American Finance Association

Predicting Volatility in the Foreign Exchange Market

Author(s): Philippe Jorion
Source: The Journal of Finance, Vol. 50, No. 2 (Jun., 1995), pp. 507-528
Published by: Wiley for the American Finance Association
Stable URL: https://www.jstor.org/stable/2329417
Accessed: 29-03-2020 01:01 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Finance Association, Wiley are collaborating with JSTOR to digitize, preserve and
extend access to The Journal of Finance

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
THE JOURNAL OF FINANCE * VOL. L, NO. 2 * JUNE 1995

Predicting Volatility in the Foreign

Exchange Market

PHILIPPE JORION*

ABSTRACT

Measures of volatility implied in option prices are widely believed to be the best
available volatility forecasts. In this article, we examine the information content
and predictive power of implied standard deviations (ISDs) derived from Chicago
Mercantile Exchange options on foreign currency futures. The article finds that
statistical time-series models, even when given the advantage of "ex post" parame-
ter estimates, are outperformed by ISDs. ISDs, however, also appear to be biased
volatility forecasts. Using simulations to investigate the robustness of these results,
the article finds that measurement errors and statistical problems can substantially
distort inferences. Even accounting for these, however, ISDs appear to be too
variable relative to future volatility.

IT IS WIDELY BELIEVED that the volatility implied in option prices is the

market's best estimate of future volatility. After all, if it were not, one could
devise a trading strategy that could generate profits by identifying mispriced
options.
The purpose of this article is to investigate the information content and
predictive power of volatility implied in options on foreign currencies. The
paper focuses on options on currency futures, traded on the Chicago Mercan-
tile Exchange (CME). Because of the depth and liquidity of CME futures and
options, traded side-by-side in the same market, implied standard deviations
(ISDs) allow for clean tests of the predictive power of implied volatilities.
Early studies of the information content of ISDs have generally found that
these contain substantial information for future volatility. Latane and
Rendleman (1976), Chiras and Manaster (1978), and Beckers (1981), for
example, regress future volatility on the weighted implied volatility across a
broad sample of Chicago Board Options Exchange (CBOE) stocks, and find
that options contain volatility forecasts that are more accurate than historical
measures.1 These studies, performed shortly after the 1973 beginning of the
CBOE option market, could only use a relatively short time span, and
therefore focused on cross-sections rather than time-series predictions.

* Jorion is from the Graduate School of Management, University of California at Irvine.

Thanks are due to David Bates, Hendrik Bessembinder, Michael Brennan, Stephen Figlewski,
Steven Heston, two referees, and seminar participants at UC-Irvine, UCLA, Georgetown Univer-
sity, the Universite Libre de Bruxelles, the University of Maryland, the University of Wisconsin-
Madison, and the Cornell conference on derivatives for useful comments. I am also grateful to the
Institute for Quantitative Research in Finance for financial support.
The slope coefficients, however, were generally around 0.5, instead of unity.

507

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
508 The Journal of Finance

More recently, research has turned to the analysis of volatility in a time-

series framework. Scott and Tucker (1989) report some predictive ability in
ISDs measured from PHLX currency options, but their methodology does not
allow formal tests of hypotheses.2 Day and Lewis (1992) analyze options on
the S&P 100 Index from 1983 to 1989, and find that the ISD has significant
information content for weekly volatility, although not necessarily higher
than that of time-series models. This approach, however, ignores the term
structure of volatility since the return horizon is not matched with the life of
the option. To address this problem, Canina and Figlewski (1993) regress the
volatility over the remaining contract life against the implied volatility of
S&P 100 Index options over 1983 to 1986. They report that ISDs have little
predictive power for future volatility, and are significantly biased forecasts.
Furthermore, option volatilities appear to be even worse than simple histori-
cal measures. Finally, Lamoureux and Lastrapes (1993) focus on individual
stock options, carefully measuring prices and matching the forecast horizon,
and find that historical time-series contain predictive information over and
above that of implied volatilities. They view their result as a rejection of the
joint hypotheses of market efficiency and of the Black-Scholes (BS) class of
option pricing models.
As with all efficiency tests, these results can be interpreted either as
indicative of misleading test procedures, or as indicative of inefficient pro-
cessing of information by option markets. The present article investigates
whether these results carry over to the currency option markets, and pro-
vides a detailed analysis of the impact of faulty test procedures due to
measurement errors or inappropriate statistical inferences.
Previous work on the information content of ISD has paid little attention to
the effect of measurement errors in reported prices. These can cause major
problems with estimation of volatility. S&P 100 options, for instance, al-
though actively traded, suffer from several shortcomings. Even with time-
stamped quotes on the underlying S&P 100 Index, it is unlikely that all 100
underlying prices reflect trades that are simultaneous with the option trade.
Given the sheer size of the underlying portfolio, some of the quotes must be
"stale," which creates measurement problems for the option volatility. In
addition, the arbitrage between options and the underlying asset is difficult
to implement because of the transaction costs involved in going long or short
100 stocks simultaneously. Because of the lower transaction costs involved
when trading options, news tends to be more rapidly disseminated in option
prices than in the prices of underlying stocks, which causes measurement
errors in the volatility. Finally, as in all markets, quoted prices may be at the
bid, the ask, or in between, which creates an additional errors-in-variables
problem. Harvey and Whaley (1991) also find that differences in closing

2 Scott and Tucker (1989) present one OLS regression with 5 currencies, 3 maturities
different dates. Because of correlations across observations, the usual OLS standard errors are
severely biased, thereby invalidating hypothesis tests.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 509

times, bid-ask spreads, and to some extent infrequent trading contribute to

biases in the daily autocorrelation of ISDs.
Another potential of source of biases is an inappropriate option pricing
model. In their original derivation, Black and Scholes (1973) assume a
nonstochastic volatility for the underlying asset. Apparently, there is some
inconsistency in recovering an implied volatility from inverting the BS model,
and then proceeding to study the stochastic behavior of volatility. In theory,
one would want to invert an option pricing model consistent with stochastic
volatility. However, no previously published article has done this, for three
reasons. First, current stochastic volatility models involve costly numerical
simulations. It is necessary, not only to compute the option price as a function
of all parameters, but also iterate to find an implied instantaneous volatility.
Second, these models assume a specific time-series process for the volatility.
This model may be misspecified and involves the estimation of additional
parameters, which introduces a supplementary source of error. Finally, for
the short-term at-the-money options considered here, the BS model is very
close to linear in the average volatility, and yields estimates virtually identi-
cal to those of a stochastic volatility model.3
The present study focuses on options on foreign currency futures. These are
actively traded contracts, with average notional volume now approaching $5
billion daily. In addition, both the underlying asset and the options are
traded side-by-side on the CME, and close at the same time. Because of the
low cost of transacting between the two markets, options and futures prices
are likely to be highly synchronized, which alleviates the problem of nonsyn-
chronous quotes afflicting S&P 100 Stock Index options. As an added advan-
tage, all CME closing quotes are carefully scrutinized by the exchange
because they are used for daily settlement, and therefore less likely to suffer
from clerical measurement errors. The currencies selected for this study are
the three most actively traded currencies: the German deutsche mark (DM),
the Japanese yen (JY), and the Swiss franc (SF).
This article investigates both the informational content and the predictive
power of volatility implied in option prices. Informational content is mea-
sured in terms of the ability of the explanatory variable to forecast 1-day
volatility. Tests of predictive power, in contrast, focus on the volatility over
the remaining days of the option contract. ISDs are pitted against time-series
models such as a moving average and Generalized Autoregressive Condi-
tional Heteroskedastic model (GARCH). In order to gain maximum efficiency
within a limited sample period, daily observations are used, and standard
errors are corrected for the data overlap.
Even in the highly synchronized currency futures and options markets,
ISDs are still. affected by measurement errors such as bid-ask spreads. In
addition, the predictive power regressions suffer from small sample biases,

3Fleming (1993) summarizes the theoretical argument for using the Black-Scholes ISD as a
rational forecast of the average volatility over the life of the option. Section II below also shows
that the biases are small for such options.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 510 The Journal of Finance

which render the interpretation difficult. The extent of these problems is

illustrated by Monte Carlo simulations that show that inferences can be
affected by nonnegligible biases. The simulations also provide a method to
gauge whether ISDs can be considered the best available forecasts of future
volatility.
This article is organized as follows. The basic regression setup is presented
in Section 1. Section II shows how the two option pricing models are used to
compute the implied volatility. Section III describes the data and basic
statistics. Predictive regressions are presented in Section IV and interpreted
in light of the Monte-Carlo simulations in Section V. Finally, the last section
contains some concluding observations.

I. Testing Predictive Power

The ISD is widely believed to be the best available forecast of the volatility of
returns over the remaining contract life. Define 0t,T to be the realized
volatility over the remaining life of the contract, measured from day t to day
T. In what follows, the future daily variance is taken from the arithmetic
average of squared returns over future trading days, o-t 2 = (1/(T -
t))ET-tRt2+, without adjustment for the mean.4
The predictive power of a volatility forecast can be estimated by regressing
the realized volatility on forecast volatility:

t = a + b& A tT (1)
where St = (tISD is the volatility forecast measured on day t, taken as the
implied volatility. As in typical tests of market efficiency, if the ISD were the
best available forecast of future volatilities, we would expect the intercept to
be zero and the slope coefficient to be unity.
This framework can be extended to comparisons of the predictive power of
the ISD with that of a simple time-series model, o-tTS. The latter can be tak
as a simple moving average, estimated for example over a 20-day window,
2 =
't2 = E R2frm
(1/20)E 21mr
ER2- i + , or from a more sophistic
the voluminous literature documenting time-variation in second moments in
the foreign exchange market, the paper adopts the GARCH(1, 1) specification,
because it is a parsimonious representation that seems to fit the data
relatively well.5 The GARCH model, developed by Engle (1982) and extended
by Bollerslev (1986), posits that the variance of returns follows a predictable
process, driven by the latest squared innovation and by the previous condi-
tional variance:

t = U + rt, Ft N(0, ht), ht = a0 + alrt_1 ? htf 1 (2)

4With daily data, the average term E(R2) dominates the term E(R)2 by a typical factor of
700 to one. Therefore, ignoring expected returns is unlikely to cause a perceptible bias in the
volatility estimate.
5 For applications, see for instance Hsieh (1989), and Giovannini and Jorion (1989).

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 511

where Rt is the nominal return, rt is the de-meaned return and ht its

conditional variance, measured at time t. To insure invertibility, the sum of
parameters (a1 + 1) must be less than unity; when this is the case, the
unconditional, long-run, variance is given by ao/(l - a1 - ,B).
When pitted against a time-series model, as in the following regression,

?t,T = a + b1otISD + b2 tTS + 8t,T (3)

we would expect o- TS to have no incremental predictive power beyond that in

ISD. In other words, the coefficient b2 should be close to zero.
The previous specification precisely matches the ISD with the volatility
over the remaining contract life, which is appropriate if one wants to test the
forecasting accuracy of implied volatilities. Another, related, issue is the
information content of the daily ISD for volatility over the next day, which
can be tested as before,

R - a + b(t sD+rt?i. (4)

This approach is similar in spirit to the work of Day and Lewis (1992), who
analyze the information content of the ISD with respect to future weekly
stock index volatility. This setup, however, addresses the question of whether
there is some useful information in o ISD, rather than whether o ISD is the
best available forecast of future volatility. Since the forecast horizon is not
matched with the realized returns, we only require the slope coefficient b to
be positive, and not necessarily unity.
Because currency options are relatively recent financial instruments, the
limited data span makes it important to fully utilize all the information in
the sample. Therefore, daily data are used in the regressions. With horizons
up to three months, however, using all the daily data introduces overlaps in
the error terms, which causes downward bias in the usual ordinary least
squares (OLS) standard errors. The point estimates of the coefficients, how-
ever, are still consistently estimated. Intuitively, the reason for this bias is
that OLS estimation assumes that each day brings a completely independent
observation, and therefore that the large number of observations permits
precise estimation of the coefficients, when in fact most of the new observa-
tions are redundant.
Hansen (1982) provides a correction that extends White's (1980) method to
deal with heteroskedasticity and properly deals with serial correlation.6 The
corrected Hansen-White (HW) variance-covariance matrix of estimated coeffi-
cients is given by

I = (X'X)-1l(X'X)-1, (5)

6 Hansen and Hodrick (1980) first apply this me

forward rates measured at monthly intervals. Canina and Figlewski (1993) apply this method to
volatility forecasts.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 512 The Journal of Finance

where fl = E[ X'se'X/T] is consistently estimated, using the OLS residuals

A

, by

fit= sLet
t
XtXt + E EQ(s, t)ASA

with Q(s, t) defined as an indicator function equal to unity if there is overlap

between returns at s and t, and zero otherwise. Note that in the case where
the residuals are homoskedastic and do not overlap, E[ s2] = s2, and Q(s, t)
is always zero, so that the covariance matrix collapses to the usual OLS
covariance matrix s2(XX)-Y1.

II. Computing Implied Volatilities

Implied volatilities are derived from the Black (1976) model for European
options on futures,

rr ln(FIK) 0-17T
c = [FN(dl) - KN(d2)]err, dd2 = d (7)

where F is the futures rate, K is the strike price, i- is the time to option
expiration, r is the risk-free rate, taken as the Eurodollar rate, and o- the
volatility. For a given option price, inverting the pricing model yields an
implied standard deviation. Because Beckers (1981) shows that using only
at-the money options was preferable to various other weighting schemes, only
at-the-money calls and puts are considered here. In addition, these are the
most actively traded, and therefore least likely to suffer from nonsimultane-
ity problems. On any given day, the ISD is computed as the arithmetic
average of that obtained from the two closest at-the-money call and put
options. Averaging a call and a put alleviates some of the measurement
problems.
Since CME options are of the American type, using a European model
introduces a small upward bias in the estimated volatility. This bias is
generally considered small for short-maturity options.7 For instance, with
typical parameter values,8 using a European model overestimates a 12

7 As shown by Whaley (1986), the bias depends on the level of the interest rate, the volatility,
the time to expiration, and the degree to which the option is in-the-money. Shastri and Tandon
(1986) use numerical procedures to show that biases in measured implied volatilities are
generally minor for short-term at-the-money option. However, some error is also introduced
because the nearest at-the-money option changes from day to day, which creates some spurious
movements in ISDs.
8 With a futures prices of 50 cents, a strike price of 50, a U.S. interest rate of 6 percent, 50
calendar days to expiration, and a true volatility of 12 percent, the value of an American and
European call is 0.8799 and 0.8786, respectively. Inverting the American call value using a
European model yields an apparent volatility of 12.02 percent. With the same parameters but 95
days to expiration, the estimated volatility is 12.04 percent. With 5 days to expiration, it is 12.00
percent.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 513

percent true volatility by reporting a value of about 12.02 percent. The

difference, however, is much less than typical bid-ask spreads, when quoted
in terms of volatility, and in any event tends to bias the slope coefficient
downward by a very small amount. Given the direction of the bias, findings
of significance indicate that rejection of the hypothesis of no forecasting
ability for ISDs would also occur with volatilities derived from a more
accurate American model.9
Another potential misspecification is that the BS model is, stricto sensu,
inconsistent with stochastic volatilities. If volatility changes in a determinis-
tic fashion, the implied volatility can be construed as an average volatility
over the remaining life of the option. But if volatility is stochastic, the
arbitrage argument behind the BS option pricing model fails.
Recent papers by Hull and White (1987), Scott (1987), and Wiggins (1987)
have examined the pricing of options on assets with stochastic volatility. The
general approach to pricing options in these papers is to treat the volatility as
a random state variable. In order to derive tractable results, the innovations
in volatility and returns are generally assumed to be uncorrelated; prices are
then calculated by Monte Carlo simulation. Scott (1988) and Chesney and
Scott (1989), for instance, present a careful empirical analysis of the random
variance model (implemented on a Cray supercomputer), and find that the
random variance model actually provides a worse fit to market prices than
the Black-Scholes model using ISDs. For U.S. stock options, differences are
only on the order of 2 cents, which is much lower than typical bid-ask spreads
of 5 to 25 cents. Duan (1995) extends the risk neutral valuation to the case
where logarithmic returns follow a GARCH process. Under some combination
of preferences and distribution assumptions, he derives a GARCH option
pricing model, but the magnitude of the bias, computed by simulations, is
very small, at most 10 to 15 cents for an option on a $100 underlying asset.
Because of the computational costs involved in pricing options, no pub-
lished research has ever recovered the implied (instantaneous) standard
deviation from a stochastic volatility model.10 It is only recently that Heston
(1993) has developed a closed-form solution that efficiently computes option
values under stochastic volatility. These models, however, introduce addi-
tional difficulties. The time-series process for the volatility may be misspeci-
fied. Also, the models involve the estimation of additional parameters, which
introduces a supplementary source of error. Finally, the gain from these
stochastic volatility models is limited because the mispricing is very small for

9With the above numbers, using a European model decreases the estimated slope coefficient
from unity to 0.998. This bias, however, cannot account for the slope coefficients below.
10 Melino and Turnbull (1990), for instance, compare option prices derived from Black-Scholes
and a stochastic volatility model, using parameters derived from the time-series process, and
find that the stochastic volatility model provides a better fit to options than the standard model
using historical volatility. They do not, however, consider a Black-Scholes model with implied
volatility.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 514 The Journal of Finance

short-term at-the-money options.11 Thus, the BS approach provides a good

working approximation to more complicated stochastic volatility models for
the options considered in this paper.

III. Data and Descriptive Statistics

The data are taken from the CME's closing quotes for currency futures and
options on futures. The currencies covered are the German deutsche mark
(DM), Japanese yen (JY), and Swiss franc (SF). These are the most active
contracts on the CME.12 For the DM, the sample period covers January 1985
to February 1992, which represents more than 7 years of daily data, or about
1810 observations. The data start in July 1986 for the JY, and in March 1985
for the SF.
We now turn to the timing of observations and the issue of matching ISDs
with future volatility. For each option contract, the implied volatility is
matched with the sequence of price movements on the underlying futures
contract until option expiration. The "realized"-or "future"-volatility is
then taken from the variance of continuously compounded futures returns.13
In what follows, the observation interval is taken as a trading day. If
volatility was primarily the result of exogenous information, and if this flow
was constant for every day, the weekend variance should be three times the
weekday variance. It is well-known, however, that the variance over week-
ends is only slightly higher than over weekdays. Baillie and Bollerslev (1989),
for instance, report that the average daily variance is 0.448 for the DM/$
rate, and that the Friday-to-Monday variance is 0.569, which is closer to the
daily variance than to the variance over three weekdays. As suggested by
French (1984), consistent with these findings, the volatility over the remain-
ing life of the contract was converted to a daily volatility using the number of
trading days, instead of calendar days. Annualized volatility measures are
obtained by multiplying by the square root of 252, which is the number of
trading days in a calendar year. Interest rates, as always, are associated with
calendar days.
Contracts expirations follow the usual March-June-September-December
cycle. On the first day of the expiration month, which is the time around
which most rollovers into the next contract occur, the option series switches

11 An earlier version of the paper provides tests using Heston's stochastic volatility model,
finds that the model cannot explain the bias in implied volatility reported below. A major
difficulty in the implementation of this model is that it requires knowledge of additional
parameters for the volatility process. Some of these can be derived from the time-series of
exchange rates, using the GMM method proposed by Bates and Penacchi (1991), but are subject
to estimation error.
12 Over the period 1985 to 1991, the average daily volume of futures contracts was 30,340
the deutsche mark, 22,290 for the Japanese yen, 21,850 for the Swiss franc, 11,540 for the
British pound, and 4,150 for the Canadian dollar.
13 The horizon of option contracts varies from 3 to 100 calendar days, or from 3 to 70 tradi
days.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 515

into the next quarterly contract. This generates a time-series of continuous

1-day returns, implied volatility, and matching realized volatility.
Table I provides a brief description of the data. Means, standard deviations,
and autocorrelations are presented for returns and different measures of
volatility: next day volatility, volatility realized over the contract life, option
implied volatility, and the two time-series models.
As expected, daily returns appear to be uncorrelated; for the DM, their
standard deviation is 0.787 percent per day, or 12.5 percent per annum. The
1-day volatility, in contrast, shows small positive but significant autocorrela-
tion at various lags, which suggests that volatility is persistent.
Looking next at statistics for the realized (future) volatility, we find num-
bers that are typical of an averaged series with data overlap: very high
autocorrelation and smaller standard deviation, about 0.197, than daily
volatility, which is about 0.534 for the DM. There is, however, still some
positive autocorrelation beyond the 100-day lag, which is the longest possible
data overlap.
Finally, the implied volatility displays a similar pattern of autocorrelation;
the lower standard deviation of 0.142 for the DM is consistent with the fact
that the latter series is an expectation of the previous series. The two series
have very similar average values: for instance, 0.773 and 0.745, respectively,
for the DM. Based on this table, there is no indication that ISDs systemati-
cally underestimate or overestimate future volatility.
The results are qualitatively similar across currencies. It should also be
kept in mind, when comparing regressions across these three currencies, that
the results are not independent since the three exchange rates are positively
correlated. 14
Figure 1 displays the time-variation in ISD for the DM, measured in
percent per annum, as well as the corresponding realized volatility. The
volatility averages about 12 percent per annum and displays substantial
time-variation, which seems to be reflected in both series.
Estimates of the GARCH(1, 1) process are presented in Table II. In line
with previous research, the GARCH model is highly significant, with a X
statistic exceeding 60 for all currencies. This is much higher than the 1
percent upper fractile of the chi-square, which is 9.2. There is no question,
therefore, that realized volatility does change over time. The process is
persistent but also stationary, with values of (a1 + ,B) around 0.96 for the
DM and SF. This number implies that a shock to the variance has a half-life
of ln(0.5)/ln(0.96), which is about 17 days.

IV. Forecasting Regressions

We first test the information content of the ISD for next day volatility with
regression equation (4). Results are presented in Table III, with one panel for

14 The pairwise correlations in the futures price change over the 1985-1992 period ar
DM/JY, 0.705, DM/SF, 0.898, and JY/SF, 0.682. The DM and SF thus move more closely to
each other.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
516 The Journal of Finance

SF Jy

DM

TableI

Autocrelain

DescriptvSa:RundVolyM

andSwisfrc(F).Peobuy28,19tJ5DMl6Yh

GARCHvolatiy0.854126937 MA(20)volatiy.83975416 ISDvolatiy0.831497265 Futrevolaiy0.891264357 1-Dayvolti0.64578392 1-Dayretun0.2986537 GARCHvolatiy0.751286943- MA(20)volatiy.67489135- ISDvolatiy0.65819274- Futrevolaiy0.65318972-4 1-Dayvolti0.59487623 1-Dayretun0.8735246 GARCHvolatiy0.78156943 MA(20)volatiy.75493816 ISDvolatiy0.731429586 Futrevolaiy0.745196832 1-Dayvolti0.57834692 1-Dayretun0.3782564
MeanStd.Dv123450
avergothps20dy),GARCHcnilfx.umk(DMJY volatiy(ermngfscdp),ISDMA20 Means,tdrviouclf1-yxp
This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 517

25%

Realized

15%

50% V

Implied
0%oo
85 86 87 88 89 90 91 92

Figure 1. Implied and realized volatility: Mark. Time-series of volatility implied from
option prices and of volatility subsequently realized over the life of the option contract.

each currency. Since the dependent variables do not overlap, standard errors
are simply computed by OLS. The table shows that the ISD contains a
substantial amount of information for currency movements the following day.
In the first panel, the slope coefficient is 0.852 for the DM, and highly
significant. Similarly high coefficients are obtained for the JY and SF.
Dividing the sample period into two subperiods leads to similar conclusions.
Forecasting with a time-series model also produces significant results. The
slope of the 20-day moving average is 0.395, and that of the GARCH 1-day
forecast is 0.598 for the DM; both are significant for all currencies. The
GARCH model appears less biased than the simple moving average model,
although its explanatory power, expressed in terms of R2, is slightly lower.
This is in spite of the fact that the parameters of the GARCH models were
estimated over the whole period 1985 to 1992, and that we should conse-
quently expect this model to outperform the MA(20) volatility. Although the
GARCH parameters are generally stable over time, there is no guarantee
that market participants would have been able to use the GARCH model as
early as in 1985 (especially since the model was not yet published).
The next lines in the DM panel pit the ISD against the time-series models.
There is no information content of the time-series models beyond that in
options for all three currencies. These results are reassuring because they
indicate that option traders form better expectations of risk over the next day
than statistical models, even when the latter are based on ex post parameter
values. To some extent, these results were expected, since time-series models
are unable to account for events such as regular announcements of macroeco-
nomics indicators, meeting of G-7 finance ministers, and so on. Because the
timing of these events is known by the foreign exchange market, we would
expect options to provide better forecasts than simple time-series models.
Predictability, however, implies that the ISD is an unbiased forecast over
the remaining option life, not over the next day. There could very well exist a

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
518 The Journal of Finance

(1.43)2976 (0.4)267 (1.63)07

(1.4)6382 (0.49)72586 (1.65)408793 TableI

ht=o+alr21?- Rt=u+,?N(Oh)

ofsigncaedGARCHpr.tumy
EstimaonfGARCH(1,)Prce

SFNormal0.29374165 JYNormal0.82491657 DMNormal0.32619485

GARCH0.24351968 GARCH0.85697123 GARCH0.287651349 Long-RuVlatiyX2()

CurencyModlat,f(%p)Lg-ikhv
1985(DM),July6YarchSF.Asmptoi-neTx2 aredutschmk(DM),JpnyYSwifF.Pob2819ar wherRtisdfnauocy,lv.Ce

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 519

Table III

Information Content Regressions

VRt1 =a+ b5t + 1

where Rt+ 1, the futures return over the next day, is regressed against the volatilit
This includes the implied standard deviation (ISD) from option prices, a moving average (MA)
with a moving window of 20 trading days, and the GARCH time-series model, which is the
conditional volatility for the next day based on parameters in Table II. Periods end on February
28, 1992, and start in January 1985 (DM), July 1986 (JY), and March 1985 (SF). OLS standard
errors are in parentheses.

Slopes On

Currency a ISD MA(20) GARCH R2

DM - 0.080 0.852* 0.0515

(0.068) (0.086)
0.282* 0.395* 0.0315
(0.041) (0.052)
0.113 0.598* 0.0304
(0.063) (0.079)
- 0.069 0.784* 0.055 0.0518
(0.069) (0.126) (0.074)
-0.095 0.785* 0.085 0.0518
(0.070) (0.123) (0.112)

JY - 0.006 0.783* 0.0355

(0.072) (0.108)
0.283* 0.335* 0.0206
(0.043) (0.061)
0.119 0.560* 0.0190
(0.075) (0.107)
0.001 0.678* 0.093 0.0365
(0.072) (0.140) (0.079)
- 0.076 0.668* 0.210 0.0374
(0.083) (0.128) (0.125)

SF -0.041 0.854* 0.0406

(0.080) (0.099)
0.335* 0.373* 0.0235
(0.049) (0.057)
0.071 0.672* 0.0223
(0.091) (0.106)
- 0.039 0.840* 0.011 0.0406
(0.083) (0.150) (0.086)
- 0.040 0.858* - 0.005 0.0406
(0.092) (0.148) (0.157)

* Significantly different from zero at

"term structure" of volatility if the market knows that information will be

released at specific points in the future. With monthly trade figure announce-
ments, for instance, the volatility could be higher on the third Thursday of
every month, and lower in surrounding days. Because the volatility is not

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
520 The Journal of Finance

necessarily constant over all contract days, it may not be appropriate to treat
the current ISD as a forecast of next day's volatility.
The predictive power hypothesis is tested using equation (1), with results
presented in Table IV. Here, the GARCH forecast is obtained by successively
solving for the expected variance for each remaining day, then averaging over
all days. Comparing to Table III, we observe that the R2s are systematically
higher than before, which was to be expected, since there is more noise in
daily volatility than in an average measure.
Table IV shows that ISDs contain a substantial amount of information for
future volatility. The slope coefficient is 0.547 for the DM, with an HW
t-statistic of 3.96 for the hypothesis that ,( = 0, which is significantly higher
than zero. The results are remarkably similar across currencies. ISDs, how-
ever, also appear to be biased predictors of future volatility. The estimate of
the slope coefficient is less than unity, and the associated HW t-statistic is
3.29 for the hypothesis that 1 - ,3 = 0, which is statistically significant. The
slope coefficient less than unity combined with a positive intercept suggests
that ISDs are too volatile: when ISDs are high relative to the average, they
should be scaled down; when ISDs are low relative to the average, they
should be brought up toward the intercept.
For all currencies, the MA(20) and GARCH time-series forecasts have
lower explanatory power than ISDs. As in the case of the information content
regressions in Table III, the slope coefficients become smaller and insignifi-
cant when using both the ISD and a time-series model in the same regres-
sion. The R2 of the regressions is also noticeably higher when including ISDs.
For the DM, for example, the R2 increases from 0.0499 for the GARCH model
to 0.1599 when including the ISD. This is a substantial improvement. Results
for the SF and the JY are very similar. In each case, the additional coefficient
on the MA(20) or the GARCH volatility is not significant; most of the
information content lies in ISDs. These results indicate that options provide
informative forecasts of future volatility that are superior to those of time-
series models.
These results are in sharp contrast to those of Canina and Figlewski
(1993), who report slope coeffilcients on ISDs and on a MA(60) model of 0.229
and 0.573, respectively. Combining both forecasts in the same regression,
they find that the ISD coefficient drops to 0.077 and becomes insignificant.
The poor performance of option volatilities, therefore, is sharply reversed in
the foreign currency option market.
Since it is difficult to argue that option traders are smarter in the foreign
exchange market than in the U.S. stock market, the most logical interpreta-
tion of the differences in the results is that S&P 100 Index option ISDs have
been measured with substantial error because of stale prices and because of
the difficulty of arbitraging between the option and underlying stock mar-
kets.
As an illustration of the seriousness of the problem, consider the impact of
changes in underlying asset prices on the implied volatility. For a given

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 521

Table IV

Predictability Regressions

St,T = a + b&AT + 8t,T

where U:t,T, the volatility over the remaining life of the option contract, is re
volatility forecast 7tT. This includes the implied standard deviation (ISD) fr
moving average (MA) with a moving window of 20 trading days, and the GARCH time-series
model, which is the conditional volatility for the next day based on parameters in Table II.
Periods end on February 28, 1992, and start in January 1985 (DM), July 1986 (JY), and March
1985 (SF). Regressions use daily observations, and standard errors are corrected for the induced
overlap and heteroskedasticity using the Hansen-White (HW) procedure. Asymptotic HW stan-
dard errors in parentheses.

Slopes On

Currency a ISD MA(20) GARCH R2

DM 0.323* 0.547*' 0.1564

(0.115) (0.138)
0.602* 0.190 0.0540
(0.084) (0.099)
0.366 0.478*' 0.0499
(0.191) (0.227)
0.303* 0.669* - 0.099 0.1632
(0.112) (0.165) (0.101)
0.401* 0.622* -0.173 0.1599
(0.152) (0.167) (0.201)

JY 0.327* 0.496*' 0.0965

(0.118) (0.181)
0.563* 0.134' 0.0223
(0.074) (0.102)
-0.063 1.017* 0.0495
(0.323) (0.458)
0.322* 0.578* - 0.073 0.1004
(0.117) (0.204) (0.111)
0.042 0.421* 0.474 0.1051
(0.289) (0.177) (0.399)

SF 0.392* 0.520* 0.1454

(0.149) (0.175)
0.658* 0. 182T 0.0542
(0.087) (0.099)
0.250 0.650* 0.0581
(0.267) (0.305)
0.370* 0.647* - 0.097 0.1521
(0.146) (0.187) (0.090)
0.526* 0.609* -0.240 0.1490
(0.210) (0.201) (0.262)

* Significantly different from zero

Significantly different from unity at the 5 percent level.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
522 The Journal of Finance

movement in the price S, a- will change by

do- dc AS

A f=Sdc dS S .(8

For one-month at-the-money options, the delta is about 0.5, and (dc/do-)/S
is about 0.10, using annualized volatility. Now consider that the standard
deviation of the S&P 100 Stock Index is about 20 percent, which is 1.3
percent on a daily basis. If some of the 100 stock prices are measured with
errors due to stale prices or bid-ask spreads, the error between the measured
and true index values could easily reach AS/S = 0.25 percent.
Stoll and Whaley (1990), for instance, demonstrate that the S&P 500 Index
suffers from serious measurement problems due to infrequent trading and
bid/ask price effects. As a result, index prices display very high autocorrela-
tion over 5 to 15 minutes lags, and are led by futures prices. In addition,
Harvey and Whaley (1992) report that the stock market and option markets
close at 3:00 and 3:15, respectively, and that the index volatility over this
15-minute interval is typically 0.18 percent.
Therefore, using an error of 0.25 percent in the measurement of the
underlying spot price appears conservative. Using this number, the error in
implied volatility is (1/0.1) x 0.5 x 0.25 percent = 1.2 percent per annum.
This error is very substantial, because it is on the same order of magnitude as
the daily variation in actual volatility, and may therefore seriously affect
measurements of ISDs. In contrast, these errors are likely to be much smaller
for options on currency futures, since there is only one underlying asset, since
both the underlying futures and option markets close at the same time, and
since both contracts are actively traded. This may explain why implied
volatilities are much better predictors of future volatility in the currency
futures market than in the S&P 100 Index option market. It would be
interesting to check whether ISDs derived from options on S&P 500 futures
provide better forecasts than those obtained from the S&P 100 cash options.

V. Simulations

The previous regressions suggest that the ISD is an informative, albeit

biased, volatility forecast. Their interpretation, however, is beset by econo-
metric problems. As indicated in Table I, implied and future volatility are
highly autocorrelated processes. In such a situation, we could expect to find
small-sample biases such as those reported by Dickey and Fuller (1979).
When regressing a random variable on its lagged value, they found that the
slope coefficient was biased down toward zero in small samples; the bias is
generally greatest when the true value of the slope is unity, but also occurs
for lower values. The driving factor in their results is the high degree of
persistence in the regression variable. Given the persistence of option volatili-
ties, the question is whether this bias also occurs here.
A second problem with the above regressions is that the HW correction is
only valid asymptotically. In this case, "large" number of observations is not

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 523

measured by the total sample size, which is about 1,800, but rather by the
number of truly independent forecasts, which is about 7 years of data
multiplied by four contracts per year. Thus we have no measure of how good
the asymptotic approximation is.
Finally, and perhaps most importantly, the ISD is still measured with
error. Because a closing option quote can represent a bid price, an ask price,
or an intermediate price, the observed ISD measures the true ISD with some
error. This can be judged from the over the counter market, where foreign
currency options are directly quoted in terms of bid/ask volatility, such as
"14.2-14.6." Thus a typical spread is 0.4 percent per annum. Conversations
with CME traders have confirmed that, for active at-the-money options,
spreads are also around 0.3 percent in terms of volatility. Even when averag-
ing over two options, the error drops from 0.4 to 0.28 percent per annum,
assuming independent positions within the bid-ask spread. If measurement
errors are independent of regression residuals, this errors-in-variables prob-
lem will bias the estimated slope coefficient downward. A priori, it is difficult
to judge the extent of the bias, since it depends on the relative size of
measurement error and of the true time-variation in expected volatility.
To assess whether these problems affect the results, we present simula-
tions that are constructed under the null hypothesis of predictability. We
need to design an experiment where the characteristics of volatility closely
resemble those of actual data: (i) the future distribution of returns from each
day forward must be derived from the present volatility position, (ii) the
pseudoimplied volatility must be the average volatility consistent with the
underlying volatility process, (iii) each new day resets the whole future
pattern of volatility, and the realized volatility is measured from the time-
series of ex post returns.
Given that the GARCH(1, 1) model provides an acceptable parsimonious
representation of the statistical behavior of volatility, this is the model
chosen for the underlying daily volatility. We also keep the actual spacing of
contract expiration over the 1985 to 1992 sample period. Since the slope
coefficient was the highest for the DM among the currencies examined, the
simulations only use DM data.
The pseudo-volatility is measured as follows. Using the actual GARCH
parameters in Table II, we generate each day ts conditional volatility over
the next day. The variance is also projected for all remaining contract days,
using

ht+l = oao + alr 2 + 3ht

Et[ht+2] = ao + (a1 + f)ht+1,
Et[ht+3] = ao + (a1 + f3)Et[ht+2],

and so on, which converges to Et[h] = ao/(l - a1 - 83). The average forecast
variance on day t is therefore

ht,T = (ht+1 + Et[ht+2] + ? +Et[hT])/(T - t). (9)

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 524 The Journal of Finance

The following day, we sample from a standard normal distribution15 the

normalized conditional return z* , where the asterisk refers to the fact that
this is a pseudorandom variable. The return the following day is generated
from r* = z*hi, and so on.
Using the actual contract sequence, a pseudovariance is measured over the
remaining life of the contract o*2, which is then matched against the
average forecast variance. Since the same structure is used to generate the
daily innovations and the forecasts, the model is built under the assumption
of predictability.
After obtaining as many observations as we had in the original sample, we
regress

ot,T= a + bVtT + , t,T, (10)

and record the slope coefficient and associated HW t-statistics. We repeat the
experiment 5,000 times, and compare the actual statistics to their empirical
distribution, by reporting the "empirical" p-value as the proportion of times
the observed statistic was exceeded under the null.
In addition, the simulation experiment can be easily modified to incorpo-
rate the effect of errors-in-variables. When regressing 0tT over ,h/t, a
zero-mean random variable uniformly distributed over the bid-ask spread
interval can be added to the regressor.
Table V reports the relevant quantiles, means, and standard deviations of
the bootstrapped distributions of regression coefficients and t-statistics under
the null hypothesis. With no noise added to the volatility, there is a down-
ward bias in the slope coefficient: its mean and median values are both 0.907,
instead of unity. The bias is substantial, since it is on the order of magnitude
of one standard error, even with 1,810 observations. The standard deviation
of the simulated bs is 0.160, which is only slightly higher than the 0.138
estimate of standard error in Table IV.
The actual slope of 0.547, reported in Table IV, however, falls in the
extreme left tail of the empirical distribution, with an empirical p-value of
1.1 percent. Similarly, the HW t-statistic for the hypothesis that 1 - / = 0 is
3.29, which is only exceeded in 1.9 percent of the sample. Alternatively, the
empirical distribution of b s shows that the value of 0.547 is (0.907 -
0.547)/0.160, or 2.3 standard deviations away from the mean. Although muc
lower than the tHW of 3.29, this is still significant. Based on this experime

15 Actually, there is evidence that the normalized residuals rt/ Xht have slightly fatter
than a normal distribution for daily returns. We also have bootstrapped those from the empi
distribution for the simulations. Because there are more observations in the tails, this lead
distributions of /'s, t's that have fatter tails than under the normality assumption. Usin
bootstrapped sample, we find that the empirical p-value changes from 0.0114 to 0.0384 for th
distribution. Thus rejections still occur when allowing for nonnormal conditional distributi
Note that a misspecification of the distribution function is not a problem here. White (198
shown that maximum likelihood estimation using a normal distribution yields consistent es
mates of the mean and variance of distributions for which these quantities are finite.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 525

sampleof108rnt4yd.

TableV

o=,Ta+brt?-

PanelA:Sprd=0,iz186

PanelD:Sprd=0.%,iz18
PanelC:Sprd=2.0%,iz186 PanelB:Sprd=0.4%,iz18

SimulatonsfPredcbyRgUN:DM

Simulatedrgsonfhvy,Tc:

tHW(l-/=0)1.52893647 tHW(,8=0)2.1954367 /30.41562897 tHW(1-/3=0).6495278 tHW(,8=0)2.3461579 /30.47561892 tHW(1-/3=0).4856297 tHW(,B=0)2.83546971 /30.59641827 tHW(l-3=0)1.4926758 tHW(,=0)2.83549176 ,B0.54167928 Staisc0.159MenDvp-Vlu
FractilesofSd.AuEmp
oftime-srbvan.Thulpcd1,806gJy5F92 term,disbunfolyva0.4%phg;c-kSz thempircal-vu,wsonfbdx.US=04% ofthebcinadHs-Wrmplyg.Twu GARCHproces.ntaxiylhmuTbd, sampledfroitbunwgGARCH(1,)cvh;y Thesimulaton5,0rpcdyfb.Fv*2

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
526 The Journal of Finance

we would conclude that there is some evidence that option volatilities are
biased volatility forecasts, assuming that ISDs were measured without error.
The lower panels in the table investigate the impact of measurement
errors. When adding noise with mean zero and uniformly distributed on a 0.4
percent per annum range, which is a typical bid-ask spread, the fractiles of
the statistics undergo only small downward shifts. The mean of the slope
coefficient, for instance, drops to 0.901, and the p-value of the observed
statistic increases from 1.1 percent to 1.2 percent. Thus the impact of typical
bid-ask spread is small. In order to obtain empirical p-values that approach
conventional 5 percent levels, we need to increase the measurement error to
2.0 percent, which is an exceedingly large spread. The third panel in the table
shows that the mean then drops to 0.787, with a p-value for the observed /3
of 4.2 percent, and for the t-statistic of 9.1 percent.
Finally, the lower panel repeats the first experiment with only four years of
data, which is similar to the sample size used by Canina and Figlewski
(1993). As expected, the distribution of /3s is wider than for the whole sample,
and empirical p-values now approach the usual 5 percent level. Given the
very low slopes observed in their study, however, it is unlikely that the
observed statistics would fall well inside the empirical distributions reported
here.
Overall, the simulation evidence demonstrates that substantial biases exist
in regression tests of the predictive power of ISDs. These biases, however, are
unable to explain the observed regression results, unless one accepts rather
large measurement errors.

VI. Conclusions

Previous work on the information content of S&P 100 Index options has
claimed that ISDs were biased forecasts of future volatility, and were found
to be sometimes worse than the simplest statistical models. These results can
be given two possible interpretations: either the test procedure is faulty, or
option markets are inefficient.
Faulty test procedures can be due to (i) measurement errors, (ii) inappro-
priate statistical inferences, or (iii) using the wrong pricing model. The option
pricing model can be wrong if volatility or interest rates are stochastic, if the
value of early exercise is misspecified, or if underlying asset prices do not
follow a diffusion process.
We use implied volatilities inverted from the BS model for short-term
at-the-money options on the DM, JY, and SF. Previous work has shown that
the BS model and stochastic volatility models product ISDs that are quite
similar for such options; differences are only on the order of bid-ask spreads.
The present article focuses on measurement errors and statistical infer-
ences as possible explanations for these biases. Because options on currency
futures are traded side-by-side with the underlying futures, they are less
likely to suffer from the measurement problems affecting cash index options.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 Predicting Volatility in the Foreign Exchange Market 527

We still recognize, however, the effect of bid-ask spreads, and provide simula-
tions to fully account for small-sample biases usually ignored in regression
tests of predictive ability. We show that measurement errors can substan-
tially distort inferences.
In contrast with stock index options, we find that statistical time-series
models, even when given the advantage of ex post parameter estimates, are
outperformed by option-implied forecasts. These results hold for the three
currencies under consideration. Even when accounting for possible measure-
ment errors and statistical problems, we still find, however, that ISDs are
biased volatility forecasts. The direction of the bias is such that options ISDs
appear too variable relative to future volatility.16 A linear transformation of
the ISD provides a superior forecast of exchange rate volatility. These results
are consistent with those of Fleming (1993), who analyzes S&P 100 Index
options, and also reports that ISDs dominate moving-average volatility, using
a more precise estimation technique.
The ultimate test of whether the reported biases are economically, as
opposed to statistically, significant is to simulate a dynamic hedging trading
rule that attempts to take advantage of mispriced options. The question then
is whether such a rule, after transaction costs, generates significant economic
profits. This is left for future research.

16 Overreactions in stock index options have also been reported by Stein (1989).

REFERENCES

Baillie, R., and T. Bollerslev, 1989, The message in daily exchange rates: A conditional-variance
tale, Journal of Business and Economic Statistics 7, 297-306.
Bates, D., and G. Pennacchi, 1991, Estimation of a heteroskedastic state space model of asset
prices, Mimeo, University of Pennsylvania.
Beckers, S., 1981, Standard deviations implied in option prices as predictors of future stock pric
variability, Journal of Banking and Finance, 363-381.
Black, F., 1976, The pricing of commodity options, Journal of Financial Economics 3, 167-179.
Black, F., and M. Scholes, 1973, The pricing of options and corporate liabilities, Journal of
Political Economy 81, 637-659.
Bollerslev, T., 1986, Generalized autoregressive conditional heteroskedasticity, Journal of
Econometrics 31, 307-327.
Canina, L., and S. Figlewski, 1993, The informational content of implied volatility, Revie
Financial Studies 6, 659-681.
Chesney, M., and L. Scott, 1989, Pricing European currency options: A comparison of the
modified Block-Scholes model and a random variance model," Journal of Financial and
Quantitative Analysis 24, 267-284.
Chiras, D., and S. Manaster, 1978, The information content of option prices and a test of m
efficiency, Journal of Financial Economics 10, 29-58.
Day, T., and C. Lewis, 1992, Stock market volatility and the information content of stock index
options, Journal of Econometrics 52, 267-287.
Dickey, D., and W. Fuller, 1979, Distribution of estimators for autoregressive time series with a
unit root, Journal of American Statistical Association 74, 427-431.
Duan, J.-C., 1995, The GARCH option pricing model, Mathematical Finance 5, 13-32.
Engle, R., 1982, Autoregressive conditional heteroskedasticity with estimates of the variance of
United Kingdom inflation, Econometrica 50, 987-1007.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms
 528 The Journal of Finance

Fleming, J., 1993, The quality of market volatility forecasts implied by S&P 100 index option
prices, Mimeo, Rice University.
French, D., 1984, The weekend effect and the distribution of stock prices: Implications for option
pricing, Journal of Financial Economics 13, 547-559.
Giovannini, A., and P. Jorion, 1989, The time-variation of risk and return in the foreign exchange
and stock markets, Journal of Finance 44, 307-325.
Hansen, L., 1982, Large sample properties of generalized method of moments estimators,
Econometrica 50, 1029-1054.
, and R. Hodrick, 1980, Forward exchange rates as optimal predictors of future spot
rates, Journal of Political Economy 88, 829-853.
Harvey, C., and R. Whaley, 1991, S&P 100 index option volatility, Journal of Finance 46,
1551-1561.
, 1992, Market volatility prediction and the efficiency of the S&P 100 index option
market, Journal of Financial Economics 31, 43-74.
Heston, S., 1993, A closed-form solution for options with stochastic volatility with applications to
bond and currency options, The Review of Financial Studies 6, 327-343.
Hsieh, D. 1989, Modeling heteroskedasticity in daily foreign exchange rates, Journal of Business
and Economic Statistics 7, 307-317.
Hull, J., and A. White, 1987, The pricing of options on assets with stochastic volatility, Journal
of Finance 42, 281-300.
Lamoureux, D., and W. Lastrapes, 1993, Forecasting stock-return variance: Toward an under-
standing of stochastic implied volatilities, Review of Financial Studies 6, 293-326.
Latane, H., and R. Rendleman, 1976, Standard deviation of stock price ratios implied by option
premia, Journal of Finance 31, 369-382.
Melino, A., and S. Turnbull, 1990, Pricing foreign currency options with stochastic volatility,
Journal of Econometrics 45, 239-265.
Scott, E., and A. Tucker, 1989, Predicting currency return volatility, Journal of Banking and
Finance 13, 839-851.
Scott, L., 1988, Random variance option pricing: Empirical tests of the model and delta-sigma
hedging, Mimeo, University of Illinois.
Scott, L., 1987, Option pricing when the variance changes randomly: Theory, estimation and an
application, Journal of Financial and Quantitative Analysis 22, 419-438.
Shastri, K., and K. Tandon, 1986, On the use of European models to price American options on
foreign currency, Journal of Futures Market 6, 93-108.
Stein, J., 1989, Overreactions in the options market, Journal of Finance 44, 1011-1023.
Stoll, H., and R. Whaley, 1990, The dynamics of stock index and stock index futures returns,"
Journal of Financial and Quantitative Analysis 25, 441-468.
Whaley, R., 1986, Valuation of American futures options: Theory and empirical tests, Journal of
Finance 41, 127-150.
White, H., 1980, A heteroskedastic-consistent covariance matrix estimator and a direct test for
heteroskedasticity, Econometrica 48, 817-838.
, 1982, Maximum likelihood estimation of misspecified models, Econometrica 48, 817-838.
Wiggins, J., 1987, Options values under stochastic volatility: Theory and empirical estimates,
Journal of Financial Economics 19, 351-372.

This content downloaded from 128.36.7.248 on Sun, 29 Mar 2020 01:01:44 UTC
All use subject to https://about.jstor.org/terms