj,t h] < .
i 1 j r 1

To establish the first inequality it is sufficient to show that

r r
var[ iêt h] < var[ iẽt h].
i 1 i 1

We showed earlier that for large h,

var( ẽt h) Q var(C (L) t (L)ut) (Q S) ,

where Q var(êt h), S var(C (L) t (L)ut), and from which it follows that
r r
var[ iẽt h] var[ iêt h] trace( S ) > 0,
i 1 i 1

because S is positive definite. To establish the second inequality, recall that

h
êt h
êt h 1
Ch i t i
C(L) ,
t h
i 1

so that
h 1 h 1
var[(1 L)êt h] Cj Cj C(1) C(1) as h .
j 0 j 0

Let CN-r(1) be the last N-r rows of C(1); then altogether we have
r N
var[ iêt h] var[(1 L)êj,t h] trace( Q ) trace(CN r(1) CN r(1) ) < ,
i 1 j r 1
 - 17 -

and the proof is complete.

4.4 The Bivariate Example, Revisited

In our simple bivariate example all we have to do to put the system in the
triangular form sketched above is to switch x and y in the autoregressive representation,
yielding
1 yt 0 vt
.
0 1 L xt µ t

For the system forecasts we have

êy,t h
êx,t h
êy,t h
êx,t h
ˆ
trace MSE E
2
v
2
.
tri
(1 L)êx,t h (1 L)êx,t h

For the univariate forecasts we have

ẽy,t h
ẽx,t h
ẽy,t h
ẽx,t h
˜
trace MSE E (2 )
2 2
.
tri v
(1 L)ẽx,t h (1 L)ẽx,t h

Thus we see that the trace MSEtri ratio does not approach one as the horizon increases; in
particular, it is constant and above one for all h,
˜
trace MSE 1 (2 )q (1 )q
tri
1 > 1, h.
ˆ
trace MSE 1 q 1 q
tri

2 2
In Figure 2 we plot the trace MSEtri ratio vs. h, for =1 and v q 1 . In this case, the
ratio is simply a constant (> 1) for all h since the systems contains no short-run dynamics.

In summary, although the long-horizon performances of the system and univariate

forecasts seem identical under the conventional trace MSE ratio, they differ under the
trace MSEtri ratio. The system forecast is superior to the univariate forecast under trace
MSEtri, because the system forecast is accurate in the conventional “small MSE” sense
and it hangs together correctly, i.e. it makes full use of the information in the cointegrating
relationship. We stress that abandoning MSE and adopting MSEtri marks a change of loss
function, and thus preferences. If the forecaster’s loss function truly is trace MSE then
using trace MSEtri might not make sense. On the other hand trace MSE is often adopted
without much thought, and an underlying theme of our analysis is precisely that thought
should be given to the choice of loss function.
 - 18 -

5. UNDERSTANDING EARLIER MONTE CARLO STUDIES

Here we clarify the interpretation of earlier influential Monte Carlo work, in

particular Engle and Yoo (1987), as well as Reinsel and Ahn (1992), Clements and
Hendry (1993), and Lin and Tsay (1996), among others. We do so by performing a Monte
Carlo analysis of our own, which reconciles our theoretical results and the apparently
conflicting Monte Carlo results reported in the literature, and we show how the existing
Monte Carlo analyses have been misinterpreted. Throughout, we use our simple bivariate
system (which is very similar to the one used by Engle and Yoo), with parameters set to
2 2
=1, µ=0 and v 1 . We use a sample size of 100 and perform 4000 Monte Carlo
replications. In keeping with our earlier discussion, we assume a known cointegrating
vector, but we estimate all other parameters. This simple design allows us to make our
point forcefully and with a minimum of clutter, and the results are robust to changes in
parameter values and sample size.

Let us first consider an analog of our theoretical results, except that we now
estimate parameters instead of assuming them known. In Figure 3 we plot the trace MSE
ratio and the trace MSEtri ratio against the forecast horizon, h. Using estimated parameters
changes none of the theoretical results reached earlier under the assumption of known
parameters. Use of the trace MSE ratio obscures the long-horizon benefits of imposing
cointegration, whereas use of trace MSEtri reveals those benefits clearly.

How then can we reconcile our results with those of Engle and Yoo (1987) and the
many subsequent authors who conclude that imposing cointegration produces superior
long-horizon forecasts? The answer is two-part: Engle and Yoo make a different and
harder-to-interpret comparison than we do, and they misinterpret the outcome of their
Monte Carlo experiments.

First consider the forecast comparison. We have thus far compared forecasts from
univariate models (which impose integration) to forecasts from the cointegrated system
(which impose both integration and cointegration). Thus a comparison of the forecasting
results isolates the effects of imposing cointegration. Engle and Yoo, in contrast, compare
forecasts from a VAR in levels (which imposeneither integration nor cointegration) to
forecasts from the cointegrated system (which imposeboth integration and cointegration).
Thus differences in forecasting performance in the Engle-Yoo setup cannot necessarily be
attributed to the imposition of cointegration—instead, they may simply be due to
imposition of integration, irrespective of whether cointegration is imposed.

Now consider the interpretation of the results. The VAR in levels is of course
integrated, but estimating the system in levels entails estimating the unit root. Although
many estimators are consistent, an exact finite-sample unit root is a zero-probability event.
Unfortunately, even a slight and inevitable deviation of the estimated root from unity
pollutes forecasts from the estimated model, and the pollution increases with h. This in
 - 19 -

turn causes the MSE ratio to increase in h when comparing a levels VAR forecast to a
system forecast or any other forecast that explicitly imposes unit roots. The problem is
exacerbated by bias of the Dickey-Fuller-Hurwicz type; see Stine and Shaman (1989),
Pope (1990), Abadir (1993) and Abadir, Hadri and Tzavalis (1996) for detailed
treatments.

It is no surprise that forecasts from the VAR estimated in levels perform poorly,
with performance worsening with horizon, as shown Figure 4. It is tempting to attribute
the poor performance of the VAR in levels to its failure to impose cointegration, as do
Engle and Yoo. The fact is, however, that the VAR in levels performs poorly because it
fails to impose integration, not because it fails to impose cointegration—estimation of the
cointegrated system simply imposes the correct level of integrationa priori. To see this,
consider Figure 5, in which we compare the forecasts from an estimated VAR in
differences to the forecasts from the estimated cointegrated system. At long horizons, the
forecasts from the VAR in differences, which impose integration but completely ignore
cointegration, perform just as well. In contrast, if we instead evaluate forecast accuracy
with the trace MSEtri ratio that we have advocated, the forecasts from the VAR in
differences compare poorly at all horizons to those from the cointegrated system, as
shown in Figure 6.

In the simple bivariate system, we are restricted to studying models with exactly
one unit root and one cointegration relationship. It is also of interest to examine richer
systems; conveniently, the literature already contains relevant (but unnoticed) evidence,
which is entirely consistent with our theoretical results. Reinsel and Ahn (1992) and Lin
and Tsay (1996), in particular, provide Monte Carlo evidence on the comparative
forecasting performance of competing estimated models. Both study a four-variable
VAR(2), with two unit roots and two cointegrating relationships. Their results clearly
suggest that under the trace MSE accuracy measure, one need only worry about imposing
enough unit roots on the system. Imposing three (one too many) unit roots is harmless at
any horizon, and imposing four unit roots (two too many, so that the VAR is in
differences) is harmless at long horizons. As long as one imposes enough unit roots, at
least two in this case, the trace MSE ratio will invariably go to one as the horizon
increases.

6. SUMMARY AND CONCLUDING REMARKS

First, we have shown that imposing cointegration does not improve long-horizon
forecast accuracy when forecasts of cointegrated variables are evaluated using the
standard trace MSE ratio. Ironically enough, although cointegration implies restrictions on
low-frequency dynamics, imposing cointegration is helpful for short- but not long-horizon
forecasting, in contrast to the impression created in the literature. Imposition of
cointegration on an estimated system, when the system is in fact cointegrated, helps the
accuracy of long-horizon forecasts relative to those from systems estimated in levels with
no restrictions, but that is because of the imposition of integration, not cointegration.
Univariate forecasts in differences do just as well! We hasten to add, of course, that the
result is conditional on the assumption that the univariate representations of all variables
 - 20 -

do in fact contain unit roots. Differencing a stationary variable with roots close to unity
has potentially dire consequences for long-horizon forecasting, as argued forcefully by
Lin and Tsay (1996).

Second, we have shown that the variance of the cointegrating combination of the
long-horizon forecast errors is finite regardless of whether cointegration is imposed. The
variance of the error in forecasting the cointegrating combinationis smaller, however, for
the cointegrated system forecast errors. This suggests that accuracy measures that value
the preservation of long-run relationships should be defined, in part, on the cointegrating
combinations of the forecast errors. We explored one such accuracy measure based on the
triangular representation of the cointegrated system.

Third, we showed that our theoretical results are entirely consistent with several
well-known Monte Carlo analyses, whose interpretation we clarified. The existing Monte
Carlo results are correct, but their widespread interpretation is not. Imposition of
integration, not cointegration, is responsible for the repeated finding that the long-horizon
forecasting performance of cointegrated systems is better than that of VARs in levels.

We hasten to add that the message of this paper isnot that cointegration is of no
value in forecasting. First, even under the conventional trace MSE accuracy measure,
imposing cointegration does improve forecasts. Our message is simply that under the
conventional accuracy measure it does so at short and moderate, not long, horizons, in
contrast to the folk wisdom. Second, in our view, imposing cointegration certainlymay be
of value in long-horizon forecasting—the problem is simply that standard forecast
accuracy measures do not reveal it.

The upshot is that in forecast evaluation we need to think hard about what
characteristics make a good forecast good, and how best to measure those characteristics.
In that respect this paper is in the tradition of our earlier work, such as Diebold and
Mariano (1995), Diebold and Lopez (1996), and Christoffersen and Diebold (1996, 1998),
in which we argue the virtues of tailoring accuracy measures in applied forecasting to the
specifics of the problem at hand. Seemingly omnibus measures such as trace MSE,
although certainly useful in many situations, are inadequate in others.

In closing, we emphasize that the particular alternative to trace MSE that we

examine in this paper, trace MSEtri, is but one among many possibilities, and we look
forward to exploring variations in future research. The key insight, it seems to us, is that if
we value preservation of cointegrating relationships in long-horizon forecasts, then so too
should our accuracy measures, and trace MSEtri is a natural loss function that does so.

Interestingly, it is possible to process the trace MSE differently to obtain an

accuracy measure that ranks the system forecasts as superior to the univariate forecasts,
even as the forecast horizon goes to infinity. One obvious candidate is the trace MSE
difference, as opposed to the trace MSE ratio. It follows from the results of Section 2 that
the trace MSE difference is positive and does not approach zero as the forecast horizon
grows. As stressed above, however, it seems more natural to work with alternatives to
 - 21 -

trace MSE that explicitly value preservation of cointegrating relationships, rather then
simply processing the trace MSE differently. As the forecast horizon grows, the trace
MSE difference becomes negligible relative to either the system or the univariate trace
MSE, so that the trace MSE difference would appear to place too little value on preserving
cointegrating relationships.

ACKNOWLEDGMENTS

We thank the Co-Editor (Ruey Tsay), an Associate Editor, and two referees for
detailed and constructive comments. Helpful discussion was also provided by Dave
DeJong, Rob Engle, Clive Granger, Bruce Hansen, Dennis Hoffman, Laura Kodres, Jim
Stock, Charlie Thomas, Ken Wallis, Chuck Whiteman, Mike Wickens, Tao Zha, and
participants at the July 1996 NBER/NSF conference on Forecasting and Empirical
Methods in Macroeconomics. All remaining inadequacies are ours alone. We thank the
International Monetary Fund, the National Science Foundation, the Sloan Foundation and
the University of Pennsylvania Research Foundation for support. The views in this article
do not necessarily represent those of the International Monetary Fund.
Figure 1. Trace MSE Ratio of Univariate vs. System Forecasts Plotted Against
Forecast Horizon, Bivariate System with Cointegration Parameter =1 and Signal to
Noise Ratio q=1
Figure 2. Trace MSE tri Ratio of Univariate vs. System Forecasts Plotted Against
Forecast Horizon, Bivariate System with Cointegration Parameter =1 and Signal to
Noise Ratio q=1
Figure 3. Trace MSE Ratio and Trace MSE tri Ratio of Univariate vs. System
Forecasts Plotted Against Forecast Horizon, Bivariate System with Estimated
Parameters
Figure 4. Trace MSE Ratio of Levels VAR vs. Cointegrated System Forecasts Plotted
Against the Forecast Horizon, Bivariate System with Estimated Parameters
Figure 5. Trace MSE Ratio of Differenced VAR vs. Cointegrated System Forecasts
Plotted Against Forecast Horizon, Bivariate System with Estimated Parameters
Figure 6. Trace MSE tri Ratio of Differenced VAR vs. Cointegrated System Forecasts
Plotted Against Forecast Horizon, Bivariate System with Estimated Parameters
 References

Abadir, K.M. (1993), “OLS Bias in a Nonstationary Autoregression,”Econometric

Theory, 9, 81-93.

Abadir, K.M., Hadri, K. and Tzavalis, E. (1996), “The Influence of VAR Dimensions on
Estimator Biases,” Economics Discussion Paper 96/14, University of York.

Banerjee, A., Dolado, J., Galbraith, J.W. and Hendry, D.F. (1993),Co-integration, Error-
correction, and the Econometric Analysis of Non-stationary Data. Oxford: Oxford
University Press.

Brandner, P. and Kunst, R.M. (1990), “Forecasting Vector Autoregressions - The

Influence of Cointegration: A Monte Carlo Study,” Research Memorandum No.
265, Institute for Advanced Studies, Vienna.

Campbell, J.Y. and Shiller, R.J. (1987), "Cointegration and Tests of Present Value
Models," Journal of Political Economy, 95, 1062-1088.

Christoffersen, P.F. and Diebold, F.X. (1996), "Further Results on Forecasting and Model
Selection Under Asymmetric Loss," Journal of Applied Econometrics, 11, 561-
571.

)))))) (1998), "Optimal Prediction Under Asymmetric Loss,"Econometric Theory, in

press.

Clements, M.P. and Hendry, D.F. (1993), “On the Limitations of Comparing Mean Square
Forecast Errors,” Journal of Forecasting, 12, 617-637.

(1994), “Towards a Theory of Economic Forecasting,” in C.P. Hargreaves (ed.),

)))))))
Nonstationary Time Series and Cointegration. Oxford: Oxford University Press.

(1995), “Forecasting in Cointegrated Systems,”Journal of Applied Econometrics,

)))))))
10, 127-146.

Diebold, F.X. and Lopez, J. (1996), "Forecast Evaluation and Combination," in G.S.
Maddala and C.R. Rao (eds.), Handbook of Statistics. Amsterdam: North-Holland,
241-268.

Diebold, F.X. and Mariano, R.S. (1995), "Comparing Predictive Accuracy,"Journal of

Business and Economic Statistics, 13, 253-265.

Engle, R.F. and Yoo, B.S. (1987), “Forecasting and Testing in Cointegrated Systems,”
Journal of Econometrics, 35, 143-159.

Granger, C.W.J. (1996), “Can We Improve the Perceived Quality of Economic

Forecasts?,” Journal of Applied Econometrics, 11, 455-473.
Granger, C.W.J., and Newbold, P. (1986), Forecasting Economic Time Series, Second
Edition. New York: Academic Press.

Hoffman, D.L. and Rasche, R.H. (1996), "Assessing Forecast Performance in a

Cointegrated System," Journal of Applied Econometrics, 11, 495-517.

Horvath, M.T.K. and Watson, M.W. (1995), “Testing for Cointegration When Some of
the Cointegrating Vectors are Known,” Econometric Theory, 11, 984-1014.

Lin, J.-L. and Tsay, R.S. (1996), “Cointegration Constraints and Forecasting: An
Empirical Examination,” Journal of Applied Econometrics, 11, 519-538.

Phillips, P.C.B. (1991), “Optimal Inference in Cointegrated Systems,”Econometrica, 59,

283-306.

Pope, A.L. (1990), “Biases of Estimators in Multivariate Non-Gaussian Autoregressions,”

Journal of Time Series Analysis, 11, 249-258.

Reinsel, G.C. and Ahn, S.K. (1992), “Vector Autoregressive Models with Unit Roots and
Reduced Rank Structure: Estimation, Likelihood Ratio Test, and Forecasting,”
Journal of Time Series Analysis, 13, 353-375.

Stine, R.A. and Shaman, P. (1989), “A Fixed Point Characterization for Bias of
Autoregressive Estimators,” Annals of Statistics, 17, 1275-1284.

Stock, J.H. (1995), “Point Forecasts and Prediction Intervals for Long Horizon Forecasts,”
Manuscript, J.F.K. School of Government, Harvard University.

Stock, J.H. and Watson, M.W. (1988), “Testing for Common Trends,”Journal of the
American Statistical Association, 83, 1097-1107.

Watson, M.W. (1994), “Vector Autoregressions and Cointegration,” in R.F. Engle and D.
McFadden (eds.), Handbook of Econometrics, Vol. IV, Chapter 47. Amsterdam:
North-Holland.

Wickens, M.R. (1996), “Interpreting Cointegrating Vectors and Common Stochastic

Trends,” Journal of Econometrics, 74, 255-271.

Zivot, E. (1996), “The Power of Single Equation Tests for Cointegration when the
Cointegrating Vector is Prespecified,” Manuscript, Department of Economics,
University of Washington.