Econometric Research and Special Studies Department

(EURO) Exchange Rate Predictability and Monetary Fundamentals in a Small

Multi-Country Panel

J.J.J. Groen

Research Memorandum WO&E no. 664

August 2001

De Nederlandsche Bank
(EURO) EXCHANGE RATE PREDICTABILITY AND MONETARY FUNDAMENTALS IN A SMALL
MULTI-COUNTRY PANEL

J.J.J. Groen

This paper has benefitted from comments by Peter van Els, Peter Vlaar and seminar participants at De
Nederlandsche Bank.

De Nederlandsche Bank NV
Econometric Research and
Special Studies Department
Research Memorandum WO&E no. 664/0119 P.O. Box 98
1000 AB AMSTERDAM
August 2001 The Netherlands
 ABSTRACT

(EURO) Exchange Rate Predictability and Monetary Fundamentals in a Small Multi-Country Panel

J.J.J. Groen

In this paper a panel of vector error correction models based on a common long-run relationship is
utilized to test whether the DM exchange rates of Canada, Japan and the United States comply in the
long-run with a rational expectations-based monetary exchange rate model. Compared to existing coin-
tegration frameworks our approach indicates that the aforementioned exchange rates are indeed consis-
tent with the monetary exchange rate model based on a common long-run relationship. We also analyze
the out-of-sample fit of this common long-run exchange rate model relative to naive random walk-based
forecasts through several forecasting evaluation measures. These forecasting evaluations indicate that
the monetary model-based common long-run model is superior to both random walk-based forecasts and
standard cointegrated VAR model-based forecasts.

Keywords: Panel cointegration testing, nominal exchange rates, exchange rate predictability.

JEL Codes: C12, C23, F31

SAMENVATTING

(EURO) Wisselkoersvoorpelbaarheid en Monetaire Fundamentele Variabelen in een Kleine Panel Data-

set over Meerdere Landen
J.J.J. Groen

In dit rapport wordt gebruik gemaakt van een panel van vector foutencorrectiemodellen op basis van een
gezamelijke lange termijn relatie om te toetsen of het lange termijn gedrag van de DM wisselkoersen van
Canada, Japan and de V.S. in overeenstemming is met het monetaire wissekoersmodel. In tegenstelling
tot bestaande methoden geeft een dergelijke benadering aan dat de voornoemde wissekoersen een lange
termijn gedrag kennen die in overeenstemming is met het monetaire model. Ook wordt de voorspel-
baarheid buiten de steekproef van deze gezamelijke lange termijn wisselkoersrelatie ten opzichte van
het toevalswandelingmodel van de wisselkoers geanalyseerd aan de hand van verschillende maatstaven.
De resultaten van deze analyse geven aan dat het op het monetaire model gebaseerde gezamelijke lange
termijnverband van de wissekoers op lange termijn een superieure voorspelkracht heeft ten opzichte van
het toevalswandelingmodel en standaard tijdreeksmodellen.

Trefwoorden: Panel cointegratie, nominale wisselkoersen, wisselkoersvoorspelbaarheid.

JEL Codes: C12, C23, F31

 –1–

1 INTRODUCTION

If monetary exchange rate models are valid representations of long-run exchange rate behaviour then

the exchange rate will only wander off from its monetary fundamentals in the short-run. The purpose

of this study is twofold. First, we want to establish whether the monetary exchange rate model has

any empirical validity for the behaviour of major exchange rates. Also, we want to exploit the effect

that exchange rate movements in the long-run are tied by monetary fundamentals in order to forecast

future exchange rate behaviour, especially vis-à-vis alternative approaches. Mark (1995) and Chinn and

Meese (1995) already tried to exploit the aforementioned effect and they claim that current monetary

fundamentals-based disequilibria can predict the exchange rate four years ahead, both in-sample and

out-of-sample for the U.S. dollar exchange rates of Canada, Germany, Japan, Switzerland and the U.K.

over the 1973-1991 period. This long horizon predictability, however, breaks down in Groen (1999)

for out-of-sample tests on these exchange rates within a 1973-1994 sample. Groen relates his result

to the absence of cointegration between exchange rates and monetary fundamentals, i.e. a stationary

linear combination of the exchange rate and its monetary fundamentals is absent. Hence, the issue of

cointegration is of major importance in establishing a predictable link between exchange rates and their

monetary fundamentals.

Studies like MacDonald and Taylor (1993) have succesfully tested the validity of monetary exchange

rate models through cointegration techniques, especially within the cointegrated vector autoregressive

[VAR] framework of Johansen (1991). They find evidence for cointegration between the log of the

exchange rate and the log of the fundamentals in a forward-looking monetary exchange rate model for

the U.S. dollar-Deutsche Mark exchange rate over the period 1976-1990. Based on the same techniques,

however, Sarantis (1994) and Groen (2000) do not find evidence for cointegration based on the monetary

model for a large number of OECD-countries’ exchange rates relative to the U.S. dollar, the Deutsche

Mark [DM] and the pound-sterling. On the other hand, Groen (2000) finds that the use of cross-section

regressions for a large number of countries, or tests for cointegration within a fixed individual effect

multi-country panel data model, results in more empirical evidence in favor of the monetary exchange

rate model. Hence, the analysis of the monetary exchange rate model across a multiple of countries

simultaneously seems to provide more positive results relative to the single country monetary model.
 –2–

Inspired by the multi-country panel-based results in Groen (2000), Mark and Sul (1998) both test for

cointegration based on the monetary model and for exchange rate predictability in a sample of bilateral

U.S. dollar exchange rates for 17 OECD countries over the period 1973-1995. The results in Mark and

Sul indicate that there is cointegration based on the monetary model and that monetary fundamentals

significantly predict future exchange rate returns using panel regression estimates with fixed time ef-

fects. Note that the panel cointegration framework described in Groen (2000) and Mark and Sul (1998)

are essentially pooled versions of the two-step Engle and Granger (1987) procedure and therefore only

allows for a limited amount of cross-country heterogeneity. From the power analyses in Groen (2000)

and Groen (2001) it becomes clear that significant cross-country differences in short-run dynamics may

impede the finding of cointegration for panels with a small number of countries within this panel Engle-

Granger framework. One way to deal with this problem is to use the panel cointegration framework

developed by Groen and Kleibergen (2001) which is based on a panel of the individual vector error

correction [VEC] models across the constituting countries. An attractive feature of the aforementioned

framework is that one can test for a common number of cointegrating vectors per country where the error

correction coefficients are assumed to be heterogeneous. Hence, we can test whether or not the long-run

relationships are identical across the countries. The power analysis in Groen (2001) indeed confirms that

the Groen and Kleibergen (2001) framework is better able to deal with heterogeneous short-run dynam-

ics in a panel with a limited amount of countries than the approach followed by Groen (2000) and Mark

and Sul (1998).

In this paper we analyze the empirical validity of the monetary exchange rate model for the DM ex-

change rates of Canada, Japan and the U.S. for quarterly data covering the 1975-2000 period. The recent

controversy with respect to the volatile behaviour of the external value of the Euro provided a rationale

for the use of DM exchange rates. We follow Arnold and de Vries (2000) who suggests that the his-

torical behaviour of EMU money demand can best be approximated by extrapolating the behaviour of

the country which mimics the ECB line of conduct best instead of constructing artificial synthetic EMU

data. Hence, we chose DM exchange rates as the subject of our empirical analysis.

From the results in Groen (1999) and Berkowitz and Giorgianni (2001) it becomes clear that, first, one

has to have cointegration based on the monetary exchange rate model in order to find a significant out-

of-sample fit of this structural exchange rate model. We therefore first test whether there is cointegration
 –3–

based on the monetary exchange rate model for the entire sample both through the purely time series-

based Johansen (1991) cointegrated VEC framework as well as through the panel cointegration frame-

work from Groen and Kleibergen (2001). Next, following the seminal work of Meese and Rogoff (1983)

we analyze the out-of-sampe performance of our cointegrated and panel cointegrated exchange rate mod-

els relative to that of the random walk model. One can also argue that in an efficient currency market the

out-of-sample fit of the cointegrated and panel cointegrated exchange rate models should be measured

in terms of the way in which the long-run forecasts of the composing series are bounded together. Also,

present-day monetary model-based error correction terms should have predictive power for future growth

rates of the monetary fundamentals. In order to analyze these aspects of the models we also utilize the

multivariate forecasting evaluation measure as proposed by Christoffersen and Diebold (1998).

In the remainder of this paper, Section 2 summarizes the cointegration restrictions implied by the mon-

etary exchange rate model and shows what its implications are in an efficient market context. Section 3

discusses the two (panel) cointegration procedures used, followed by an application of the two proce-

dures on our data. The methodology and the results of the out-of-sample forecasting evaluation of the

monetary exchange rate model can be found in Section 4. Finally, Section 5 contains conclusions.
 –4–

2 THE MONETARY EXCHANGE RATE MODEL AND EFFICIENT MARKETS

Assume that both purchasing power parity [PPP],

  
pt
ϑ pt st ϖt  (1)

and uncovered interest rate parity [UIP],

 
Et st  1   st
it  it   (2)

hold. In (1) ϑ is an intercept term, ϖt is a stationary zero-mean disturbance, pt and pt are the logarithms

of the home and foreign aggregate price indexes respectively, and st is the log exchange rate of the number

of domestic currency units in terms of one foreign currency unit. The symbol Et in (2) is the expectations

operator based on the currently available information set, and it and it are the nominal interest rates at

home and abroad. The aggregate price levels are determined through the quantity relation Mt Vt
Pt Yt or

in logarithms for both the domestic and foreign cases:

 
mt vt
pt yt
(3)
 
mt vt
pt yt 

with mt (mt ), vt (vt ), pt (pt ) and yt (yt ) the logarithms of the domestic (foreign) money supply Mt (Mt ),

the domestic (foreign) money velocity Vt (Vt ), the domestic (foreign) price level Pt (Pt ) and domestic

(foreign) real income Yt (Yt ), respectively. Money velocity measures how many times a unit of money

is spend in period t, i.e Vt
Pt Yt 
Mt . Define now the demand for real balances in logarithms at home

and abroad as
 
mt  pt
η φyt  ωit νt
(4)
 
mt  pt
η φyt  ωit νt 

where the income elasticity φ  0, the interest semi-elasticity ω  0 and νt , νt are stationary, zero mean

disturbances.1 Given the definition of money velocity and (4), we can write the log money velocity

1 In the empirical analysis we shall use M3 aggregates on a quarterly basis. The interest rate used here is therefore

a short-term interest rate and as a consequence it may also serve as the own return on M3, implying a positive
effect on money demand. Hence, the sign of the interest could be ambiguous. However, as will become clear later
on the interest rate vanishes from our relationships and thus the sign of it in (4) is not of real importance.
 –5–

function as a linear function of log real income and the nominal interest rate it :

vt  η  φ 1  yt  ωit  νt 
(5)
vt  η  φ 1 yt  ωit  νt 

Combining (1) and (2) with (3) and (5) yields a rudimentary monetary exchange rate model:

st  c   mt mt  φ  yt yt   ωEt  ∆st  1   ζt  (6)

where c  η η ϑ, ζt  νt νt ϖt , and ∆   1 L  is the first difference or lag operator. Assuming

rational expectations and using the law of iterated expectations, recursive forward substitution based on

(6) results in:2

∞
 ∞

1 ω i
1 ω i

1  ω i∑ 1  ω i∑
st  Et  ft  i    c  Et  ζt  i  
 0 1  ω  0 1  ω
 (7)
1 ∞ ω i
 c  1 ω ∑ 1 ω Et  ft  i  
 i 0 

where:

ft   mt mt  φ  yt yt  

The asset market approach implies, through (7), that when the monetary fundamentals are I(1) the ex-

change rate also has a unit root. However, this approach does allow for temporary deviations between the

log exchange rate st and the monetary fundamentals ft . This result can be derived, in accordance with

Campbell and Shiller (1987) and Campbell et al. (1997, Chapter 7), by subtracting ft from both sides of

equation (7), and using

∞

1 ω i

1  ω i∑
ft  ft 
 0 1  ω

2 See also Mussa (1976).

 –6–

so that

∞
1 ω i

ω∑
st  ft  c  Et ft !  ft "
1  1  ω
i
i 1
∞
1 ω i i

1  ω i∑ ∑ Et
 c ∆ ft ! j "
 1  1  ω j 1
∞ ∞
(8)
1 ω i
 c  1  ω ∑ Et ∆ ft ! j " ∑ 1  ω 
j 1 i j 
∞
ω j
 c ∑ Et ∆ ft ! j " #
j 1  1  ω

Based on the spread in (8) the monetary exchange rate model has the following implications within

efficient foreign exchange markets:

1 through the specification of st  ft in (8) it becomes obvious that exchange rate dynamics is domi-

nated by “news"and thus exhibits a low signal-to-noise ratio,

2 st  ft is stationary and thus st , mt  mt$ " and yt  yt$ " are cointegrated with one cointegrating

vector or equilibrium relationship,

3 in st  ft the cointegrating vector complies with the following restrictions: β % 1  1 φ" & ,

4 the currently observed exchange rate level st has predictive power for the future growth rates in

mt  mt$ " and yt  yt$ " .

 –7–

3 COINTEGRATION TESTS OF THE MONETARY MODEL

Several methods are available to test whether or not the log exchange rate st is cointegrated with the log

monetary fundamentals ' mt ( mt) * and ' yt ( yt) * . Traditionally, these methods are based on time series

data and the most freqently used method is the VAR-based framework of Johansen (1991). However,

Otero and Smith (2000) show in Monte Carlo studies that the power of this particular approach to reject

the null of no cointegration in the face of a persistent alternative depends on the span of the data sample.

As the post-Bretton Woods sample covers a quite short time span one would expect that methods such

as those of Johansen (1991) have difficulty in verifying the cointegration restriction as summarized in

(8). Alternatively, one could apply panel-based techniques in which inference is based on an artificially

extended number of observations. In Section 3.1 we discuss such a panel-based approach and contrast

it with the time series-based framework of Johansen (1991). Both cointegration frameworks are subse-

quently applied in Section 3.2 on our sample of exchange rates and monetary fundamentals of Canada,

Japan and the U.S. relative to Germany in order to test the cointegration restriction of the monetary

exchange rate model within an efficient foreign exchange market.

3.1 Cointegration Tests for Time Series and Panels

For testing for cointegration between st , ' mt ( mt) * and ' yt ( yt) * we can use the VEC framework of

Johansen (1991), i.e.

3 p1 1
∆Xt + ∑ χs D̄s - α . β/ ( β0/ 0 Zt 1 1- ∑ Γ j ∆Xt 1 j- εit 2 (9)
s, 1 j, 1

In (9) the 3 3 1 vector Xt equals

Xt + ' st ' mt ( mt) * ' yt ( yt) * * / 4

∆Xt + Xt ( Xt 1 1 , Zt 1 1 + ' Xt/ 1 1 1 * / , D̄s is a zero-mean seasonal dummy and εit is a 3 3 1 vector of white

noise disturbances. The 1 3 r vector β0 is a vector of intercept terms, α and β are 3 3 r matrices of

adjustment parameters and cointegrating vectors, respectively, and r is the cointegrating rank value of
 –8–

VEC model (9). Note that this specification of the deterministic part of (9) implies that the intercepts

appear only in the long-run relationships, which is in compliance with (8).

The Johansen (1991) likelihood ratio statistic for the null of r cointegrating vectors versus the alternative

of a stationary VAR, which has a non-standard asymptotic distribution, can be used to determine the

poper value of the cointegrating rank r in (9). After the proper cointegrating rank is determined likeli-

hood ratio tests are used to test restrictions on the r cointegrating vectors. As these tests are conducted

conditional on the cointegrating rank they have standard limiting distributions. Validity of the monetary

model within VEC (9) implies a reduced rank value r 5 1 and a cointegrating vector, normalized on st ,

equal to 6 β̄798 β̄70 : 5%6 1 8 1 φ 8 c: , which complies with 8 φ ; 0 in (7).

VEC model (9) can therefore be used to separately test the cointegration restriction in (8) for each

bilateral exchange rate in our sample. However, both the log exchange rates sit and the corresponding

spreads sit 8 fit exhibit in reality co-movements across countries due to contemporaneous correlation.

This contemporaneous correlation is due to the fact that we analyze bilateral exchange rates relative

to the same numeraire and movements related to the base country therefore induces contemporanaous

correlation. Also, common factors such as herd behaviour, contagion and so on, which are not part

of our model, can cause the aforementioned co-movements. Groen (2000) shows that within a cross-

section of 14 OECD countries long-run DM exchange rate changes are related to long-run changes in

the corresponding monetary fundamentals based on common parameter values. Hence, based on these

arguments one has to analyze all the bilateral exchange rates in our sample simultaneously as applying

the VEC (9) for each exchange rate separately will then be based on inefficiently estimated parameters

and thus decrease the power of this approach. Alternatively like Groen and Kleibergen (2001) one can

stack VEC models like (9), constructed for each of our N exchange rates, into one system:

∆X1t 5 α1 ? β71 8 β70 = 1 @ Z1 = t A

A
1 χ1 = s D̄s ∑ j< 1 1 Γ j = 1 ∆X1 = t A > ε1t
p 1
∑3s< > 1 > j
.. .. .. .. ..
. . . . . (10)
p A 1
∆XNt 5 ∑3s< 1 χN = s D̄s > αN ? β7N 8 β70 = N @ ZN = t A 1 > ∑ j < N 1 Γ j = N ∆XN = t A j > εNt

and estimate this system under the assumption of a common cointegrating rank r1 5CB B B 5 rN 5 r. Groen

and Kleibergen (2001) show that maximum likelihood estimation of the parameters in a panel VEC

model such as (10) can be achieved by applying iterative Generalized Method of Moment [GMM] es-
 –9–

timators. Again, as in the time series approach of Johansen (1991), the corresponding likelihood ratio

tests on the common cointegrating rank value r have a non-standard limiting distribution. We consider

and test different possible specifications of the cointegrating vectors in our panel VEC model which can

be defined as:

Definition 3.1. In the N-country panel VEC model which corresponds with (10) we test the following

hypotheses:

– B(r D E A tests the null hypothesis that for each cross-sectional unit i the cointegrating rank value

equals r versus the alternative hypothesis that for each cross-sectional unit i we have a full rank value

while not assuming β1 FCG G G F βN in (10).

– C H r D E A tests the null hypothesis that for each cross-sectional unit i the cointegrating rank value

equals r versus the alternative hypothesis that for each cross-sectional unit i we have a full rank value

assuming β1 FCG G G F βN F β in (10).

When the monetary exchange rate model is appropriate within panel VEC (10) we thus expect to find the

following test results

– (10) has a common cointegrating rank value of r F 1 combined with β1 FIG G G F βN F β,

– the common cointegrating vector must comply with the parameter restrictions of the monetary

model, and these restrictions are tested through a likelihood ratio statistic which is conditional on the

common cointegrating rank r and thus has a standard (χ2 ) asymptotic distribution.

3.2 Testing the Cointegration and Parameter Restrictions

The first step in the empirical analysis of the monetary exchange rate model is to test the cointegration

and parameter restrictions implied by this monetary model, see (8). The panel VEC model (10) and, as a

benchmark, the pure time series framework of Johansen (1991) are used to test the validity of the afore-

mentioned restrictions. We analyze the quarterly Euro exchange rates and the accompanying monetary

fundamentals of Canada, Japan and the U.S. for the period 1975-2000. These monetary fundamentals

are based on M3 monetary aggregates and real GDP, where M3 was preferred over M1 as the M3-based
 – 10 –

results provided more stable estimation results.3

The focus on the Euro-data raises the issue of what data one should use for the pre-EMU era which

started in 1999. It is common practice to use weighted averages of the individual data of the member

countries. However, as formalized by Arnold and de Vries (2000) the use of these weighted averages can

overstate the stability in the underlying relative money demand function on which the monetary exchange

rate model is based. Crucial in this line of reasoning is the stability of intra-EMU exchange rates. In

the pre-EMU era these intra-EMU rates were not stable and thus the law of large numbers applies for

the aforementioned weighted averages through which we find a stable synthetic EMU money demand

function. Intra-EMU rates, however, are stable from 1999 onwards resulting in strong co-movements of

member state money aggregates and because of this the law of large numbers no longer applies. Hence,

in the true EMU period area-wide money aggregates exhibit per definition a different behaviour than

during the pre-EMU period. As an alternative Arnold and de Vries (2000) and Hubrich and Vlaar (2001)

propose to extrapolate the behaviour of the country whose central bank design is closest to the design of

the European Central Bank. As there is a consensus in the literature that the ECB design is inspired by

that of the Bundesbank, we approximate the pre-EMU behaviour by German behaviour.

We first construct VEC models as in (9) for Canada, Japan and the U.S.4 The lag order p for each of these

countries is determined through the minimum of the Akaike Information Criterion [AIC] computed for

each of p J 0 K L L L K 8. After estimating the VEC model in its unrestricted form we construct, sequentially,

the Johansen (1991) likelihood ratio tests or ‘trace statistics’ of the null hypothesis r J 0 K L L L K 2 versus

the alternative r J 3 (i.e. a stationary VAR), and these are reported in Table 1. The results in Table 1

show that except for Japan we find that the log exchange rate is cointegrated with the log monetary

fundamentals as the null of no cointegration is rejected and the null of 1 cointegrating vector is accepted

for Canada and the U.S.

The discussion of the monetary model in Section 2, however, showed that cointegration based on a single

equilibrium relationship is not enough for the monetary model to be empirically valid. It should also be

the case that within the cointegrating vector β JNM βs β O m P mQ R β O y P yQ RTS β0 U V the restriction βs W β O m P mQ R J 0

3 For a more detailed description of the data and its sources see the Data Appendix.
4
Unreported unit root tests indicate that st , X mt Y mtZ [ and X yt Y ytZ [ are I(1) for our three exchange rates, which is
in compliance with e.g. MacDonald and Taylor (1993) and de Vries (1994).
 – 11 –

should be valid, which implies that the normalized cointegrating vector corresponds with
\] _`ba%c
β̄
1 ^ 1 φ ^ cd e f (11)
^ β̄0
a
a
as suggested in Section 2. We compute the likelihood ratio statistic for H0 : βs g β h m i mj k 0 versus

H 1 : βs g β h m i mj a these are reported under the column heading ‘Mon.

kml 0 for Canada and the U.S., and
El.’ in Table 1. As the restriction βs g β h m i c mj k 0 only applies to one parameter in the cointegrating

vector its asymptotic distribution equals a χ2 1d distribution. Based on this distribution the results in the

last column of Table 1 indicate that we have to reject the parameter restriction of the monetary exchange

rate model. Thus, at different stages we have to reject the empirical validity of the monetary exchange

rate model for the DM-exchange rates of Canada, Japan and the U.S. within an individual analysis.

However, deviations with respect to the monetary model can be persistent enough such that they are very

hard to deal with for an individual exchange rate given the short post-Bretton Woods time span. An

alternative would be to analyze our three exchange rates simultaneously within a stacked system or panel

of the individual VEC models such as in (10). Firstly, we test the common cointegration rank value across

Canada, Japan and the U.S. within (10) through the common cointegration rank likelihood ratio statistics

of Groen and Kleibergen (2001) based on the lag orders of the individual systems in Table 1.5 These

statistics are reported in Table 2 and the first row of this table contains both the computed value of the

likelihood ratio statistic for the null of no cointegration across the three countries and the corresponding

critical values. The result in this first row indicates that the null of no cointegration across the countries

has to be rejected. a In the second row of Table 2 we report the test results for the null of a common

cointegation rank r 1 based on country specific cointegating vectors. From the values in this row one

can observe that this null cannot be rejected. An interesting testable hypothesis is to test whether the

single cointegrating vectors of Canada, Japan and the U.S. are, apart from the intercept term, identical

a
across these countries. The fourth row of Table 2 contains the outcome of a likelihood ratio test of the

joint hypothesis of a common cointegrating rank r 1 combined with a common cointegrating vector

(apart from the intercept). This test reveals that we have across Canada, Japan and the U.S. cointegration

based on a single cointegrating vector per country with parameters which are homogeneous across the

5 The usage of the individual lag orders is based on the ‘bottom-up’ modelling strategy for restricted VAR models
from Lütkepohl (1993, pp.182-183) such that the appropriate lag orders for the panel VEC model are set equal to
the optimal AIC-based lag order for each country-specific VEC sub-model individually.
 – 12 –

exchange rates.

Although tests on the empirical equivalent of (10) point to the acceptance of the cointegration restriction

of the monetary exchange rate model, we now also have to check whether we find evidence for the

parameter restrictions of the monetary model. The first column of Table 3 contains the normalized

cointegrating vector for (10) estimated under r n 1 and a common cointegrating vector. A relative money

elasticity of 0.83 is close to the theoretical value of 1, whereas the relative income elasticity of o 3 p 02 has

a sign which is in compliance with the theory.6 In order to test whether or not the relative money elasticity

equals the theoretically valid value of 1, we conduct the likelihood ratio test for βs q β r m s mt u n 0 within
the common cointegrating vector and the corresponding result can be found in the last row of Table 3.

This test result verifies that the estimated relative money elasticity is not significantly diferent from

1, and the corresponding normalized common cointegrating vector can be found in the last column of

Table 3. Hence, our panel VEC model for the DM-exchange rates of Canada, Japan and the U.S. provides

empirical evidence for the validity of both the cointegration and parameter restrictions of the monetary

exchange rate model.

6 Although the sign of the relative income elasticity is appropriate, the absolute value of this elasticity is much

higher than in previous studies. This is most likely caused by other factors not included in our monetary model,
such as the Balassa (1964)-Samuelson (1964) effect.
 – 13 –

4 PREDICTING EXCHANGE RATE MOVEMENTS

Since the seminal paper of Meese and Rogoff (1983) it has become common practice in international

finance to subdue structural exchange rate models to an out-of-sample forecasting competition vis-à-

vis non-structural models such as the random walk model. This section deals with the out-of-sample

performance of our monetary model-based panel VEC model for the DM-exchange rates of Canada,

Japan and the U.S. relative to both non-structural models and purely time series-based approaches. The

methodology through which we measure this out-of-sample performance is described in Section 4.1.

Section 4.2 presents the results of our forecasting evaluations.

4.1 Methodology

Meese and Rogoff (1983) compared post-sample predictions of several monetary exchange rate models

for the log level of the exchange rate with those of a random walk or ‘no change’ model at forecasting

horizons up to 1 year. Mark (1995) and Chinn and Meese (1995) conducted a similar exercise in which

they compared the out-of-sample exchange rate change predictions of current error-correction terms with

the predicted change, i.e. a zero change, of the random walk model at horizons up to 4 years. As this

has become standard in exchange rate economics, we shall also follow this approach and compare the

out-of-sample exchange rate forecasts (in levels) of (10) for Canada, Japan and the U.S. with those of a

random walk. Our evaluation criterion for the log exchange rate level is the root of the mean of squared

forecast errors [RMSE]

T| h
z wy 1
y T { t0 { h t∑
RMSE vx e2s ~ t  h€ (12)
} to

where t0 is the first observation in the forecast period, h is the forecasting horizon and es ~ t  h is the

forecast error of the model-generated prediction of the log exchange rate level relative to the observed

log exchange rate level.

If our monetary model-based panel VEC model has empirical validity it should outperform, according to
 – 14 –

the aforementioned approach, the random walk-based forecasts. However, these two approaches both im-

pose an identical order of integration, i.e. I(1), which resembles the order of integration of log exchange

rates under a floating regime. For such a case Christoffersen and Diebold (1998) prove that based on

commonly used forecasting evaluation measures, such as the mean of squared forecasting erros [MSE],

cointegrated VAR model-based forecasts and forecasts based on unrelated integrated series have an iden-

tical forecasting performance. Thus, we would expect that our monetary model-based panel VEC model

and the random walk model behave identically for the log exchange rate out-of-sample. Next, within an

efficient foreign exchange rate market one would not expect that, if (long-run) currency pricing in the

market is based on the monetary model, currency forecasts based on the current level of the exchange

rate should be significantly worse than monetary model-based forecasts. This is due to the fact that the

current exchange rate level should on average reflect the true future developments in the monetary funda-

mentals, see (7) in Section 2. Hence, one may doubt whether a forecasting contest between the monetary

model-based panel VEC model and a naive random walk model for the exchange rate properly measures

the success of the monetary exchange rate model in an out-of-sample context within an efficient market.

As an alternative approach we measure the systemic implications of a valid monetary model within an

efficient currency market. Based on the line of reasoning in Section 2 these systemic implications are

that

1 the out-of-sample forecasts for st ,  mt ‚ mtƒ „ and  yt ‚ ytƒ „ generated by the monetary model-based

panel VEC model should move together in the long-run,

2 the spread or error correction term  st ‚ ft „ predicts future growth rates of  mt ‚ mtƒ „ and  yt ‚ ytƒ „ .

One criterion which is able to measure whether or not our forecasting models comply with the two above

mentioned implications is the ‘triangular MSE’ criterion of Christoffersen and Diebold (1998):

T‡ h
1
T ‚ t0 ‚ h t∑
MSEtri … trace † Aet ‰ h etŠ ‰ h AŠ ‹Œ (13)
ˆ t0

In (13) et ‰ h …  es  t ‰ h eŽ m‡ m   t‰ h eŽ y ‡ y   t ‰ h „ Š is the k ‘ 1 vector of forecast errors for the log levels of the

exchange rate and the monetary fundamentals, and the k ‘ k matrix A equals
–—
φ —
1 ‚ 1
A …“’”
”•
0  1 ‚ L„ 0 ˜ Œ
0 0  1 ‚ L„
 – 15 –

Using the random walk model of the log exchange rate level as a benchmark for our monetary model-

based forecasts is, however, inappropriate when we analyze the forecasting behaviour through (13) as

such a benchmark does not say anything about the behaviour of the monetary fundamentals. Therefore,

under (13) we use as a benchmark for the monetary model-based panel VEC model

∆sit ™ εsit
pi
m ¡ m£ ¤
∆ š m › mœ  it ™ δi1 ž ∑ γi1 j ∆ š m › mœ  i  t ¡ j ž εit¢
jŸ 1 (14)
pi
y ¡ y£ ¤
∆ š y › yœ  it ™ δi2 ž ∑ γi2 j ∆ š y › yœ  i   t ¡ j ž εit¢ ¥ i ™ 1¥ ¦ ¦ ¦ ¥ N t ™ 1 ¥ ¦ ¦ ¦§¥ T ¥
jŸ 1

and we estimate (14) across i ™ 1 ¥ ¦ ¦ ¦ ¥ N with feasible Generalized Least Squares [FGLS]. The restricted

VAR model (14) imposes the same order of integration as our monetary panel VEC model, but each of the

exchange rate and the monetary fundamentals are assumed to be unrelated, i.e. they are not cointegrated.

In (14) the log exchange rate is, as in the standard approach, assumed to be generated by a random walk.

The forecasts are generated in a recursive manner. Suppose that our first forecast has to be generated

at observation t0 (t0 ¨ T ). Consequently, we first estimate each of our forecasting models on a sample

which runs up to t0 .7 Based on these estimates we generate our forecasts for the levels at all forecasting

horizons h, where for h © 1 the forecasts in case of the monetary panel VEC model and (14) are generated

in a dynamic manner, i.e. the forecasts for the previous quarter are used to generate the forecasts for the

current quarter. These two steps are repeated for the observations t0 ž 1 ¥ t0 ž 2 ¥ ¦ ¦ ¦ ¥ T › h. In order to be

able to evaluate the behaviour of our monetary model-based panel VEC model, we construct based on

our recursively generated predictions the ratio of RMSE (12) for the monetary panel VEC model relative

to the random walk and the ratio of MSEtri (13) for the monetary panel VEC model relative to (14) across

i ™ 1 ¥ ¦ ¦ ¦ ¥ N. For our monetary model-based panel VEC model to be valid these ratios should be smaller

than 1.

7 Obviously, the random walk model for st does not involve any estimation.
 – 16 –

4.2 Forecasting Evaluation

We evaluate the forecasting performance of our monetary model-based panel VEC model, i.e. (10) with

r ª 1 based on a common cointegating vector and a unity relative money elasticity imposed, across to

forecast periods 1989.I-1998.IV and 1989.I-2000.IV. In the first forecast period we solely evaluate the

monetary panel VEC model with pre-EMU data, whereas EMU data are also included in the evaluation

in the second forecast period. As in the recent papers on exchange rate predictability such as Mark (1995)

and Chinn and Meese (1995), we consider as forecasting horizons (in quarters) h ª 1, 4, 8, 12 and 16.

Mark (1995) and Chinn and Meese (1995) concluded that at forecasting horizons of three to four years

monetary fundamentals predict exchange rate movements better than the random walk model. Groen

(1999) and Berkowitz and Giorgianni (2001), amongst others, showed that these results were spurious

as these forecasts were not based on cointegrated models. This is the reason why we did the elaborate

cointegration analysis on the monetary model for our three DM-exchange rates in Section 3.2, and based

on the results in that subsection we can expect to have positive results for exchange rate predictions

based on the panel VEC model. The results for the exchange rate level predictions in the first forecasting

period can be found in the first two columns of Table 4. These two columns report RMSE ratios for

the log exchange rate level of an individual time series VEC model (first column) or the monetary panel

VEC model (second column) relative to random walk forecasts. From these results it can be seen that at

forecasting horizons of three and four years only the monetary panel VEC model outperforms the ran-

dom walk model. Qualitatively, the results for the second forecasting period, see the first two columns of

Table 5, are the same. These results seem at first sight to indicate that in the long-run there is predictabil-

ity in exchange rate movements based on monetary fundamentals. However, Berkowitz and Giorgianni

(2001) elegantly show that exchange rate predictability at horizons h « 1 depend on the existence of

exchange rate predictability at h ª 1. From Tables 4 and 5 it is clear that even within the monetary panel

VEC models there is no exchange rate predictability at h ª 1, and thus the long horizon results are not

feasible. On the other hand, Christoffersen and Diebold (1998) show that as long as the proper order of

integration is imposed different forecasting models will exhibit on average an identical predictive per-

formance. Hence, measuring the out-of-sample fit of our monetary panel VEC model in terms of the

exchange rate solely does not provide us with a definite source regarding the empirical validity of the
 – 17 –

monetary exchange rate model.

Alternatively, one can measure whether or not our monetary panel VEC model mimics the systemic

predictive features as implied by an efficient currency market, see Sections 2 and 4.1, in an out-of-

sample context through (13). Christoffersen and Diebold (1998) show in general for cointegrated VARs

that under the true hypothesis of cointegration the MSEtri ratio (13) of the cointegrated VAR relative to

univariate time series forecasts (such as in (14)) becomes smaller than 1 as h ¬ ∞. Therefore, from the

results for our first forecasting pre-EMU evaluation period in the last two columns of Table 4 we can

conclude that the monetary panel VEC model behaves in compliance with an efficient currency market.

From this table it can be seen that for all three exchange rates the MSEtri ratio achieves a value smaller

than 1 when the forecasting horizon h increases, especially in the case of Canada. This situation does

not change when we consider the second forecasting sample, which includes the ‘real’ EMU era, as

becomes apparent from the last two columns from Table 5. Note that as in the case of the RMSE ratios

the purely time series based VEC models perform severely worse than either the random walk model

or the univariate system (14) which confirms the conclusion from Groen (1999) that forecasting models

should be based on empirically valid cointegration restrictions.

 – 18 –

5 CONCLUSIONS

We investigate in this paper both the in-sample as well as the out-of-sample fit of the monetary exchange

rate model for the Euro-exchange rates of Canada, Japan and the U.S. We follow both the standard pure

time series-based cointegrated VAR approach and the panel VEC approach of Groen and Kleibergen

(2001) to analyze the cointegration and long-run parameter restrictions of the monetary exchange rate

model. Our empirical investigation indicates that modelling the three aforementioned Euro-rates simul-

taneously within a panel structure is necessary in order to validate the empirical appropriateness of the

monetary exchange rate model.

The out-of-sample forecasting evaluation confirms that a multi-country panel data structure is necessary

in order to find any empirical evidence pro the monetary model. A comparison of the monetary model-

based panel VEC exchange rate forecasts with random walk model exchange rate forecasts shows that

the monetary model-based panel approach yields superior forecasts at horizons of three and four years.

This predictive superiority vis-à-vis random walk forecasts even occurs at shorter horizons in the case of

the U.S. and Canada. However, within an efficient currency market it need not be the case that structural

model-based predictions outperform random walk predictions. Hence, we have analyzed the predictive

implications of an efficient currency market: forecasts of the exchange rate and the monetary funda-

mentals should be tied together in the long-run and the current level of the monetary model-based error

correction term predicts the future growth in the monetary fundamentals. The corresponding multivari-

ate forecasting evaluation criterion, derived form Christoffersen and Diebold (1998), again confirms the

empirical superiority of monetary model-based panel approach relative to both non-structural univariate

forecasts and pure time series-based versions of the monetary model.

The ESCB assigns a special role to money in the first pillar of its monetary policy strategy by announcing

a reference value for the monetary aggregate M3 and analysing actual monetary developments against

this reference path. Deviations are taken into account in the monetary policy decision process, see

Issing et al. (2001). Our empirical results, taken at face value, seem to offer some support for this

approach given the fact that exchange rates seem to be driven partly by relative monetary developments

in particular in the medium to long run (with exchange rate changes in turn feeding into domestic prices).
 – 19 –

The empirical results are also consistent with the view that deviations of current exchange rates from

their equilibrium values partly reflect expectations regarding future excess money growth rates relative

to third countries. Hence, the pre-announcement of a reference value for M3 which is consistent with the

ESCB’s definition of price stability may contribute to a development of the Euro exchange rate which

indeed supports the achievement of price stability. However, our empirical results also indicated that in

the short run the monetary influence on the exchange rate may be dominated by other factors.

In future research we first want to focus on the computation of proper standard errors of our forecasting

criteria, i.e. the RMSE and MSEtri ratios, through appropriate bootstrap experiments. Also, extensions

of the monetary model, for example by including accumulated current account deficits, may provide a

useful follow-up to the analysis presented here.

 – 20 –

DATA APPENDIX

The main source for our dataset are the International Financial Statistics [IFS] from the IMF, and we use

quarterly data which start in the first quarter of 1975 and end in the last quarter of 2000. As motivated

in the text, we use German behaviour as a proxy for EMU behaviour in the pre-EMU era, i.e. the period

before 1999. For the ‘Euro’ exchange rates of Canada, Japan and the U.S. this implies that we use the

observed Euro exchange rates, calculated for Canada and Japan through cross-rates relative to the U.S.

(based on line ‘ae’ in IFS), from 1999 onwards. For the period 1975-1998 we use the Deutsche Mark

[DM] exchange rates for these three countries multiplied with the Euro conversion rate of the DM (i.e. 1

Euro = 1.95583 DM).8

As our measure of money we use the IMF definition of M3, i.e. ‘money’ (line 34 in IFS) plus ‘quasi-

money’ (line 35 in IFS), and this measure of money is not seasonally adjusted. For our three countries

we compute the relative M3 money supplies vis-à-vis Germany for the entire sample, where German

M3 is converted in Euro’s through the fixed Euro-DM conversion rate. Real income is proxied by gross

domestic product [GDP] in volume terms (line 99B in IFS), and again the relative real income series

are constructed relative to Germany. Note that for German M3 and real GDP we correct the series for

the unification by fitting an autoregressive model to the log first differences with an unification dummy

for 1991.I (and seasonal dummies in case of M3) included. As a next step we rescale the level of the

pre-unification series with the estimated coefficient of the aforementioned unification dummy.

8 See the website of the European Central Bank, http://www.ecb.int.

 – 21 –

REFERENCES

Arnold, I.J.M. and C.G. de Vries, 2000, Endogeneity in european money demand, European Journal
of Political Economy, 16, 587–609.

Balassa, B., 1964, The purchasing power parity doctrine: A reappraisal, Journal of Political Economy,
72, 584–596.

Berkowitz, J. and L. Giorgianni, 2001, Long-horizon exchange rate predictability?, Review of Eco-
nomics and Statistics, 83, 81–91.

Campbell, J.Y., A.W. Lo, and A.C. MacKinlay, 1997, The Econometrics of Financial Markets, Prince-
ton: Princeton University Press.

Campbell, J.Y. and R.J. Shiller, 1987, Cointegration and tests of present value models, Journal of
Political Economy, 95, 1062–1088.

Chinn, M.D. and R.A. Meese, 1995, Banking on currency forecasts: How predictable is change in
money?, Journal of International Economics, 38, 161–178.

Christoffersen, P.F. and F.X. Diebold, 1998, Cointegration and long-horizon forecasting, Journal of
Business & Economic Statistics, 16, 450–458.

Engle, R.F. and C.W.J. Granger, 1987, Co-integration and error correction: Representation, estimation
and testing, Econometrica, 55, 251–276.

Groen, J.J.J. and F. Kleibergen, 2001, Likelihood-based cointegration analysis in panels of vector
error correction models, Research Memorandum WO&E 646, De Nederlandsche Bank, available at
http://www.dnb.nl, second round revision at Journal of Business & Economic Statistics.

Groen, J.J.J., 1999, Long horizon predictability of exchange rates: Is it for real?, Empirical Economics,
24, 451–469.

Groen, J.J.J., 2000, The monetary exchange rate model as a long-run phenomenon, Journal of Interna-
tional Economics, 52, 299–319.

Groen, J.J.J., 2001, Cointegration and the monetary exchange rate model revisited, mimeo, De Neder-
landsche Bank.

Hubrich, K. and P.J.G. Vlaar, 2001, Monetary transmission in germany: Lessons for the euro area,
DNB Staff Reports 57, De Nederlandsche Bank, http://www.dnb.nl.

Issing, O., V. Gaspar, I. Angeloni, and O. Tristani, 2001, Monetary Policy in the Euro Area, Cam-
bridge: Cambridge University Press.

Johansen, S., 1991, Estimation and hypothesis testing of cointegration vectors in gaussian vector au-
toregressive models, Econometrica, 59, 1551–1580.

Johansen, S., 1996, Likelihood-Based Inference in Cointegrated Vector Autoregressive Models, Oxford:
Oxford University Press, 2nd edition.

Lütkepohl, H., 1993, Introduction to Multiple Time Series Analysis, Berlin: Springer-Verlag, 2nd edi-
tion.
 – 22 –

MacDonald, R. and M.P. Taylor, 1993, The monetary approach to the exchange rate: Rational expec-
tations, long-run equilibrium and forecasting, IMF Staff Papers, 40, 89–107.

Mark, N.C. and D. Sul, 1998, Nominal exchange rates and monetary fundamentals: Evidence from a
seventeen country panel, mimeo, Department of Economics, Ohio State University.

Mark, N.C., 1995, Exchange rates and fundamentals: Evidence on long-horizon predictability, Ameri-
can Economic Review, 85, 201–218.

Meese, R.A. and K. Rogoff, 1983, Empirical exchange rate models of the seventies: Do they fit out of
sample?, Journal of International Economics, 14, 3–74.

Mussa, M., 1976, The exchange rate, the balance of payments and monetary and fiscal policy under a
regime of controlled floating, Scandinavian Journal of Economics, 78, 229–248.

Otero, J. and J. Smith, 2000, Testing for cointegration: Power versus frequency of observation - further
monte carlo results, Economics Letters, 67, 5–9.

Samuelson, P.A., 1964, Theoretical notes on trade problems, Review of Economics and Statistics, 46,
145–154.

Sarantis, N., 1994, The monetary exchange rate model in the long run: An empirical investigation,
Weltwirtschaftliches Archiv, 130, 698–711.

de Vries, C.G, 1994, Stylized facts of nominal exchange rate returns, in F. van der Ploeg (ed.), The
Handbook of International Macroeconomics, Oxford: Blackwell.
 – 23 –

Table 1 Cointegration tests for the individual countries, 1975:I-2000:IVa

Lags LR(0 ­ 3) LR(1 ­ 3) LR(2 ­ 3) MON.EL.

Canada/Germany 3 41 ® 12¯ ¯ ¯ 11 ® 98 3 ® 37 14 ® 98
°
0 ® 00±

Japan/Germany 3 27 ® 98 10 ® 03 3 ® 73 –

U.S./Germany 4 39 ® 27¯ ¯ 6 ® 96 2 ® 19 15 ® 75
°
0 ® 00±

90% 31 ® 88 17 ® 79 7 ® 50
95% 34 ® 80 19 ® 99 9 ® 13
99% 40 ® 84 24 ® 74 12 ® 73

a The column denoted with ‘Lags’ contains the order of first differences in (9) determined with
AIC. LR(r ² 3) denotes the values of the Johansen (1991) likelihood ratio test statistic for H0 :
rank(Π) = r versus H1 : rank(Π) = 3 in (9). The symbol ³ (³ ³ ) [³ ³ ³ ] indicates rejection of H0 at
the 10% (5%) [1%] significance level. The row ‘90%’ (‘95%’) [‘99%’] contains the asymptotic
90% (95%) [99%] quantile for LR(r ² 3) under the null, see Johansen (1996, Table 15.2). The
column denoted with “MON.EL."contains, if r ´ 1 is accepted, the likelihood ratio test of
the restriction β̂s µ β̂ ¶ m · m¸ ¹ ´ 0 on the non-normalized cointegrating vector (see text) and the
corresponding χ2 º 1» p-values are reported in parentheses.
 – 24 –

Table 2 Cointegration rank tests in the panel VEC

models, 1975:I-2000:IVa

Empirical 90% 95% 99%

LR ¼ B(0) ½ A¾ 95 ¿ 29À À 85.97 90 ¿ 38 99 ¿ 03

LR ¼ B(1) ½ A¾ 42 ¿ 92 46.03 49 ¿ 27 55 ¿ 76
LR ¼ B(2) ½ A¾ 11 ¿ 27 18.35 20 ¿ 63 25 ¿ 36
LR ¼ C(1) ½ A¾ 48 ¿ 61 50.70 54 ¿ 15 61 ¿ 03
LR ¼ C(2) ½ A¾ 17 ¿ 04 23.40 25 ¿ 99 31 ¿ 33

a Rows denoted with “LR Á B(r) Â AÃ "and “LR Á C(r) Â AÃ "contain

likelihood ratio tests on the null of a common cointegrating
rank value of r assuming heterogeneous and homogeneous
cointegrating vectors respectively, see also Definition 3.1. The
columns “90%"(“95%") [“99%"] are the 90% (95%) [99%]
quantiles of the appropriate limiting distribution computed
with the procedures from Groen (2001) and Ä Å Ä Ä Æ Ç Ä Ä Ä È indicates
a rejection of the null hypothesis at these quantiles.

Table 3 Normalized common cointegrating vectors

in the panel VEC model, 1975:I-2000:IVa

β̄ÉC MON.EL.
sit 1 1
Ê
mit Ë mtÀ Ì Ë 0 ¿ 83 Ë 1
Ê
yit Ë ytÀ Ì 3 ¿ 02 3 ¿ 16

Ê
LR ¼ βs Í β Î m Ï mÐ Ñ ÌÒ 0¾ 0 ¿ 29 —
Ê
0 ¿ 59Ì

a Column β̄ÓC provide the normalized estimated common

cointegrating vectors, and column “MON.EL"is the nor-
malized cointegrating vectors with proportionality between
the log exchange rate and the log relative money supply
imposed. Row LR Á Ô βs Õ β m Ö m×
Å Æ ØÚÙ 0Ã a contains like-
lihood ratio test of the restriction on the respective non-
normalized common cointegrating vectors, and the corre-
sponding χ2 Ô 1 p-value is reported in parentheses.
Ø
 – 25 –

Table 4 Forecasting evaluation of monetary fundamentals, 1989:I-

1998:IVa

RMSE Ratios MSEtri Ratios

h Ind.VEC Pan.VEC Ind.VEC Pan.VEC

Ca/Ger 1 1 Û 088 0 Û 961 1 Û 050 0 Û 913

4 1 Û 225 0 Û 973 2 Û 023 0 Û 708
8 1 Û 653 0 Û 904 2 Û 130 0 Û 494
12 1 Û 736 0 Û 837 2 Û 131 0 Û 321
16 1 Û 881 0 Û 870 2 Û 085 0 Û 232

Jp/Ger 1 1 Û 018 1 Û 068 1 Û 062 1 Û 307

4 0 Û 979 1 Û 026 2 Û 561 2 Û 489
8 1 Û 191 1 Û 133 1 Û 064 0 Û 671
12 1 Û 288 0 Û 987 1 Û 241 0 Û 840
16 1 Û 392 0 Û 983 1 Û 125 0 Û 792

U.S./Ger 1 1 Û 249 1 Û 048 1 Û 497 1 Û 061

4 1 Û 141 1 Û 030 3 Û 265 2 Û 669
8 1 Û 240 0 Û 971 1 Û 774 0 Û 846
12 1 Û 359 0 Û 946 1 Û 666 0 Û 799
16 1 Û 776 0 Û 866 1 Û 486 0 Û 679

a The entries in the table are the RMSE ratio of monetary fundamentals versus
random walk predictions in case of predicted exchange rate levels, see (12), and
the triangular MSE ratio of fundamentals versus univariate forecasts, see (13) and
(14). The forecasting horizons (in quarters) can be found under the heading “h".
Columns with “Ind.VEC"report the outcomes for the individual country VEC
models and “Pan.VEC"those for the panel VEC model.
 – 26 –

Table 5 Forecasting evaluation of monetary fundamentals, 1989:I-

2000:IVa

RMSE Ratios MSEtri Ratios

h Ind.VEC Pan.VEC Ind.VEC Pan.VEC

Ca/Ger 1 1 Ü 197 0 Ü 997 1 Ü 204 0 Ü 917

4 1 Ü 206 0 Ü 967 1 Ü 814 0 Ü 753
8 1 Ü 641 0 Ü 878 2 Ü 044 0 Ü 526
12 1 Ü 732 0 Ü 812 2 Ü 118 0 Ü 358
16 1 Ü 844 0 Ü 781 2 Ü 099 0 Ü 284

Jp/Ger 1 0 Ü 969 1 Ü 024 1 Ü 038 1 Ü 240

4 0 Ü 972 1 Ü 009 2 Ü 751 2 Ü 576
8 1 Ü 071 1 Ü 013 1 Ü 197 0 Ü 682
12 1 Ü 117 0 Ü 845 1 Ü 054 0 Ü 689
16 1 Ü 161 0 Ü 868 1 Ü 039 0 Ü 680

U.S./Ger 1 1 Ü 215 1 Ü 019 1 Ü 416 1 Ü 052

4 1 Ü 125 0 Ü 999 3 Ü 127 2 Ü 514
8 1 Ü 199 0 Ü 883 1 Ü 762 0 Ü 876
12 1 Ü 279 0 Ü 802 1 Ü 639 0 Ü 906
16 1 Ü 411 0 Ü 663 1 Ü 457 0 Ü 921

a See the notes in Table 4.