Sei sulla pagina 1di 20

Journal of International Economics 19 (1985) 47-66.

North-Holland

CONDITIONAL VARIANCE AND THE RISK PREMIUM


IN THE FOREIGN EXCHANGE MARKET
Ian DOMOWITZ
Northwestern University, Evanston, IL 60201, USA

C r a i g S. H A K K I O *
Federal Reserve Bank of Kansas City, Kansas City, MO 64198, USA

Received November 1983, revised version received September 1984


We investigate the existence of a risk premium in the foreign exchange market, based on the
conditional variance of market forecast errors. The forecast errors are assumed to follow the
ARCH process introduced by Engle (1982). Estimation and diagnostic testing of the model are
discussed, and results are presented for the currencies of the United Kingdom, France, Germany,
Japan and Switzerland.

1. Introduction
T h e existence of a risk p r e m i u m in the foreign e x c h a n g e m a r k e t p r o v i d e s
an alternative h y p o t h e s i s to the p r o p o s i t i o n t h a t the f o r w a r d rate is a n
u n b i a s e d p r e d i c t o r of the future s p o t rate, w i t h o u t sacrificing the n o t i o n o f
m a r k e t efficiency. M o t i v a t e d b y this link to questions of efficiency in the
f o r w a r d m a r k e t a n d b y a g r o w i n g b o d y of e m p i r i c a l evidence a g a i n s t the
u n b i a s e d n e s s hypothesis, there has been c o n s i d e r a b l e interest in e m p i r i c a l l y
t r a c t a b l e theories of a risk p r e m i u m . W i t h o u t such a theory, there is no w a y
to empiric/ally distinguish between a n inefficient m a r k e t a n d a t i m e - v a r y i n g
risk p r e m i u m . In this p a p e r we a t t e m p t to shed s o m e light o n this issue b y
p r e s e n t i n g a n empirical m o d e l of the risk p r e m i u m as a function of the
c o n d i t i o n a l variance of m a r k e t forecast errors.
Several m o d e l s have been p r o p o s e d t h a t g e n e r a t e a risk p r e m i u m in the
foreign exchange m a r k e t . 1 I n the t w o - p e r i o d , o p t i m i z i n g m o d e l s of S t o c k m a n
*We would like to thank Pekka Ahtiala, Robert Flood, Arnold Harberger, Robert Hodrick,
Dennis Kraft and two anonymous referees for helpful comments, and Tom Holmes for the
programming involved in estimation of the model. We would also like to thank the University
Research Grants Committee of Northwestern University for their financial support of the
project. The views expressed herein are solely those of the authors and do not necessarily reflect
the views of the Federal Reserve Bank of Kansas City or the Federal Reserve System.
tThe introduction will not be an exhaustive summary of the literature on risk premium.
Rather, we will refer to some 'typical' papers. For a more complete survey, see Levich (1983).
0022-1996/85/$3.30 1985, Elsevier Science Publishers B.V. (North-Holland)

48

I. Domowitz and C.S. Hakkio, Conditional variance

(1978) and Frenkel and Razin (1982), the risk premium depends on the
concavity of the utility function and the probability distribution of the
exogenous processes. Assuming 'that the moment of these probability distributions are relatively stable' [Stockman (1978, p. 167)1, the risk premium will
be constant. Tests for a constant risk premium are mixed: Frenkel (1978)
finds the risk premium to be statistically insignificant, Frenkel (1982) finds
the risk premium to be significant for some currencies, and Stockman (1978)
finds that the risk premium changes sign. This last result, and the findings by
Hansen and Hodrick (1983) and Cumby and Obstfeld (1982) of serially
correlated forecast errors, indicate that a time-varying risk premium is
needed to explain these deviations from simple market efficiency.
Several models exist that generate a time-varying risk premium. As
Stockman (1978, p. 167) recognizes, if the probability distributions are not
constant, then the risk premium will change. The portfolio balance models of
Frankel (1979, 1982) and Dornbusch (1982) generate a risk premium that is a
function of the quantity of nominal government bonds and the concavity of
the utility function. Frankel (1982) fails, however, to find evidence of a risk
premium. Then, there are equilibrium, dynamic, optimizing models of asset
pricing. Lucas (1982) presents such a model in an international context, and
Hodrick and Srivastava (1984) extend it to price forward contracts. The risk
premium depends on the conditional covariance of the intermporal marginal
rates of substitution and, therefore, implicitly on the concavity of the utility
function and the probability distribution function of the exogenous processes.
This fact makes empirical tests of such models quite complex. Because of
this, Hansen and Hodrick (1983) and Hodrick and Srivastava (1984) test
models that are only loosely related to the theoretical model. These papers
provide weak support for a time-varying risk premium. 2
In light of these results we find that considerable controversy exists
regarding the existence and empirical relevance of a time-varying risk
premium. As stated earlier, a major stumbling block in the estimation of a
time-varying risk premium is the lack of an economic model of the risk
premium which is empirically tractable. In section 2, the investigation of a
time-varying risk premium is motivated by an example based on the
intertemporal capital asset pricing model. The model is highly stylized,
however, and direct estimation and testing of the model is difficult. Such
difficulties were noted above. This problem is not resolved in the present
paper. Instead, section 3 presents an econometric model in which the risk
premium is a function of the conditional variance of the error in forecasting
the spot rate using the forward rate. This specification captures some major
aspects of risk in a foreign exchange contract. The forecast errors are
assumed to follow the ARCH process introduced by Engle (1982), which
2The results of Hansen and Hodrick (1983) break down when the sample period is extended.

I. Domowitz and C.S. Hakkio, Conditional variance

49

defines the conditional variance of the error as a function of prior information, which includes past forecast errors. This particular model also
captures several empirical regularities noted in the estimation of exchange
rate models, including conditional heteroscedasticity in the forecast errors
[-Cumby and Obstfeld (1982) and Hodrick and Srivastava (1984)] and fattailed behavior [Friedman and Vandersteel (1982)].
Estimation of the general model is discussed in section 4. Since estimation
requires an iterative maximum likelihood procedure, prior tests of the least
squares model are carried out and reported in section 5. Based on these
results, estimates of the complete model for the currencies of the United
Kingdom, France, Germany, Japan, and Switzerland for the period 1973 to
1982 are reported in section 6. The model is subjected to general specification tests in order to determine the adequacy of the statistical specification.
Statistical evidence concerning the existence of a risk premium is presented
and discussed in section 7. Finally, conclusions are summarized and suggestions for future research are given in section 8.

2. A model of the risk premium in the foreign exchange market


In this section an asset pricing model with a time-varying risk premium is
presented. The model is developed in Lucas (1982) and is extended by
Hodrick and Srivastava (1984). Rather than rederive the results in these
papers, a particular example is presented that yields an exchange rate
equation with a time-varying risk premium that is a function of the
conditional variance of domestic and foreign money. In addition, the error
term is heteroscedastic. The set-up of the model is as follows. There are two
countries (the United States and England), two goods (x and y) and two
monies (M and N; dollars and pounds). Consumers in the United States
receive an endowment ~t of good x and nothing of good y; consumers in
England receive nothing of good x and an endowment of r/t of good y.
Agents of each country demand both dollars and pounds, the demand being
motivated by a cash-in-advance constraint. The current period utility function,
the same for all agents, is assumed to be of Cobb-Douglas form:
U(x,y) =Ax~y 1-~.

(1)

Agents are assumed to maximize an intertemporal utility function of the


form

~, p'tJ(x,, y,)

(2)

t=0

subject to a budget constraint and two finance constraints. Finally, endowments and the money supplies are assumed to follow conditional Gaussian

I. Domowitz and C.S. Hakkio, Conditional variance

50

AR(1) stochastic processes:


In ~t = Pl In ~t- 1 "[- Ult,

(3a)

In rh = P2 In rh_ i + u2t,

(3b)

In M t = ? i In M t_ 1 + u at,

(3c)

In N t =?2 In N t_ 1 + u4t,

(3d)

u; = (ul,, u2,, us,, u,t),

(4a)

u, I I,_1 "" N ( O , H,),

(4b)

o'=(o

o o o),

Ht = diag (h 1It, h22,, ha3t, h44t).


It is assumed that there are zero covariances between the stochastic
processes? Given this model, it can be shown that, for example, ~t is
conditionally log-normal with mean exp (p 1 In ~t - 1 + ( 1/2) h 11t).
Equating the relative price of good y to the marginal rate of substitution
yields the following expression for the spot exchange rate: 4

It should be noted that, contrary to many rational expectations models of


the exchange rate [for example, Mussa (1979)], the exchange rate does not
incorporate expectations of the future, s This result follows from the relation
between the time decisions are made and the resolution of uncertainty: all
decisions are made after the resolution of uncertainty, so there is no
speculative component to the money demand function [see Lucas (1982, p.
342)].
The forward exchange rate, F~, can be derived from the interest rate parity
condition. For this example [see Hodrick and Srivastava (1984)]:
alt is possible to generalize the model to allow for nonzero covarianee. This would lead to the
risk premium also depending upon the conditional covariances.
4See Lueas (1982) or Hoddek and Srivastava (1984) for a general expression.
5To see that this is not a result of our specific example, the general equation is

s, = M, ~, u,(~, ~,)
N, ~, u,(~,,~,)"

I. Domowitz and C.S. Hakkio, Conditional oariance

EtQ~, + I

F, = st etQg 1'

51

(6)

Q~+ 1 = ~(~t+ 1/th+ ,)'/(Nt+ l/t/t + 1)


(~t/tlt)~/(Nt/rh)
,

(7a)

Q~I =/7(r/t + ,/~,+

1) 1 -"/(Mr + 1/~,+ 1)
(r/t/~t) 1 _ :l(Mtl~t )
,

(7b)

where QM 1 and Q~+I are simply the intertemporal marginal rates of


substitution of dollars and pounds.
Eqs. (5)-(7) can be used to derive linear expressions for the expectation of
the logarithm of the future spot rate, EtlnSt+l, and the logarithm of the
forward rate, In Ft: ~
Et In St +1 = In [(ct- 1)/ct] + ~I In M t - 7 2 In Nt

(8)

l n F t = l n [ ( ~ - l ) / ~ ] + ? 1 1 n M t - T z l n N , - 1 / 2 h s s , t + 1 + l / 2 h 4 4 , , + 1.

(9)

Subtracting (9) from (8) yields the following expression for the risk
premium, defined as Et In St + 1 - In Ft:
E, In S, + 1 - I n F, = 1/2 [h s s.t + 1 -h44., + 1].

(10)

Eq. (10) indicates that the risk premium depends upon the conditional
variances of the forecast errors of the domestic and foreign money supplies.
An increase in the conditional variance of domestic money, h33,t+l , increases
the conditional variance of domestic prices and, therefore, increases the
expected future purchasing power of the dollar (l/P). Consequently, the
expected return on a long position in foreign exchange (the risk premium)
must rise to compensate investors. Equivalently, there is an increase in the
demand for future dollars which drives down the forward rate, leading to an
increase in the risk premium.
Using eqs. (8) and (9) it is easy to show that
(In St + 1 - In St) = RPt + (ln F t - In St) + st + 1,
RP, = 1/2[hss.t+ i

h44,t+

(11)

11,

~t+l=U3t+l--U4t+l"
6To obtain (9) we need an equation for In E,Q M 1 and In EtQ~+1. By using the results for alognormal distribution we can show, for example, that
E,Q~ 1=exp {ln fl-ct(l - P l ) I n ~,-(1 -~)(1 -P2) In ~/t

+(i -71) In Mt+(o~2/2)hl 1,+ t + ( ( a - I)2/2)h22,+ l + hsat+ 1}.

52

I. Domowitz and C.S. Hakkio, Conditional variance

Given the stochastic processes in eq. (3), it is clear that in a regression of the
rate of depreciation on the forward premium there will be a time-varying risk
premium and the error term of the regression will exhibit conditional
heteroscedasticity.

3. An econometric model of the risk premium


The model in the previous section is highly stylized and quite restrictive in
its assumptions on preferences and technology. In spite of these limitations,
the model is capable of generating a time-varying risk premium, but it is
quite difficult to estimate the Lucas model, even in such a simplified form. In
this section a model is presented that is estimable and that captures some
aspects of risk in the foreign exchange market.
As noted in the previous section, in a regression of the rate of depreciation
on the forward premium, the risk premium depends on the conditional
variances of the exogenous variables, and the error term is heteroscedastic.
Instead of estimating such a multiple equation model [eqs. (11), (3a)--(3d)], a
simpler, single equation model is considered. The risk premium is assumed to
depend on the conditional variance of 8t in (11), as opposed to the
conditional variance of only some components of ft. The forecast errors, tt,
are assumed to follow the ARCH process introduced by Engle (1982),
described by (14) and (15) below, in which the conditional variance is a
function of past information, which includes past squared forecast errors. In
fact, "it can be shown that the error term in eq. (11) will also follow an
ARCH process, conditional on its past, if the exogenous variables, eqs. (3a)(3d), follow an ARCH process.
There are several reasons for choosing this particular representation. The
ARCH model is a convenient specification for incorpoarating heteroscedasticity into the estimation procedure. This is desirable, because Cumby and
Obstfeld (1982) and Hodrick and Srivastava (1984) provide evidence that the
forecast error is heteroscedastic. Mussa (1979, pp. 11-12) also observes that
'for many exchange rates, there appear to be periods of quiescence in which
day-to-day and week-to-week movements are very small, and periods of
turbulence in which day-to-day movements are large'. Friedman and Vandersteel (1982) show that daily exchange rate changes exhibit leptokurtosis, i.e.
fat tails, and state (p. 185) that one 'interpretation is that both the trend and
volatility o f [exchange rate] movements are affected by changing economic
and institutional factors'. These two aspects of exchange rates, periods of
quiescence followed by periods of turbulence, and fat tails, are captured by
the ARCH model.
With this introduction and justification, consider the following model:
S t + _ _- 1 St = R P t + [31 Ft - St + ~t+
St
~
1,

(12)

I. Domowitz and C.S. Hakkio, Conditional variance

53

RP, = 8o + Oh,+ 1,

(13)

e,+l II , ~ N( 0, h2+ 1),

(14)

h?+l=~0+ ~ ~i~?+l-~+z,cb,

(15)

i=1

where It represents information available at time t, and zt is some vector of


variables contained in that information set. ( S t + l - S t ) / S t and ( F t - S t ) / S t are
used instead of ln St+ l - I n St and In F t - l n S t. The correlation between the
two pairs of variables exceeds 0.999 in all cases, except for the French rate of
depreciation, in which case the correlation equals 0.988.
Eqs. (12)-(15) allow tests of several hypotheses concerning the time-series
behavior of the risk premium. According to the model in section 2, fl~ should
equal 1 and et+ 1 should be white noise, independent of whether or not a risk
premium exists. A test of the hypothesis that 0 = 0 is a test of whether the
conditional variance matters in determining the deviation of the forward
rate from the expected future spot rate. Maintaining fl~= 1 and et+l white
noise, to = 0 and 0 = 0 implies a zero risk premium, 8o ~ 0 and 0 = 0 implies a
nonzero constant risk premium, and t0 5 0 and 0 5 0 implies a time-varying
risk premium. Eq. (15) provides a general representation for the conditional
variance, and hence the risk premium: it depends on the previous p lagged
squared forecast errors and a variable zt, that can be specified below.
Consider two implications of eq. (15) for the behavior of the risk premium.
First, according to eq. (13) the only time-series movement of the risk
premium is due to the movement of the conditional variance. Second, the
risk premium can be positive or negative and can switch signs, depending on
the values of 8o and 0. This is important since, for example, Stockman (1978,
p. 172) found that 'the point estimates of the risk premium change signs',
when he estimates a hypothesized constant risk premium over different
subsamples. Suppose, for example, that 8 0 < 0 and 0>0. Then, for small
forecast errors the risk premium will be negative (long positions in forward
foreign currency require an expected loss), while for large forecast errors the
risk premium may turn positive (long positions in forward foreign currency
require an expected profit).

4. The ARCH-in-mean regression model


Consider a generalization of (12)--(15), which we call the ARCH-in-mean
(AIM) regression model, originally introduced by Engle, Lilien and Robins
(1982). For the case in which the conditional variance is described by a pthorder linear lag structure in squared forecast errors and some exogenous
variables, the specification is given by

I. Domowitz and C.S. Hakkio, Conditional variance

54

Ytlx,, I,_l N(x,fl + h,O,h2),

(16)

h t = a o + ~,, ai~2_i+zi_i~b,

(17)

i=I

(18)

~,= Y,- x,fl-h,O,

where xt and zt are i x k and 1 x s vectors of weakly exogenous and possibly


lagged dependent variables. The log-likelihood function is
T

L = T - ' ~ L,,

(19)

t=l

L, = - l o g h , - k # / h L
in which the constant term has been omitted. This likelihood can be
maximized with respect to the unknown parameters /7, 0, a, and ~b.
Maximization is not as straightforward as for the original ARCH model
proposed by Engle (1982). In that model, 0 = 0 , and it can be shown that the
information matrix is block-diagonal between the (a, ~b) and fl parameters.
The information matrix is not block diagonal here, and all parameters must
be estimated simultaneously. Also, if 0 = 0 , the ordinary least squares
estimator of fl is consistent, and if the xt can be treated as fixed constants,
the least squares standard errors are correct. Neither of these statements is
true if 0 # 0.
The maximum likelihood estimator is calculated by solving the first-order
conditions. The derivative with respect to fl is:

r [-x;st 10ht ['~2

aht Ont"]

(20)

The first term in (20) corresponds to the first-order condition for an


exogenous heteroscedasticity correction. The second term arises in the
ARCH model of Engle (1982), because the conditional variance is a function
of the fl's. The last term appears because the ht is now a part of the mean
regression function. The derivatives of ht with respect to all the parameters of
the model have a recursive structure. Expressions for these derivatives are
given in an appendix, available from the authors. The derivatives of the
likelihood function with respect to the other parameters have a similar form.
Gathering terms, they may be written as

OL
~ [-5, 1 Oh, 7
0"-'0= T-t

(21)

I. Domowitz and C.S. Hakkio, Conditional variance

55

OL T- 1 ~ 1 Oht
--=Oa
,=1 h ~ ~"

(22)

OL
r 1 Oht
0---~= T-1 ,= IEh~-~ ~''

(23)

where y,= e2/h2 - 1 + Oedh,.


The information matrix is calculated as minus the expected value of the
conditional expectation of the Hessian matrix. These calculations are also
given in the appendix. Consistent estimates of the information matrix are
obtained by evaluating the relevant expressions at the maximum likelihood
estimates t, 0, a, and t~. Finally, it can be shown that under suitable
regularity conditions: 7

~/-T(~-~)I~N(O,I-I),

as T---,oo,

(24)

where ~ = (fl, 0, a, ~b), ~ = (L 0, 4, q~), and I is the complete information matrix.


A test for the presence of lagged squared errors in both the mean and the
variance of a regression model is given by Engle, Lilien and Robins (1982).
The test can be derived by evaluating (22) and the information matrix under
the null hypothesis, H o : 0 q = ... = % = 0 , and applying the Lagrange multiplier (LM) principle. The test statistic is

LM=2--~7'wlw'w

- ~ 0 2 (w'x(x'x) -lx'w)l-lw'y,

(25)

where w is a matrix with rows wt=(1,e2_l .... ,82_p) and y is a vector .with
elements Yt, evaluated under the null hypothesis. This expression can be
calculated directly using the least squares residuals from an OLS regression
of Y on x. Unfortunately, if x includes a constant term, 0 is unidentified
under the null. This creates an ambiguity in the test procedure, and is
discussed in Davies (1977) and Engle, Lilien and Robins (1982). Any choice
of 0 results in a test of correct asymptotic size. The test may be expected to
have less than optimal power asymptotically. A choice for 0 5 0 is discussed
in the next section. If 0 is chosen to be zero, the test simplifies to the ARCH
test derived in Engle (1982). In that case the test statistic is computed as TR 2
of a regression of 52 on w,. In both cases, the LM statistic is distributed as
X2(p) asymptotically, under the null hypothesis.
VSufticient conditions may be found in Crowder (1976) and Domowitz and White (1982).
Restrictions on h2 are given by Engie (1982, theorem 3).

I. Domowitz and C.S. Hakkio, Conditional variance

56

5. The least squares model


I n this section, O L S e s t i m a t e s for a r e g r e s s i o n of the rate of d e p r e c i a t i o n
o n the f o r w a r d p r e m i u m are p r e s e n t e d . Tests are c o n d u c t e d for the p r e s e n c e
of A R C H i n the e r r o r s a n d A I M effects, as well as for serial c o r r e l a t i o n . O L S
e s t i m a t e s of

St+l-St

Ft--St +

(26)

are g i v e n i n t a b l e 1, w h e r e ( S t + t - S t ) / S t is the rate of d e p r e c i a t i o n a n d


( F t - S t ) / S t is the f o r w a r d p r e m i u m . T h e d a t a are for the p e r i o d J u n e 1973
to S e p t e m b e r 1982; see the d a t a a p p e n d i x (at the e n d of the p a p e r ) for m o r e
detail. T h e e s t i m a t e s of fl0 a n d /fit in (26) are s i m i l a r to those r e p o r t e d b y
o t h e r s [see F r e n k e l (1982) a n d C u m b y a n d O b s t f e l d (1982)]. I n eq. (26) the
e s t i m a t e of fit r a n g e s f r o m - 1.092 to 0 . 3 2 2 )

Table 1
Tests for serial correlation and ARCH OLS estimation: (St+~-S))/S,=flo+

#t[(Ft-S,)/S,] + 8,+~.
Parameter

United
Kingdom

France

Germany

Japan

Switzerland

fie

-0.002
(0.003)

-0.003
(0.003)

0.002
(0.004)

0.001
(0.003)

0.010
(0.006)

fit

- 1.056
(0.741)
0.247
0.212
0.730
0.883

0.322
(0.662)
0.334
0.443
0.298
0.001

0.234
(0.959)
0.578
0.469
0.172
0.004

-0.200
(0.482)
0.006
0.038
0.037
0.002

- 1.092
(0.922)
0.652
0.754
0.521
0.961

Serial-1
Serial-2
ARCH-1
ARCH-2

Notes: Standard errors are in parentheses below each coefficient. The serial
correlation and ARCH test statistics reported are the marginal signiticance level. Serial-1 is an LM test for serial correlation of order 4;
serial-2 is an LM test for serial correlation of order 4 that is robust to
heteroscedasticity; ARCH-1 is an LM test for ARCH errors of order 4,
under the assumption that 0=0. ARCH 2 is an LM test for ARCH
errors of order 4, under the assumption that 0~0.

SFama (1983) defines P,= - R P t and writes In Ff = Et In St+ x+ Pt, Then, in a regression of the
rate of depreciation on the forward premium, he shows that the slope coefficient can be written
as
vat (Er In S)+ i - In St) + coy (P, Et In St + i - In St)

fll

var(Pr)+ var(EtlnSc+1_inS,)+ 2cov(P, EflnSt+1_inSr ).

Therefore, fl~ = I only if the risk premium is constant.

I. Domowitz and C.S. Hakkio, Conditional variance

57

Marginal significance levels for standard LM tests for fourth-order serial


correlation are given in the row marked 'Serial-l' of the table. 9 The
hypothesis of no serial correlation in the errors is rejected only in the Japan
regression, at the 5 percent level. These tests are valid only under the
maintained hypothesis of homoscedastic errors, however. As noted in section
3, there is some evidence that the errors of such regressions are indeed
heteroscedastic. The heteroscedasticity-robust LM test described in
Domowitz and Hakkio (1984) is, therefore, used to test the same null
hypothesis. Marginal significance levels are reported in the row marked
'Serial-2' of table 1. As can be seen, there is no qualitative change in the
results.l 0
The residuals from eq. (26) are also tested for fourth-order A R C H
maintaining 0 = 0 . These results are reported in the row marked ' A R C H - I ' of
table 1. The null hypothesis of no A R C H cannot be rejected, except in the
case of Japan. 11 The latter result may be spurious, however, due to the
presence of serial correlation in the residuals. If (16)-(18) is the correct model,
maintaining 0 = 0 should produce a test with poor power properties. The
score under this maintained hypothesis simply tests for nonzero correlations
between current and lagged residuals. If 0 is actually nonzero, the correlations
between lagged squared residuals and the level of current residuals should be
taken into account as well.
This is accomplished by using the statistic (25) with 0 # 0 . As noted in
the previous section, 0 is not identified under the null, since a constant is
included in the regressions. Following Engle, Lilien and Robins (1982), the
intercept under the null is used to estimate 0 as though there were no
constant in Xr This produces an absolute value for 0 of about 0.1.12 These
results are reported in the row marked 'ARCH-2' of table 1. The results
suggest that lagged squared errors may affect the mean and variance in (26)
for France, Germany, and Japan.

6. The AIM model for exchange rates


In this section the results of estimating eqs. (12)-(15) for all countries are
9Tests for first- through fifth-order serial correlation were also performed, using both the
standard LM test and the heteroscedasticity-robust LM test described below. There was no
change in the results, except for the England regressions, in which the hypothesis of no firstorder serial correlation was rejected in all cases. This result may be due to the omission of a
time-varying risk premium in the model; see section 6.
tAs noted in Domowitz and Hakkio (1984), it is possible to conduct tests for serial
correlation in the presence of heteroscedasticity, but it is not feasible, in general, to develop valid
tests for heteroscedasticity in the presence of serially correlated errors. We present the results for
the LM test for serial correlation, reported in Domowitzand Hakkio (1984).
11In several cases individual coefficients were found to be significantly different from zero in
the test regressions of squared errors on lagged squared errors, however.
t2Subsequent estimates, reported in the next section, indicate that the values of 0 used for (26)
are too low.

58

I. Domowitz and C.S. Hakkio, Conditional variance

reported, excluding any predetermined z, variables from the variance equation. The order of ARCH is set equal to four for all currencies, due to
programming constraints and based on the previous test results. Since the
parameters in the variance equation must satisfy non-negativity conditions,
previous estimates of ARCH models have generally been based on the
assumption that the variance parameters follow a triangular lag distribution
with a single parameter to be estimated and checked for conformity [e.g.
Engle (1982), Engle, Lilien and Robins (1982)]. Instead, we use squares to
impose the non-negativity constraints directly, estimating variance equations
of the form:
4

h,2=e(2+ E ~ e 2 ,

(27)

i=1

The results of estimating eqs. (12)-(15) are given in table 2. Parameter


Table 2

(S,+~-S,)/S,=Po+#,[(F,-S,)/SJ+Oh,+~,+I; h,2=~o+~,~,_,2
~ ~ +"" +~,~L,.
United
Kingdom

France

Germany

Japan

Switzerland

flO

-0.016
(0.002)

0.020
(0.022)

0.040
(0.023)

-0.020
(0.021)

0.010
(0.038)

Pl

- 1.755
(0.884)

0.381
(0.632)

0.054
(1.120)

- 0.211
(0.444)

- 1.092
(1.436)

0.325
(0.665)

-0.708
(0.798)

- 1.120

0.023
(0.003)
0.555
(0.207)

~0

tY 2

eF3

0.723

0.006

0.024
(0.003)

(0.714)
0.027
(0.004)

(0.806)
0.021
(0.004)

(0.010)
0.036
(0.003)

0.378
(0.176)

0.209
(0.213)

0.126
(0.480)

0.211 * 10-~
(0.211 I0 -~)

0.477,10 -~ - 0 . 1 0 9 , 1 0 -7
(0.190)
(0.189)

0.242,10 -7
(0.379)

0.147
(0.225)

0.162, 10-9
(0.162, 10 -9 )

0.103,10 -7
(0.365)

0.476
(0.158)

0.485
(0.144)

0.125, lO-'l
(0.457, 10 -7)

0.458
(0.171)

(0.330)

(0.523)

0.501
(0.184)

0.495
(0.160)

0.636, I0-~
(0.636, 10 -~)

RP1

0.125

0.681

0.159

0.573

0.500

C(4

0.100

0.505* 10 -7

RP2

0.024

0.612

0.275

0.017

0.518

Het

0.358

0.271

0.074

0.036

--

Info

0.742

0.485

0.107

0.810

0.742

Notes:

Standard errors are in parentheses below each coefficient. The numbers


reported for RP1, RP2, Her, and Info are marginal significance levels.
RP1 is an F test of the hypothesis fie=0 and 0=0; RP2 is an F-test of
the hypothesis fie=0, i l l = 1 and 0=0; Het is the residual based test for
remaining heteroscedastieity; Info in the information matrix test of White
(1982) for general model specification.

I. Domowitz and C.S. Hakkio, Conditional variance

59

estimates for the variance equations are reported in terms of the underlying
(unsquared) parameters.
There are several things to note in table 2. The estimates of/~o and ~x are
similar to those reported in table 1. With the exception of Switzerland, at
least one ~ is statistically significant. Finally, the coefficient on ht is estimated
imprecisely; the t-ratios for 0 range from 0.50 (for the United Kingdom) to
1.57 (for Germany).
Before conducting tests of economic hypotheses, checks on the adequacy of
the statistical model are carried out by examining the possibility of remaining heteroscedasticity and serial correlation in the errors and the validity of
the distributional assumption. Results for three test procedures are reported,
including a general heteroscedasticity test, a test for certain omitted variables
in the conditional variance function, and the White (1982) information
matrix text.
A heteroscedasticity test, constructed as a residual-based diagnostic test of
the model, may be carried out by regressing ( ~ - ~ ) / ~
on I/E, x;xt/~, and
x'tflt/~, and testing whether the coefficients of these variables are significantly
different from zero, where /z~ is the conditional variance evaluated at the
maximum likelihood estimates and ~t=Yt:--xtfl-~tO. The form of this test
regression is the same as that used to carry out a White (1980) heteroscedasticity test in a linear model estimated using generalized least squares. It is
not the same test, however, since the AIM model is nonlinear. ~a The idea
behind this specification test is that any heteroscedastic pattern might be
approximated in such a way, and that the approximation should be close
enough to yield reasonable powerJ 4
Marginal significance levels for this test are reported in table 2 in the row
marked 'Her'. The null hypothesis of homoscedasticity is rejected only for
Japan at the 5 percent level, although Germany is a borderline case, with a
marginal significance level of 0.074J 5
Wald tests for the omission of a variable in the conditional variance were
then carried out for Germany and Japan. The models were re-estimated, for
example, using (St-1-Ft-2)/St-2 as the z variable in (17). The coefficient on
this variable was very small in magnitude and was estimated with a large
13See Pagan and Hall (1983) for a discussion of the general principles leading to such a test
statistic.
~4Agaln, see Pagan and Hall (1983). They note that this test is closely related to that
conducted by regressing squared residuals on xt~, which serves as a proxy for E(Yt). The latter
test imposes a weighting pattern involving ~ to reduce the test to one with one degree of
freedom. The test might also be interpreted as a Lagrange multiplier test for remaining
heteroscedasticity after a heternscedasticity correction has been carried out. In that case,
however, the squares and cross-products of g~_~, i= 1. . . . . p, weighted by ~ , would be added to
the test regression ['see Engle (1982)].
15The test could not be successfully calculated for Switzerland. Since the coefficients were so
close to zero, 0 is essentially unidentified, resulting in the singularity of the test regression crossproduct matrix.

60

I. Domowitz and C.S. Hakkio, Conditional variance

standard error in both cases. Thus, the raw lagged forecast errors could not
further explain the conditional variance of the forecast error in the case of
G e r m a n y and Japan, failing to provide an explanation for the results
reported above. 16
The information matrix test of White (1982) is a test of the validity of the
model against any alternative which renders the usual m a x i m u m likelihood
inference techniques invalid. When the model is correctly specified, the
information matrix may be expressed either in Hessian form, -E[t:32L/t3~O~'],
or in the outer product form, E[OL/O~.OL/O~'], where ~ is the vector of
parameters being estimated. The White procedure tests the equality of these
alternative expressions. In the present context, the information matrix test for
normality is sensitive to skewness or kurtosis [White (1982)], and can also
be interpreted as a test for parameter constancy [Chesher (1983)].
The test is based on the 'indicators' OL,/O~i.OL]O~ i + 02 Lt/O~iO~i, i,j = 1. . . . . m,
where m is the total number of parameters in the model. As noted by White
(1982), it is sometimes infeasible to test all indicators jointly. This is the ease
in our model, which contains 36 indicators. We base our test on the elements
of the information matrix associated with the parameters in the mean
regression function, and use 21 indicators.
Computation of the White test can also be cumbersome due to the
presence of third derivatives of the likelihood in the test statistic. This is
especially true here, since such expressions involve recursive relations in first
and second derivatives of the conditional variance. Under the null hypothesis
of information matrix equivalence, however, Chesher (1983) shows that the
test statistic may be computed by calculating T times the raw R 2 from the
regression of a column of ones on a matrix with elements OLJO~, i = 1. . . . . m,
and O2L,/O~Oj + c3LJO~, c3LtO~i, i = 1. . . . , m; j = 1 , . . . , m.17 These expressions
are all evaluated at ~ = (, the m a x i m u m likelihood estimate.
Marginal significance levels for the information matrix test are also
reported in table 2, in the row marked 'Info'. The results are similar to those
obtained using the heteroscedasticity test, with the exception of Japan. The
hypothesis of a correctly specified model cannot be rejected for any country
in the depreciation rate model, although the result for G e r m a n y is still
marginal} s
~Other speitications for z were tried as well, yielding the same results. These included [(St-t F,- 2)/S,-1] 2, logSt-l-logF,_2, and [log S,_ ~-log Ft_ 2]2. It might also be noted that the

remaining parameter estimates were virtually unchanged when these variables were included in
the models.
tVThis is the R2 calculated as the explained sum of squares over the total sum of squares.
laAgain, the information matrix for Switzerland was virtually singular, and the test result in
this case has little significance. The conflict between the heteroscedasticity and information
matrix test results for Japan may be ascribed to finite-sampleproblems. The information test has
less than optimal power, since not all possible indicators are utilized, while the beteroscedasticity
test lacks optimality properties in a maximum likelihood setting.

I. Domowitz and C.S. Hakkio, Conditional variance

61

Thus, eqs. (12)-(15) cannot be obviously rejected by the data. It is


interesting to note that, although serial correlation seemed to be present in
the OLS formulation of the model for Japan, these is no residual serial
correlation when the conditional variance of the errors is modeled by an
ARCH process and a time-varying risk premium is taken into account.
7. Statistical evidence for a risk premium

Several tests for the presence of a risk premium in eq. (26) are reported in
this section. As noted in the discussion following eqs. (12)-(15), the absence
of a risk premium requires t o = 0 , fix=l, 0=0, and et+l to be serially
uncorrelated. Diagnostic testing indicates a lack of serial correlation in et+l.
Maintaining fl~=l, we first test the joint hypothesis t o = 0 and 0=0. The
results in table 2 (in the row marked 'RPI') indicate that this null hypothesis
cannot be rejected for any currency. We next test the expanded hypothesis
t o = 0 , fl~ = i, and 0=0. The results of this test are reported in table 2, in the
row marked 'RP2'. The results indicate that the null hypothesis of no risk
premium can be rejected for the United Kingdom and Japan, and cannot be
rejected for France, Germany, or Switzerland.
In the case of Japan, the heteroscedasticity-robust serial correlation test
reported in table 1 indicates serial correlation in the errors, which also
indicates a rejection of the unbiasedness hypothesis. Although the coefticient
on ht in the yen AIM specification is estimated with a large standard error,
such a serial correlation is not observed in the AIM model residuals. The
results of tables 1 and 2 are, therefore, in agreement with respect to the
rejection of the unbiasedness hypothesis, and suggest the importance of
further investigating a time-varying risk premium for the yen-dollar exchange rate.
It is useful to compare some of the other results in a similar manner. The
unbiasedness hypothesis for the pound is also rejected in the OLS model
based on the criterion fll:~l and on the possibility of first-order serial
correlation in the OLS residuals (see footnote 9 above). Diagnostic tests of
the AIM residuals indicate no serial correlation when ht is included, despite
the large standard error on 0, again suggesting the potential relevance of a
time-varying premium.
The story is somewhat different for Switzerland. The tests of the OLS
model and the AIM results both indicate the absence of ARCH effects and
residual correlation. Based on the OLS model, the unbiasedness hypothesis is
rejected on the grounds that fll ~ 1. The standard error of fit in the AIM
specification is high enough that the hypothesis fix = 1 cannot be rejected.
Under the circumstances, it might be sensible to resolve the conflict on the
basis of the simple OLS model. It must be kept in mind, however, that this
standard error is extremely large in both specifications, and the evidence is
too slim to draw any firm conclusions.

I. Domowitz and C.S. Hakkio, Conditional variance

62

The results in table 2 can be used to estimate and plot the risk premium.
Plots of the estimated risk premium for France and Germany are shown in
figs. 1 and 2; plots of the risk premium for the other three countries and the
plots of the conditional variance are available from the authors. The plots
indicate substantial movement of the conditional variance. In addition, the
estimated risk premium switches from positive to negative. This is consistent
with the results of Stockman (1978, p. 172) and the observation of Frankel
(1982, p. 256) that the risk premium must fluctuate 'frequently between
positive and negative, in order to account for the absence of persistent
unconditional bias'.
8. Conclusion

In this paper, the existence of a risk premium in the foreign exchange


market based on the conditional variance of market forecast errors was
examined. There have been several studies which recognize the presence of
conditional heteroscedasticity in the forecast errors and take it into account
in inference procedures. We model the conditional variance directly, based on
the ARCH specification of Engle (1982), using maximum likelihood techniques to obtain efficient estimates.
Our estimates provide some evidence of a nonzero constant risk premium
for some, but not all, currencies. The null hypothesis of no risk premium
(fl0=0, fll = 1, and 0=0) can be rejected for the United Kingdom and Japan,
but cannot be obviously rejected for Germany, France, and Switzerland.
While there is evidence against the unbiasedness hypothesis for a majority of
the currencies, there is little support for the conditional variance of the
exchange rate forecast error being an important sole determinant of the risk
premium. However, estimates of the risk premium itself can be obtained,
given our specification, and it is interesting to note that the premium is
positive or negative for different periods, in accord with the results of
Stockman (1978).
Similar problems have been encountered in other empirical studies of a
time-varying risk premium. For example, Hansen and Hodrick (1983) and
Hodrick and Srivastava (1984) consider a model in which the risk premiums
on various forward contracts are linearly related to the expected return on a
benchmark portfolio, and are proportional to each other. 19 Hodrick and
19In particular, they estimate

s~

#,leE,(R,b +1 - R{+ ~),

where a superscript i denotes currency i, Rb+l is the nominal return on a benchmark security
and flf~,_
l i Rtb+ l]/vart (Rt+
a 1) is assumed constant. Et(Rb+ 1 - R{+ i) is treated as
- cov~ [(Sfi + I - Ft)/S~,
an unobserved variable; the instruments used by H a n s e n and Hodrick (1983) were the vector
(S~-F~_ 1)/S~f_l, the instruments used by Hodrick and Srivastava (1983) were the vector (F~-S~)/S~.

63

I. Domowitz and C.S. Hakkio, Conditional variance

J u n e 1973 t o A u g u s t 1982

Percent
.24

I l- -w ii

,/,V ..I

- .24

- .48

vl

1
- .72

.96

1.20 - -

1.44

'74

1973

I
'75

'76

'77

'78

'79

'80

'81

'82

Fig. I. French risk premium.


J u n e 1973 to A u g u s t 1982

Percent
1.50
1.00
.50 - -

v~/

0
-

.50

II'

- 1.00

- 1.50

- 2.00
2.50 - -

3.00
1973

I
'74

I
'75

I
'76

I
'77

I
'78

I
'79

Fig. 2. German risk premium.

I
'80

'81

'82

64

I. Domowitz and C.S. Hakkio, Conditional variance

Srivastava (1984), reproducing the results of Hansen and Hodrick (1983), are
unable to reject the model of the risk premium for the period February 1976
to December 1980. However, when they extend the sample period to
September 1982 they reject their empirical model of the risk premium. They
state that either the model's coefficients are not constant or a different model
of risk and return is required to describe the risk premium. In this study, we
use the same data as do Hodrick and Srivastava. 2 While our results are not
directly comparable due to a different model of the risk premium, 21 the
results of section 7 suggest that time variation in the coefficients is not
necessarily the problem.
Frankel (1982) considers an international portfolio balance model based
on maximizing end-of-period wealth. Taking this model as a maintained
hypothesis, he finds that he cannot reject the hypothesis of risk neutral
agents. He argues that since he incorporates the restrictions of m e a n variance optimization, his test has greater power than do the tests of
Hansen and Hodrick (1983) and C u m b y and Obstfeld (1982). However, he
notes that several simplifying assumptions have been made [Frankel (1982, p.
260)].
In most theories of asset pricing, the risk premium depends on the
covariance matrix of exchange rate changes [Frankel (1982)1 or on the
eovariance between marginal rates of substitution [Hodriek and Srivastava
(1984)]. In either case, time variation in the premium might be expected, but
it is an entire variance-eovariance structure which must be modeled in this
way, and a multi-equation framework is called for. If one accepts the first
explanation, a generalization of the present approach to a multi-currency
model might be appropriate. In that case, the risk premium is specified as a
function of the conditional covariance matrix for all currencies. It is certainly
possible that the covariance terms m a y have more explanatory power.
An alternative approach is to consider a general model in the spirit of that
presented in section 2 of this paper. In that ease, the result that the risk
premium depends on the conditional second moments of the exogenous
2Hodrick and Srivastava use only the last 78 observations for most of their analysis. Their
tests of parameter stability indicate the possibility that a structural break occurs after the first 30
observations, for three of the five currencies. We observe what appears to be a shift in the
residual variance after these first 30 observations, but cannot reject parameter constancy based
on the tests here. Since parameter stability tests are closely linked to heteroscedasticity tests
['Pagan and Hall (1983)1, and the tests are only asymptoticallyjustified, the resolution of this
conflict appears to require further data.
21They treat the risk premium as linearly related to the vector of current forward premiums,
while we treat it as related to the conditional variance of the forecast error 6, = Yt + ~ - f i e - O h , - f l l x t
[eq. (12)]. Also, they relate the conditional variance to the vector of current forward premiums
and the squared forward premiums (but they do not use this information in estimation), while
we use squared forecast errors, ~, and a variable z, [eq. (15)]. The error term in the Hoddck
and Srivastava paper is conditionally and unconditionally normal, while ~t in our model is conditionally normal but not unconditionally normal.

I. Domowitz and C.S. Hakkio, Conditional variance

65

variables would be exploited directly. F o r each currency an exchange rate


equation, m o n e y supply equations, and real income equations would be
specified and estimated jointly, taking the form of the conditional heteroscedasticity into account.

Data appendix
The data used in this paper were supplied to us by Robert Hodrick, and
are described in H a n s e n and H o d r i c k (1984) and H o d r i c k and Srivastava
(1984). The data are averages of b i d - a s k rates taken from Bank of America
by DRI. There are 108 nonoverlapping observations, from 5 June 1973 to 17
August 1982. The data are 'approximately' monthly. The ' a p p r o x i m a t i o n '
arises for the following reason. F o r each date there are three variables: the
spot rate (St), the forward rate (Ft), and the future spot rate (FSt). The
forward rate (Ft) 'predicts' the future spot rate (FSt). The timing is that at
time t, St and F t are observed on Tuesday, and FSt is the spot rate observed
four weeks and two business days later, on a Thursday. Then, at time t-I-1,
St+l and F t + l are observed on Friday and FSt+~ is the spot rate observed
four weeks and one business day later, on ~a M o n d a y . [Notice, this forward
rate is not matched with the appropriate future spot rate; see H o d r i c k and
Srivastava (1984, footnote 13) and H a k k i o (1983).] We then repeat this cycle.
It should be noted that because of this dating F S t ~ S t + ~.

References
Chesher, A., 1983, The information matrix test: Simplified calculation via a score test
interpretation, Economic Letters 13, 45-48.
Crowder, M.J., 1976, Maximum likelihood estimation for dependent observations, Journal of the
Royal Statistical Society, Series B, 38, 45-53.
Cumby, R.E. and M. Obsffeld, 1982, International interest rate and price level linkages under
flexible exchange rates: A review of recent evidence, NBER Working Paper No. 291,
forthcoming in J.F.O. Bilson and R.C. Marston, eds., Exchange rates: Theory and practice
(University of Chicago Press, Chicago).
Davies, R.B., 1977, Hypothesis testing when a nuisance parameter is present only under the
alternative, Biometrika 64, 247-254.
Domowitz, I. and C. Hakkio, 1984, Testing for serial correlation and common factor dynamics
in the presence of heteroscedasticity, Northwestern University, revised.
Domowitz, I. and H. White, 1982, Misspecified models with dependent observations, Journal of
Econometrics 20, 35-58.
Dornbusch, R., 1982, Exchange risk and the macroeconomics of exchange rate determination, in:
R. Hawkins, R. Levich and C. Wihlborg, eds., The internationalization of financial markets
and national economic policy (JAI Press, Greenwich, Connecticut).
Engle, R.F., 1982, Autoregressive conditional heteroscedasticity with estimates of the variance of
United Kingdom inflation, Econometrica 50, 987-1007.
Engle, R.F., D.M. Lilien and R.P. Robins, 1982, Uncertainty and the term structure, U.C.S.D.
Discussion Paper 82-4.
Fama, E., 1983, Forward and spot exchange rates, manuscript, University of Chicago.
Frankel, J.A., 1979, The diversifiability of exchange risk, Journal of International Economics 9,
379-393.

66

I. Domowitz and C.S. Hakkio, Conditional variance

Frankel, J.A., 1982, In search of the exchange risk premium: A six-currency test assuming mean
variance optimization, Journal of International Money and Finance I, 255-274.
Frenkel, J.A., 1978, A monetary approach to the exchange rate: doctrinal aspects and empirical
evidence, in: J.A. Frenkel and H.G. Johnson, eds., The economics of exchange rates
(Addison-Wesley, Reading, Massachusetts).
Frenkel, J.A., 1982, Interest rates, exchange rates and the emerging strength of the U.S. dollar,
University of Chicago.
Frenkel, J.A. and A. Razin, 1982, Stochastic prices and tests of efficiency of foreign exchange
markets, Economic Letters 6, 165-170.
Friedman, D. and S. Vandersteel, 1982, Short-run fluctuations in foreign exchange rates:
evidence from the data 1973-1979, Journal of International Economics 13, 171-186.
Hakkio, C.S., 1983, Discussion of risk averse speculation in the foward foreign exchange market:
An econometric analysis of linear models, in: J.A. Frenkel, ed., Exchange rates and
international macroeconomics (University of Chicago Press, Chicago).
Hansen, L.P. and R.J. Hodrick, 1983, Risk averse speculation in the foward foreign exchange
market: An econometric analysis of linear models, in: J.A. Frenkel, ed., Exchange rates and
international macroeconomics (University of Chicago Press, Chicago).
Hodrick, R.J. and S. Srivastava, 1984, An investigation of risk and return in forward foreign
exchange, Journal of International Money and Finance 3, 5-30.
Levich, R.M., 1983, Empirical studies of exchange rates: Price behavior, rate determination and
market efficiency, NBER Working Paper No. II12.
Lucas, R.E., 1982, Interest rates and currency prices in a two-country world, Journal of
Monetary Economics I0, 335-360.
Mussa, M., 1979, Empirical regularities in the behavior of exchange rates and theories of the
foreign exchange market, Carnegie-Rochester Conference Series on Public Policy I l, 9-57.
Pagan, A.R. and A.D. Hall, 1983, Diagnostic tests as residual analysis, Australian National
University.
Stockman, A.C., 1978, Risk, information, and forward exchange rates, in: J.A. Frenkel and H.G.
Johnson, eds., The economics of exchange rates (Addison-Wesley, Reading, Massachusetts).
White, H., 1980, A heteroscedasticity-consistent covariance matrix estimator and a direct test for
heteroscedasticity, Econometrica 48, 817-838.
White, H., 1982, Maximum likelihood estimation of misspecifled models, Econometica 50, 1-25.

Potrebbero piacerti anche