Journal of International Money and Finance 1994 13(3) 375-383

Spread and volatility in spot and

forward exchange rates

TAE-HWY LEE*
Department of Economics, Louisiana State University,
Baton Rouge LA 70803, USA

This paper is concerned with modeling the conditional heteroscedasticity

of the prediction error of foreign exchange rates. As spot and forward rates
are cointegrated we use a system of error correction models for mean
prediction. To predict the variance we use a bivariate generalized auto-
regressive conditional heteroscedasticity (GARCH) model as a function of
the spread. Using daily series for seven currencies, we find that unmodeled
conditional heteroscedasticity by GARCH can generally be explained by
the squared spread. This indicates that as the spread is bigger the exchange
rates are more volatile. (JEL F31, C32).

The characterizations of exchange rate movements have important implications

for many issues in international finance and macroeconomics. It is therefore
important to carefully model any temporal variations in the volatility process as
well as in the process of change in the exchange rate series. A standard way to
specify volatility is autoregressive conditional heteroscedasticity (ARCH) due to
Engle (1982) and generalized ARCH (GARCH) of Bollerslev (1986). In this paper,
we consider an extension of the GARCH model for the error correction models
(ECM) of cointegrated series. Cointegration links long-run components of a group
of integrated series. It can be used to discuss some types of equilibrium and to
introduce those equilibria into time series models. The dynamic models for the
short-run behavior of cointegrated series are called error correction models.
An ECM may be considered as a prediction equation of the cointegrated series
as it is a conditional expectation of the first difference of the series given an
information set. Engle and Yoo (1987) show that the error correction term, which
is the short-run deviation from a long-run cointegrated relationship, has important
predictive power for conditional mean of the cointegrated series. If the error
correction term is an important variable in the conditional mean, so may it be
in the conditional variance. This may imply that if the series deviate further from
each other they are harder to predict. If disequilibrium (measured by error
correction term) is responsible for uncertainty (measured by conditional variance),

*I would like to thank Richard Baillie for providing the data used in this paper, two anonymous
referees, Ji-Chai Lin, Doug McMillin, Ted Palivos and Hector Zapata for valuable comments, and LSU
College of Business Administration for financial support.

0261-5606/94/03/0375-009 0 1994 Butterworth-Heinemann Ltd

 316 Spread and volatility in spot and forward exchange rates

conditional heteroscedasticity may be modeled with a function of several lagged

error correction terms. Examining the behavior of the variances over time as a
function of short-run deviations is reasonable when one expects increased volatility
due to shocks to the system which propagate on both the first and the second
moments. We thus use a system of ECMs for the conditional mean, and an
extended bivariate GARCH model with error correction term for the conditional
variance. The main virtue of the model lies in its capability of pointing to a
particular feature ofcointegrated series, which is the potential relationship between
disequilibrium and uncertainty. The model seems appropriate for testing for
causality in variance through the error correction term.
Using this model we investigate the predictive power of the spread between
daily spot and forward exchange rates for seven currencies in predicting
volatility of the exchange rate changes. We find that unmodeled conditional
heteroscedasticity by the GARCH(l, 1) model can be explained by a function of
the spread. The conditional variance of the prediction error is positively related
with the squared spread. This indicates that as the spread is bigger the exchange
rates are more volatile and harder to predict. The short-run error from the
cointegrating long-run relationship is therefore a useful variable in modeling
conditional variance as well as conditional mean. This may be exploited to obtain
more precise time varying confidence intervals for point forecasts of exchange
rate changes.

I. The model
Much empirical research on the risk premium in forward foreign exchange relies
on intertemporal asset pricing models such as that of Lucas (1978). An important
relationship in an international environment that can be derived from there is
the conditional pricing relation for a forward premium: &Q,+&z,+~ =O, where
E, is the conditional expectation given an information available at time t,
Q t+k=(U’(Ct+k)/U’(Ct))(Pt/P,+k) is th e marginal rate of substitution between
dates t and t + k, U’(C,) is the marginal utility of real per capita consumption C,,
P, is a price index in the domestic economy at time t, z~+~=(S/+~-F/,~) is the
forward premium, Sj is the spot exchange rate between the domestic currency
j at time t, F:,k is the forward exchange rate which is the domestic currency price
of a unit of currency j established at t for payment at t + k. Then we have
&Q, + k&Z, + k + p,(Q, + k, 2, + k) a,@r + k) C’,h + k) = 0, where PC is the conditional
correlation and or is the conditional standard deviation. The premium on forward
foreign exchange is seen to be related to the conditional second moments.’
Although asset pricing theories relates first moments to second moments, most
empirical studies (e.g. Giovannini and Jorion, 1989; Kaminsky and Peruga, 1990;
McCurdy and Morgan, 1991; Attanasio, 1991) have focused on a specification
in which first moments are explained by second moments, especially using
GARCH in mean (GARCH-M) model of Engle et al. (1987). Exceptions are
Cumby and Obstfeld (1984) and Hodrick (1989), who consider the converse. While
the GARCH-M specification is useful to explain excess return in terms of
conditional second moments, examining the converse is also interesting due to
the reason discussed before especially when the system is cointegrated. The
purpose of the paper is to examine this converse specification using a multivariate
GARCH specification in vector error correction models. Error correction terms
 TAE-HWYLEE 311

often turn out to be the most important terms for conditional means, especially
in the long horizons. Since most finance theory links conditional means to
conditional second moments, it is naturally interesting to see if the most important
variables for conditional means do in fact affect conditional variances. This allows
us to examine the potential relationship between disequilibrium and uncertainty
in the cointegrated system.
Now we turn to our econometric model. Consider a 2 x 1 vector of cointegrated
series X,. They may be considered to be generated by the following ECM

(1) AX,=~+~Z,_~+I-~ AX,_l+...+I-pAXt_-p+~t,

where A = 1 -B, B is the backshift operator, Z,E cr’X, is the magnitude of the
short-run deviation at time t, and p, y, c1are all 2 x 1 vectors of parameters and
Is are 2 x 2 matrices of parameters. c1is the cointegrating vector which determines
the long-run equilibrium relationship and y is the parameter measuring the speed
of error corrections. The error vector E,= (ei, ezt)’ is assumed to follow a bivariate
conditional normal distribution with mean zero and conditional variance-
covariance matrix H,rE(&) Ft__& where &_ 1 is the a-field generated by all
information available at time t- 1.
If the ECM (1) is correctly specified for the conditional mean E(AX, I Ft- I),
then E(E, 1Ft_ ,)=O with probability one by construction. The error E, is a
martingale difference process but is not necessarily independent. The squared
error may be predictable using the information set, and thus the conditional
variance may be a function of the variables in the information set. H, can be
written as a measurable function of the variables in gt _ 1. We may consider a
function of AX,_, such as selective elements of AX,_, AX; --s for some s > 1.
However, the short-run disequilibrium error z,_i may seem an especially
interesting one. We consider the squared z,__1 in a bivariate GARCH( 1, 1) model
suggested by Baba et al. (1991),

(2) H,=C’C-~A’E,_~E~_~A+B’H~_~B+D’DZ:_,,
where A, B, C and D are 2 x 2 matrices, C is an upper triangular matrix with cl,
c:! in its diagonal and c3 off the diagonal, and D has the same form as C with
corresponding elements dl, d2 and d3. The model will be referred to as ‘GARCH-X’.
The estimates of di may quantify the extent that this model explains the
relationship between disequilibrium and conditional volatility not explained by
GARCH(l, 1). This parameterization guarantees H, to be positive definite and it
allows the conditional covariances to change signs over time.2 In the empirical
study in the next section we suppose that A and B are diagonal matrices with
al, a2 and bi, b2 in the diagonals, respectively.

II. Empirical results

In the following analysis we use daily spot and 30-day forward exchange rate
data from the New York Foreign Exchange Market. The data begin on March
1,198O and end on January 28,1985, which constitute a total of 1245 observations.
The data are those used in Baillie and Bollerslev (1989a, 1989b). They are opening
bid prices for British Pound (BP), German Deutche mark (DM), Japanese Yen
(JY), Canadian Dollar (CD), French Franc (FF), Italian Lira (IL) and the Swiss
Franc (SF). Apart from the BP all the series are in terms of the number of US
378 Spread and volatility in spot and forward exchange rates

Dollars for one unit of foreign currency. The notation s, and ft is used for the
logarithms of the spot and forward series, respectively, so that X,=(s,f$.
Baillie and Bollerslev (1989a) present the results that the unit root hypothesis
cannot be rejected for all the series. They also report the results of the cointegration
tests of Phillips (1987) between s, + 22and ft, which are found strongly cointegrated.
However, if X, =(s t+22 ft)‘, the ECM (1) includes variables not available at time
t. As we are interested in prediction using the conditional expectations given
information at time t, we instead use X, =(st ft)’ so that the right-hand side of
equation (1) should not include the variables unknown at time t. s, and ft are
found strongly cointegrated in all currencies at less than one per cent level. We
use z, =ft -s,, which is found stationary. To save space the estimated statistics
are not reported.
The lag lengths (p) in ECM (l), chosen by the use of the Schwarz information
criterion (SIC), are 4 (BP), 4 (DM), 5 (JY), 5 (CD), 0 (FF), 0 (IL) and 4 (SF).
In Table 1, the asymptotic probability values for various specification tests for
the ECM (1) estimated by least squares assuming conditional homoscedasticity
are presented. These tests are: (a) the Ljung-Box test and the McLeod-Li (1983)
test for up to the twentieth order serial correlation in the residuals and in the
squared residuals; (b) Wooldridge’s (1990) robust regression based LM test for
autocorrelations (AR);(c) the LM test for ARCH; (d) White’s (1989) neural network
test for neglected non-linearity in conditional mean; and (e) the LM test for
GARCH-X. Using the lag lengths (p) in ECM (1) chosen by the SIC, the residuals
are not serially correlated, while the McLeod-Li statistics and the LM test
statistics for ARCH are very significant.
We estimate the model with formulations of the conditional mean E(AX, I%- ,)
as in (1) and the conditional variance H,_=E(& ( &__ 1) as in (2). Let 8 be the
vector of all the parameters in the conditional mean and conditional variance.
The parameter estimates are obtained by maximizing the quasi (normal) log-
likelihood function over 8 using scoring methods with only first numerical
derivatives being used (ci la Bollerslev and Wooldridge, 1992). At the maximizing
value of 8 (the QMLE), the asymptotic robust standard errors are obtained (d la
White, 1982; White and Domowitz, 1984; Weiss, 1986; Bollerslev and Wooldridge,
1992). In Table 2 the estimated parameters of the conditional variances are
reported for five currencies. To save space the results for the conditional means
are not reported. The asymptotic robust standard errors are in parentheses. BP
and JY are not included as we cannot find a step of increasing likelihood from
many different initial values of 8. The estimates of D are generally significant.
Nelson (1990) and Bougerol and Picard (1992) establish conditions for the
stationarity and ergodicity of the (integrated) GARCH(l, 1) process. Bollerslev
and Wooldridge (1992) provide regularity conditions under which the QMLE
will be consistent and asymptotically normal, some of which are verified by
Lumsdaine (1991). She also proves that the QMLE is consistent and asymptotically
normal without assuming a finite fourth moment of errors (which is assumed in
Weiss, 1986). Lee and Hansen (1992) show that these may hold when the
GARCH(l, 1) process is integrated or even when it is mildly explosive provided
that the conditional fourth moment of standardized error is bounded (which is a
fairly weak condition).
Based on the asymptotic normality of the QMLE we test the GARCH-X using
Lagrange multiplier (LM) test, Wald tests, and likelihood ratio (LR) tests.3 Our
 TABLE 1. Specification tests for the ECM (1) estimated by OLS.

DM CD FF IL SF

As, Af, A% Aft A‘% ASr A& Aft A& Aft

Ljung-Box(20) 0.660 0.823 0.108 0.110 0.242 0.109 0.091 0.186 0.351 0.342
McLeod-Li(20) 0.000 0.000 0.000 0.000 0.085 0.000 0.000 0.000 0.000 0.000
Neural(3) 0.139 0.321 0.020 0.273 0.080 0.007 0.483 0.171 0.290 0.341
AR(l) 0.992 0.677 0.464 0.280 0.037 0.040 0.075 0.053 0.848 0.530 _1
AR(2) 0.953 0.970 0.504 0.572 0.999 0.987 0.993 0.865 0.955 0.991 $
AR(5) 1.000 1.000 1.000 0.999 1.000 1.000 1.000 1.000 0.889 0.874 2
ARCH( 1) 0.003 0.003 0.000 0.001 0.161 0.026 0.000 0.000 0.09 1 0.163 ;
ARCH( 5) 0.000 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.000 0.000 w
ARCH( 10) 0.000 0.000 0.000 0.000 0.013 0.000 0.000 0.000 0.000 0.000
ARCH[O]-X 0.000 0.000 0.000 0.000 0.053 0.000 0.000 0.000 0.000 0.000
ARCHLI J-X 0.025 0.027 0.001 0.003 0.143 0.017 0.239 0.929 0.013 0.013
ARCH[S]-X 0.128 0.133 0.016 0.028 0.160 0.057 0.315 0.852 0.088 0.090
ARCH[lO]-X 0.106 0.106 0.049 0.071 0.159 0.077 0.379 0.947 0.101 0.108
GARCH[l,l]-X 0.038 0.005 0.004 0.552 0.985 0.027 0.020 0.163 0.003 0.012

Nofe: The number in parentheses is the degree of freedom. The number in square brackets is the order of GARCH fitted under the null hypothesis to test for GARCH-X.
All the values are asymptotic p-values. For the neural network test we use 10 phantom hidden units, 3 principal components of them, and 5 draws of the test in
computing the Hochberg Bonferroni based (see Lee et al., 1993). We do not report results for BP and JY as we could not find a step of increasing likelihood for
the ECM model with GARCH-X specification from many different initial values of f3.
380 Spread and volatility in spot and forward exchange rates

TABLE2. Quasi-maximum likelihood estimates of GARCH-X parameters.

DM CD FF IL SF

Cl 0.003 (3e-4) 5e-4 (5e-5) 0.003 (4e-4) 0.005 (2e-4) 0.001 (6e-4)
c2 9e-5 (6e-5) 4e-5 (le-5) < le-5 (0.7 15) 6e-4 (le-4) 6e-5 (8e-5)
c3 0.003 (3e-4) 5e-4 (5e-5) 0.003 (4e-4) 0.005 (2e-4) 0.001 (6e-4)
a, 0.627 (0.075) 0.441 (0.016) 0.516 (0.039) 0.409 (0.029) 0.483 (0.053)
a2 0.607 (0.072) 0.447 (0.017) 0.547 (0.044) 0.424 (0.028) 0.473 (0.052)
b, 0.758 (0.013) 0.870 (0.004) 0.782 (0.017) 0.635 (0.015) 0.880 (0.011)
bz 0.760 (0.011) 0.870 (0.005) 0.767 (0.019) 0.602 (0.025) 0.882 (0.012)
dl 0.444 (0.077) 0.312 (0.039) 0.206 (0.096) 0.076 (0.067) 0.195 (0.079)
d2 0.030 (0.009) 0.023 (0.005) 0.107 (0.036) 0.092 (0.016) 0.018 (0.005)
d3 0.450 (0.069) 0.301 (0.040) 0.136 (0.099) 0.076 (0.080) 0.202 (0.076)

Nope: Asymptoticrobust standard errors are in parentheses. For BP and JY we could not find a step of
increasing likelihood using many different parameter values.

LM test (Table 1) is based on a consequence of the null hypothesis. Let

h,,rE(ei,eit ) Ft_-1). When hiit is estimated under the null hypothesis that the
conditional variance is correctly specified, it should be that E(e$ - hiit ) Pj__ I) = 0,
i= 1,2 with probability one if the null hypothesis is true. Consequently, (es-h,)
is uncorrelated with any measurable function of the variables in Ft_i, We
construct an LM test statistic for the null hypothesis that (ez - hii,) is uncorrelated
with zf-i. From the standard asymptotic arguments the TR2 statistic has the
x2(1) distribution asymptotically when the null hypothesis is true, where T is the
number of observations and R2 is the uncentered squared multiple correlation
from a standard linear regression of (efl-hii,) on z:-i). We consider the null
hypotheses that hiit is ARCH(q), q =0, 1, 5, 10. The test with q =0 provides
information on the correlation structure between ei: and zf-i. We also consider
the null hypothesis that hiit is GARCH(l, 1). The test generally indicates that
zf- 1 may improve the prediction of X, through the conditional second moments.
The GARCH-X effect is also tested by Wald tests (Table 2) and by the LR
tests (Table 3) for five currencies. Both tests suggest the presence of the GARCH-X
effect for all five currencies. The model selection criteria (Table 3) also indicate
that the model with GARCH(l, 1)-X has a better fit than the model with
GARCH( 1, 1).
These tests may have power in some other directions and other misspecification
can possibly be detected. If the results indicate the presence of the GARCH-X
effect, they might not provide definitive evidence of neglected GARCH-X effect.
The possible presence of higher order GARCH effects or the potential effects of
other neglected variables in the information set may lead to the observed results.
Much of the model selection literature in which model choice is based on the
SIC, the Akaike information criterion (AJC), or other criteria, is concerned with
selecting parsimoniously undominated models usually in terms of likelihood and
subject to a restriction of correct specification of a model in that the residual
processes from the (first and higher) conditional moments be martingale difference
sequences. If a model is correctly specified in conditional mean and conditional
 TAE-HWY LEE 381

TABLE 3. Log-likelihood, the LR statistics and model selection criteria.

Model Log-likelihood LR statistic AIC SIC

BP Model 1 11222.65 42.85 - 22399.30 - 22281.38

Model 2 11244.08 -22434.15 - 22295.73
DM Model 1 11794.28 464.13 - 23542.56 - 23424.65
Model 2 12026.35 21.74 - 23998:69 - 23860.27
Model 3 12037.22 - 24014.43* - 23860.63*
JY Model 1 10349.67 175.71 - 20645.34 - 20506.92
Model 2 10437.53 -20813.05 -20654.12
CD Model 1 13820.89 754.03 - 27587.79 - 27449.36
Model 2 14197.91 55.22 - 28333.81 -28174.88
Model 3 14225.52 - 28383.04* - 28208.72*
FF Model 1 10020.61 2133.12 - 20027.21 - 19991.33
Model 2 11087.17 267.09 - 22152.33 - 22095.93
Model 3 11220.71 -22413.42* -22341.64*
IL Model 1 10750.58 418.65 -21487.16 -21451.27
Model 2 10959.91 80.27 -21897.81 -21841.41
Model 3 11000.04 - 21972.08* -21900.30*
SF Model 1 11187.19 1063.21 - 22328.38 - 22210.46
Model 2 11718.79 22.32 - 23383.59 -23245.16
Model 3 11729.96 - 23399.91* -23246.11*

Note: * denotes the minimum of AIC or SIC. Model 1 is the ECM estimated by least squares assuming
the conditional homoscedasticity; Model 2 is the ECM with GARCH specification; and Model 3
is the ECM with GARCH-X specification. Model 2 and Model 3 are estimated by maximizing the
normal likelihood function (QMLE) using the method of scoring. The number of parameters in
Model 1 is 4(p+ l)+ 3 where p is the number of lags in ECM (1). Model 2 has four more parameters
than Model 1 and Model 3 has three more parameters than Model 2. For BP and JY we could
not find a step ofincreasinglikelihood for Model 3 from many different initial values ofparameters.

variance, then we may use the GARCH-X specification as a useful description

of the salient aspects of the chosen phenomena.
In order to test the validity of the model, a series of specification tests for the
model standardized by hli2 were conducted (but not reported), where hiir is
estimated by GARCH(l, 1)-X. The McLeod-Li tests and LM tests for ARCH
become less significant, suggesting that some conditional heteroscedasticity is
captured by the model. However, as is well recognized in Baillie and Bollerslev
(1989b) and Hsieh (1989) among many others, daily exchange rate series show a
considerable amount of leptokurtosis even after accounting for GARCH( 1, 1). All
of the standardized residuals have mean close to zero and variance close to unity.
The LM tests indicate no serial correlation in the standardized residuals. The
skewness and the excess kurtosis of the standardized residuals become generally
smaller in modulus. From the diagnostics, we think the model using bivariate
GARCH(l, 1)-X in the system of the error correction models are reasonably well
specified.
We have seen that the conditional variances of prediction errors of the exchange
rate changes are positively related to the short-run deviation from the long-run
relationship. This means that when the spread is bigger, the exchange rates are
382 Spread and volatility in spot and forward exchange rates .

more volatile and uncertainty increases. If $_I has additional predictive power
for the changing variances of the spot and forward exchange rate changes, this
may be exploited to obtain more precise time varying confidence intervals for
point forecasts of exchange rate changes.

III. Concluding remarks

In this paper we investigate a model which seems useful in examining how the
short-run disequilibrium has an effect on uncertainty in predicting cointegrated
series. Examining the behavior of the variances over time as a function of
disequilibrium is reasonable when one expects increased volatility due to shocks
to the system which propagate on first and second moments. The model is thus
appropriate for testing for causality in variance as well as in mean through the
error correction term.
It seems that this specification would ensure a better fit and would be useful
empirically in other possible examples in economics. We have conducted the
same analysis using monthly short- and long-term interest rate series as well as
monthly interest rate series of the commercial paper and the Treasury bill. The
results indicate stront GARCH-X for those data. Hence the’model seems useful
to study the relationship between the short-run deviation from a long-run
relationship (disequilibrium) and uncertainty.

Notes
1. See Hansen and Hodrick (1983) or Mark (1985) for more details. They also derive more
explicit relationship between the first and second conditional moments under additional
assumptions on the joint distribution of Q and z and consumer preferences.
2. In order to allow z:_ 1 to have negative coefficients in (2) we also used a symmetric matrix
G instead of D’D.As the diagonal elements of G turn out to be positive for our data, we
use D’D which guarantees If, to be positive definite. The off-diagonal element of D’D may
be negative.
3. It may be noted that our LM and LR tests are not robust to departures from normality
or information equality while our Wald tests are. Robust LM tests are available in literature
(Wooldridge, 1990; Bollerslev and Wooldridge, 1992), which we do not pursue here. LR
tests do not follow an asymptotic chi-square distribution in the presence of ARCH (White,
1984, p. 76).

References

ATTANASIO,ORAZIO P., ‘Risk, Time Varying Second Moments and Market Efficiency,’ Review
of Economic Studies, May 1991,58: 479-494.
BABA,YOSHIHISA,ROBERTF. ENGLE, DENNIS F. KRAFT, AND KENNETH F. KRONER, ‘Multivariate
Simultaneous Generalized ARCH,’ UCSD Working Paper, 1991.
BAILLIE,RICHARDT., AND TIM BOLLERSLEV, ‘Common Stochastic Trends in a System of Exchange
Rates,’ The Journal of Finance, March 1984, 44: 167-181 (1989a).
BAILLIE,RICHARDT., AND TIM BOLLERSLEV, ‘The Message in Daily Exchange Rates: A Conditional
Variance Tale,‘Journal of Business and Economic Statistics, July 1989,7: 297-305 (1989b).
BOLLERSLEV,TIM, ‘Generalized Autoregressive Conditional Heteroskedasticity,’ Journal of
Econometrics, April 1986, 31: 307-327.
BOLLERSLEV,TIM, AND JEFFREYM. WOOLDRIDGE‘Quasi-Maximum Likelihood Estimation and
Inference in Dynamic Models with Time Varying Covariances.’ Econometric Reviews,
1992, ll(2): 143-172.
 TAE-HWY LEE 383

BOUGEROL,PHILIPPE, AND NICO PICARD, ‘Stationarity of GARCH Processes and of Some

Nonnegative Time Series,’ Journal of Econometrics, April/May 1992, 52: 115-127.
CUMSY,ROBERTE., AND MAURICEOBSTFELD, ‘International Interest Rate and Price Level Linkages
under Flexible Exchange Rates: A Review of Recent Evidence,’ in John F. 0. Bilson and
Richard C. Marston, eds, Exchange Rate Theory and Practice, Chicago: Univeristy of
Chicago Press, 1984.
ENGLE, ROBERT F., ‘Autoregressive Conditional Heteroskedasticity with Estimates of the
Variance of United Kingdom Inflation,’ Econometrica, July 1982, 50: 987-1007.
ENGLE, ROBERT F., DAVID M. LILIEN, AND RUSSELLP. ROBINS, ‘Estimating Time Varying Risk
Premia in the Term Structure: The ARCH-M Model,’ Econometrica, March 1987, 55:
391-407.
ENGLE ROBERT F., AND BYUNGSAMYoo, ‘Forecasting and Testing in Cointegrated System,’
Journal of Econometrics, May 1987, 35: 143-159.
GIOVANNINI,ALBERTO,AND PHILIPPEJORION,‘The Time Variation of Risk and Return in Foreign
Exchange and Stock Markets,’ Journal of Finance, June 1989,44: 307-325.
HANSEN LAS PETER, AND ROBERT J. HODRICK, ‘Risk Averse Speculation in the Forward
Foreign Exchange Market,’ in Jacob A. Frenkel, ed, Exchange Rates and International
Macroeconomics, Chicago: University of Chicago Press, 1983.
HODRICK,ROBERTJ., ‘Risk Uncertainty and Exchange Rates, Journal of Monetary Economics,
May 1989, 23: 433-459.
HSIEH, DAVID A., ‘Modeling Heteroskedasticity in Daily Foreign Exchange Rates,’ Journal of
Business and Economic Statistics, July 1989, 7: 307-317.
KAMINSKY,GRACIELA,AND RODRIGOPERUGA,‘Can a Time Varying Risk, Premium Explain Excess
Returns in the Forward Market for Foreign Exchange? Journal of International
Economics, February 1990, 28: 47-70.
LEE, SANG-WON, AND BRUCE E. HANSEN, ‘Asymptotic Theory for the GARCH(l, 1) Quasi-
Maximum Likelihood Estimator,’ University of Rochester, Working Paper, 1992.
LEE, TAE-HWY, HALBERTWHITE, AND CLIVE W. J. GRANGER,‘Testing for Neglected Nonlinearity
in Time Series Models: A Comparison of Neural Network Methods and Alternative
Methods,’ Journal of Econometrics, April 1993, 56: 269-290.
LUCAS,ROBERT E. JR., ‘Asset Prices in an Exchange Economy,’ Econometrica, November 1978,
46: 1426-1446.
LUMSDAINE,ROBIN L., ‘Asymptotic Properties of the Maximum Likelihood Estimator in
GARCH(l, 1) and IGARCH(l, 1) Models,’ Princeton University, Working Paper, 1991.
MCCURDY, THOMAS H., AND IEUAN G. MORGAN, ‘Test for a Systematic Risk Component in
Deviations from Uncovered Interest Rate Parity,’ Review of Economic Studies, May 1991,
58: 587-602.
MCLEOD, A. I., AND W. K. LI, ‘Diagnostic Checking ARMA Time Series Models Using
Squared-Residual Autocorrelations,’ Journal of Time Series Analysis, 1983, 4: 169-176.
MARK, NELSON C., ‘On Time Varying Risk Premia in the Foreign Exchange Market,’ Journal
of Monetary Economics, July 1985, 16: 3-18.
NELSON, DANIEL B., ‘Stationary and Persistence in the GARCH(l, 1) Model,’ Econometric
Theory, September 1990, 6: 318-334.
PHILLIPS,PETER C. B., ‘Time Series Regressions with Unit Roots,’ Econometrica, March 1987,
55: 277-302.
WEISS, ANDREW A., ‘Asymptotic Theory for ARCH Models: Estimation and Testing,’
Econometric Theory, April 1986, 2: 107-131.
WHITE, HALBERT, ‘Maximum Likelihood Estimation of Misspecified Models, Econometrica,
January 1982, 50: l-26.
WHITE, HALBERT,Asymptotic Theory for Econometricians, Orlando: Academic Press, 1984.
WHITE, HALBERT, ‘An Additional Hidden Unit Test for Neglected Nonlinearity in Multilayer
Feedforward Networks,’ Proceedings of the International Joint Conference on Neural
Network, Washington D.C., San Diego: SOS Printing, 1989, II: 451-455.
WHITE, HALBERT, AND IAN DOMOWITZ,‘Nonlinear Regression with Dependent Observations,’
Econometrica, January 1984, 52: 143-161.
WOOLDRIDGE,JEFFREY M., ‘A Unified Approach to Robust, Regression Based Specification
Tests,’ Econometric Theory, March 1990, 6: 17-43.