Evaluating the Predictive Accuracy of Volatility Models

Jose A. Lopez

Economic Research Department

Federal Reserve Bank of San Francisco
101 Market Street
San Francisco, CA 94705-1579
(415) 977-3894
[email protected]

Draft Date: June 30, 1999

ABSTRACT:

Statistical loss functions that generally lack economic content are commonly used for
evaluating financial volatility forecasts. In this paper, an evaluation framework based on loss
functions tailored to a users economic interests is proposed. According to these interests, the
user specifies the economic events to be forecast, the criterion with which to evaluate these
forecasts, and the subsets of the forecasts of particular interest. The volatility forecasts from a
model are then transformed into probability forecasts of the relevant events and evaluated using
the specified criteria (i.e., a probability scoring rule and calibration tests). An empirical example
using exchange rate data illustrates the framework and confirms that the choice of loss function
directly affects the forecast evaluation results.

JEL Classification: C32, C52, C53, G12

Keywords: volatility, ARCH, probability forecasts, scoring rules, exchange rates

Acknowledgments: The views expressed here are those of the author and do not necessarily reflect those of the
Federal Reserve Bank of San Francisco or the Federal Reserve System. I thank Frank Diebold, Tim Bollerslev,
Peter Christoffersen, Roberto Mariano, Lee Ohanian, Esther Ruiz and Ken West for their comments as well as
seminar participants at the Derivatives Project Meeting of the National Bureau of Economic Research, the
Econometric Society Seventh World Congress, the University of Pennsylvania, Arizona State University and the
Federal Reserve Bank of New York.
Author biography:

Jose A. Lopez is currently an economist in the Research Department at the Federal

Reserve Bank of San Francisco. He holds a Ph.D. in economics from the University of
Pennsylvania. His current research focuses on financial volatility forecasting and its applications
in the risk measurement and management systems of financial institutions.
 Evaluating the Predictive Accuracy of Volatility Models

ABSTRACT:

Statistical loss functions that generally lack economic content are commonly used for
evaluating financial volatility forecasts. In this paper, an evaluation framework based on loss
functions tailored to a users economic interests is proposed. According to these interests, the
user specifies the economic events to be forecast, the criterion with which to evaluate these
forecasts, and the subsets of the forecasts of particular interest. The volatility forecasts from a
model are then transformed into probability forecasts of the relevant events and evaluated using
the specified criteria (i.e., a probability scoring rule and calibration tests). An empirical example
using exchange rate data illustrates the framework and confirms that the choice of loss function
directly affects the forecast evaluation results.

JEL Classification: C32, C52, C53, G12

Keywords: volatility, ARCH, probability forecasts, scoring rules, exchange rates
 "Volatility forecasting is a little like predicting whether it will rain; you can be
correct in predicting the probability of rain, but still have no rain."
- Engle (1993)

I. Introduction

Although dynamics in the variance of financial time series were observed at least as early

as Mandelbrot (1963), efforts to empirically model these dynamics have only developed in the

last twenty years. GARCH models, pioneered by Engle (1982) and Bollerslev (1986), are the

volatility models most commonly used, although numerous alternatives exist. Specifically,

stochastic volatility models, which arose from the theoretical finance literature, are increasingly

employed in empirical research. For recent surveys of volatility models, see Bollerslev, Engle

and Nelson (1994) as well as Diebold and Lopez (1995).

Volatility models and their forecasts are of interest to many types of economic agents.

For example, options traders require asset volatilities to price options, and central banks or

international investors forecasting exchange rates may require interval forecasts, which are

readily derived from volatility forecasts. Given the vast number of models available, such agents

must decide which forecasts to use as well as the evaluation criterion upon which to base that

decision. As suggested by Bollerslev et al. (1994), economic loss functions that explicitly

incorporate the costs faced by volatility forecast users provide the most meaningful forecast

evaluations. Several such loss functions based on specific economic questions have been

proposed in the literature. For example, West, Edison and Cho (1993) propose a loss function

based on the expected utility from an investment decision requiring volatility forecasts. Engle et

al. (1993) propose an economic loss function based on trading profits earned in a simulated

options market; see also Engle, Kane and Noh (1996) and Noh, Engle and Kane (1994).

1
 However, since specific economic loss functions are frequently unavailable, volatility

forecast evaluation is typically conducted by minimizing a statistical loss function, such as mean

squared error (MSE). Numerous studies have used MSE to evaluate the performance of volatility

models; for example, see Taylor (1987); Akgiray (1989); Dimson and Marsh (1990); Pagan and

Schwert (1990); Lee (1991); West and Cho (1994); Kroner, Kneafsey and Claessens (1995);

Bollerslev and Ghysels (1996); and Brailsford and Faff (1996). Such forecast evaluations are

problematic since they almost exclusively use squared asset returns as a proxy for the latent

volatility process. As discussed by Andersen and Bollerslev (1997), squared returns are noisy

estimators of the actual variance dynamics and will thus limit the inference available regarding

the forecast accuracy.

In this paper, a forecast evaluation framework based on an alternative set of statistical

loss functions is proposed. Rather than selecting a simple statistical criterion as the appropriate

evaluation measure, this framework permits the forecast user to tailor their choice of criterion

according to their economic interests. Specifically, the framework is based on evaluating

probability forecasts generated from an underlying volatility model. Probability forecasts are the

object of interest for a large, general class of loss functions based on decision-theoretic problems,

such as deciding whether to take a specific action; see Granger and Pesaran (1996) as well as

Zellner and Hong (1991) for further discussion.

The user tailors the forecast evaluation by specifying three key elements of this

framework. The first element is the economic events of interest, such as a central banks interest

in whether an exchange rate will leave a specified target zone. Volatility forecasts are readily

transformed into such probability forecasts using the underlying model's assumed (or estimated)

2
distribution for the innovation term. Once generated, the forecasts are evaluated using scoring

rules, which measure the accuracy of probability forecasts with respect to whether the forecasted

events occur. Forecast users can select the scoring rule best suited to their particular interests.

The most common scoring rule (and the one used in this paper) is the quadratic probability score,

which is the analog of MSE for probability forecasts and an explicit measure of forecast

accuracy. Furthermore, forecast users can employ calibration tests to evaluate subsets of the

probability forecasts that are of particular interest, such as forecasts that lie above a certain

decision-theoretic trigger point (say, 75 percent). Calibration tests, as developed by Seillier-

Moiseiwitsch and Dawid (1993), examine the degree of equivalence between an event's predicted

and observed frequencies of occurrence within subsets of the unit interval. The specification of

these three elements forms an evaluation framework that, although based on statistical loss

functions, can be tailored to provide economically meaningful inference for a wide range of

decision-theoretic problems.

The paper is structured as follows. In Section II, the loss functions previously used for

volatility forecast evaluation are reviewed. Section III describes the three elements of the

proposed evaluation framework that are specified by the user -- the events to be forecast, the

scoring rule used to evaluate the probability forecasts, and the subsets of the forecasts of

particular interest. Section IV presents an empirical example using daily exchange rates to

illustrate the evaluation procedure. Four different sets of probability forecasts are used to

evaluate the forecasts from nine volatility models. The example provides three insights into

volatility forecast evaluation. First, the evaluations are sensitive to the loss function chosen, and

thus forecast users should consider their choice carefully. Second, the disparity between in-

3
sample and out-of-sample evaluation results clearly indicate that in-sample diagnostic tests must

be supplemented by out-of-sample analysis. Third, although not capable of providing definitive

answers on which forecasts minimize the relevant loss functions, the proposed forecast

evaluation framework provides useful and economically meaningful information for reducing the

choice set. Section V concludes and suggests directions for further research.

II. Previous Evaluations of Volatility Forecasts

Volatility models are generally expressed as

yt
t  t , t
ht zt , zt ~ i.i.d. D( 0, 1 ),

where the conditional mean t
E yt G t
1 , t
1 is the information set available at time t-1,

t is the innovation term, ht is its conditional variance and D(0,1) is a symmetric, standardized

distribution; thus, t G t
1 ~ D 0, ht . Variance dynamics are introduced into the model via
2
the specification of ht; for example, the GARCH(1,1) model sets ht
 t
1  ht
1 .

Generally, the parameters of the model are estimated over a specified in-sample period t= 1,..., T,

and volatility forecasts based on these estimates are generated over the out-of-sample period t=

T+1,...,T+T*. (Note that, in a convenient abuse of terminology, the in-sample, fitted conditional

variances are frequently referred to as in-sample forecasts).

Several in-sample diagnostics, such as analysis of standardized residuals, conditional

moment tests as per Wooldridge (1990) or the news impact curve proposed by Engle and Ng

(1993), are available for comparing volatility model specifications. However, out-of-sample

forecast accuracy provides an alternative and potentially more useful comparison; i.e., a model

4
that generates accurate forecasts must be providing a reasonable approximation of the underlying

data generating process. In the literature, volatility forecasts have been evaluated using economic

and statistical loss functions. Economic loss functions, such as the utility-based loss function

proposed by West et al. (1993) or the profit-based loss function proposed by Engle et al. (1993),

provide the most meaningful forecast evaluations since they directly incorporate the users

decision structure into the evaluation.

However, such specific economic loss functions are rarely available, and purely statistical

loss functions, such as the mean-squared error (MSE) criterion, are more commonly used to

evaluate both in-sample and out-of-sample volatility forecasts. Focusing on out-of-sample

forecasts, MSE is the average squared difference between the actual conditional variance hT+t and

the corresponding volatility forecast h . However, since hT+t is not directly observable, the
Tt
T

2 1 2 2
squared innovation Tt is used as a proxy. Thus, MSE
Tt  hTt , where
T  t1


T
hTt are the one-step-ahead forecasts based on the in-sample parameter estimates and the
t1

relevant information sets. (Note that this analysis can be extended to k-step-ahead forecasts.)

Although MSE could in some instances be the appropriate loss function for evaluating
2
conditional variance forecasts, the use of the Tt proxy is problematic. Specifically, even
2
though Tt is an unbiased estimator of hT+t, it is generally an imprecise or noisy estimator due

to its asymmetric distribution. For example, if zTt ~ N( 0, 1 ), then

2 2 2 2
E Tt G Tt
1
hTt E zTt G Tt
1
hTt since zTt ~ (1). However, since the median of a
2 2 1
(1) distribution is 0.455, Tt < h more than 50% of the time. In fact,
2 Tt
2 1 3 2 1 3
Pr Tt M hTt, hTt
Pr zTt M ,
0.2588;
2 2 2 2

5
 2
that is, even if one is willing to accept a proxy that is up to 50% different from hT+t, Tt would

fulfill this condition only 25% of the time. (See Andersen and Bollerslev (1997) for a discussion

of alternative proxies based on high-frequency data.) A further difficulty with MSE evaluation in

a heteroskedastic environment, as noted by Bollerslev et al. (1994), is that the symmetric nature

of this loss function does not sufficiently penalize nonpositive variance forecasts.

Several alternative statistical loss functions are proposed in the literature, although the

same criticisms cited above would apply to them; for example, mean absolute error (MAE),
T

1 2
MAE
Tt  hTt L. Two loss functions that penalize volatility forecasts
T t  1 L
asymmetrically are the logarithmic loss (LL) function employed by Pagan and Schwert (1990),

T

1 2 2
LL
ln Tt  ln hTt ,
T t  1

2
which penalizes inaccurate variance forecasts more heavily when Tt is low, and the

heteroskedasticity-adjusted MSE (HMSE) of Bollerslev and Ghysels (1996),

2
T
2
Tt
1
HMSE
 1 .
T t  1 hTt

Bollerslev et al. (1994) further suggest the loss function implicit in the Gaussian quasi-maximum

likelihood function often used in estimating GARCH models; that is,

T
2
Tt
1
GMLE
ln hTt  .
T t  1 hTt

Although these functions may coincide with a specific forecast users economic loss

function, they are generally devoid of economic content. Hence, economic inference based on

such forecast evaluations is limited. Below, a forecast evaluation framework based on

6
probability forecasts is proposed. This framework is also based on statistical loss functions;

however, analyzing probability forecasts generated from volatility forecasts permits the

evaluation to be tailored more closely to the users economic interests.

III. Generating and Evaluating Probability Forecasts

The proposed framework permits the user to tailor the forecast evaluation to their

particular interests in two stages. The first stage consists of transforming volatility forecasts into

probability forecasts of events of interest. (Note that this type of evaluation defines a volatility

model as a combination of both the variance dynamics specified by ht and the distribution of the

standardized residuals specified by D.) This focus on probability forecasts can be linked to a

large, general class of economic loss functions commonly used in decision-theoretic problems.

As discussed by Granger and Pesaran (1996), probability forecasts are the key input to loss

functions that gauge the economic value of taking an action, such as whether or not to purchase

an asset. In the second stage, the probability forecasts are evaluated using the statistical criteria

that most closely correspond to the users interests. The first part of this section describes how

probability forecasts are generated from volatility forecasts, and the remainder describes the

evaluation criteria employed.

3a. Transforming Volatility Forecasts into Probability Forecasts

Given that t G t
1 ~ D 0, ht , volatility forecasts can be easily transformed into

probability forecasts. The distribution D can be assumed or estimated as per Engle and

Gonzalez-Rivera (1991), and one-step-ahead probability forecasts can be easily generated by

7
integrating over this distribution. The appropriate range of integration will depend on the

economic event of interest; i.e., the user can tailor the forecast evaluation to the region of

D 0, ht of particular interest. This transformation of the volatility forecasts permits forecast

users with different (but unspecified) economic loss functions to introduce their particular

interests into the forecast evaluation, even though they will all use the same statistical tools.

Multi-step-ahead volatility forecasts are not as easily transformed into probability

forecasts. As shown by Baillie and Bollerslev (1992), variance dynamics cause the distribution

of T+k (k>1) to be a nontrivial function of T. However, a number of methods are available to

address this difficulty; for example, Baillie and Bollerslev (1992) use the Cornish-Fisher

approximation method to construct k-step-ahead confidence intervals for a GARCH(1,1) process.

Alternatively, simulation methods, as used by Geweke (1989) as well as Bollerslev and

Mikkelsen (1996), can be used to generate the desired k-step-ahead probability forecasts. Since

the forecast evaluation tools described below are appropriate for transformed volatility forecasts

of all horizons, the following discussion will focus on one-step-ahead forecasts to simplify both

notation and exposition.

The probability forecast notation is as follows. Volatility models are fit to the in-sample
T
observations yt , and Pt is the in-sample probability forecast for time t based on the
t1

parameters estimated over the entire sample and the necessary elements of T. Out-of-sample,

PT+t for t = 1,...,T* is the one-step-ahead probability forecast conditional on the parameter

estimates and Tt
1 . The subsequent discussion will focus on out-of-sample forecasts. The

event space to be examined is created by partitioning the set of all possible outcomes into N

mutually exclusive and collectively exhaustive subsets according to the forecast user's interests.

8
If N=2, a binomial event is specified, and PT+t and RT+t, an indicator variable equaling one if the

event occurs and zero otherwise, are scalar variables.

In the following discussion and empirical exercise, two general types of binomial events,

innovation and level events, will be examined. An important category of volatility forecast users

are those interested in the behavior of the innovation term Tt . This category includes, for

example, spot traders or options traders structuring hedging strategies with respect to Tt. For

such users, the economic event of interest is Tt M L, Tt, U, Tt , and the associated

probability forecast given the one-step-ahead forecast hTt is

u, Tt
L, Tt U, Tt
PTt
Pr L, Tt@Tt@U, Tt
Pr @ zTt @
f zTt dzTt,

hTt hTt l, Tt

where zT+t is the standardized innovation, f zTt is the functional form of D (0, 1) and

l, Tt, u, Tt is the standardized range of integration. Forecast users thus tailor the volatility

forecast evaluation by specifying the appropriate L, Tt, U, Tt interval.

The second type of economic events of interest are based on the behavior of the level

term yT+t. Such events would be of interest, for example, to a central bank forecasting whether an

exchange rate will remain within a target zone. In such cases, the event of interest is

yTt M Ly, Tt, Uy, Tt , and the probability forecast is

PTt
Pr Ly, Tt @ yTt @ Uy, Tt
Pr Ly, Tt  Tt @ Tt @ Uy, Tt  Tt

uy, Tt
Ly, Tt  Tt Uy, Tt  Tt

Pr @ zTt @
f zTt dzTt,

hTt hTt ly, Tt

where Tt is the forecasted conditional mean and ly, Tt, uy, Tt is the standardized range of

integration. For the central bank target zone example, Ly, Tt, Uy, Tt would equal the constant

9
interval [L,U] for all t. Granger, White and Kamstra (1989) examine interval forecasts based on
T

yT+t, in which case the endpoints Ly, Tt, Uy, Tt are set to generate a constant %
t1

confidence interval around the corresponding conditional mean forecasts; see Christoffersen

(1996) for further discussion on evaluating interval forecasts.

Once the desired probability forecasts are generated, forecast users can proceed to

evaluate them using probability scoring rules and calibration tests. These two forecast evaluation

criteria, as well as the Diebold-Mariano (1995) tests of comparative predictive accuracy, are

described in the following subsections; see Diebold and Lopez (1996) for further discussion.

3b. Probability Scoring Rules

Probability scoring rules are primarily employed in the economics literature to evaluate

business-cycle turning-point probabilities as in Diebold and Rudebusch (1989), Ghysels (1993),

and Lahiri and Wang (1994). Scoring rules measure the statistical accuracy of the forecasted

probabilities. (Note that a scoring rule should be proper; i.e., it should not cause the forecaster

to modify their actual forecasts in order to improve their expected score.) Just as with the

statistical loss functions described in Section II, a variety of proper scoring rules are available;

see Murphy and Daan (1985) as well as Winkler (1993) for further discussion. However, a key

difference between these two classes of statistical loss functions with respect to volatility

forecasts is that scoring rules are based on comparing a models forecasts to observable events,

not a proxy for the true, unobservable variance.

In practice, the forecast user should select the proper scoring rule that addresses their

particular interests most directly. In the following discussion and empirical exercise, the

10
quadratic probability score (QPS), the most common scoring rule, is used. The QPS, originally

developed by Brier (1950), is the analog of MSE for probability forecasts and thus implies a

quadratic loss function. The QPS over a forecast sample of size T* is

T

1
QPS
2 PTt  RTt ,
2

T t1

which implies that QPS M [0,2] and has a negative orientation (i.e., smaller values indicate more

accurate forecasts). This type of statistical loss function would be most relevant to economic

agents facing decision-theoretic problems that explicitly value the accuracy of the probability

forecast inputs.

3c. Comparing the Predictive Accuracy of Probability Forecasts

Scoring rules measure the accuracy of probability forecasts; hence, if the QPS for



PA, Tt T is closer to zero than the QPS for PB, Tt T , then the forecasts from model A are
t1 t1

more accurate than those from model B. However, whether this outcome is statistically

significant or an artifact of the sample at hand is unclear. Diebold and Mariano (1995) propose

several tests for determining whether the expected losses (under a general loss function) induced

by two sets of point forecasts are statistically different. These tests are readily generalized to

probability forecasts.

The null hypothesis under the loss function g is E g PA, Tt, RTt
E g PB, Tt, RTt or,

equivalently, E dTt
E g PA, Tt, RTt  g PB, Tt, RTt
0. For QPS,

dTt
2 PA, Tt  RTt 2

 2 PB, Tt  RTt 2,

where PA,T+t and PB,T+t are the probability forecasts from two volatility models. (Note that to test

11
 T

this hypothesis, dTt must be covariance stationary, a condition determined empirically.)
t1

Several statistics, both asymptotic and finite-sample, are proposed for testing this null hypothesis.

In this paper, the asymptotic mean statistic is used and defined as

d a
S
N(0,1),
~
2fd(0) / T 

where d is the sample mean of dT+t and fd(0) is the spectral density function at frequency zero

estimated using a rectangular lag window.

3d. Calibration Tests of Probability Forecasts

The calibration of probability forecasts is defined as the degree of equivalence between an

event's observed and predicted frequencies of occurrence within subsets of the unit interval. For

example, a forecaster is providing perfectly calibrated rain forecasts if it rains on 10% of the days

for which a 10% chance of rain is forecast. A simple measure of calibration is the calibration

plot, which graphs Pr(event occurs | j) against j, where j is the midpoint of one of the J

mutually exclusive and collectively exhaustive subsets of the unit interval specified by the user.

The degree to which this graph lies on the 45N line is a visual measure of calibration. Seillier-

Moiseiwitsch and Dawid (1993) present test statistics that formalize this analysis.

The null hypothesis of these tests is that the predicted frequency of occurrence for a

specified binomial event equals the observed frequency of occurrence and hence the forecasts are

well calibrated. The relevant test statistics are constructed by dividing the out-of-sample

probability forecasts into the J subsets of the unit interval relevant to the forecast user; for

example, if the deciles {[0, 0.1); [0.1, 0.2); ...; [0.9, 1]} are of interest as in Diebold and

12
 
Rudebusch (1989), then J = 10. Let j denote the midpoint of subset j, Tj the number of

forecasts in subset j, and Rj the number of observed events paired with the forecasts in subset j.

The test statistics are the subset j calibration statistics,


Rj  Tj j Rj  ej
Zj

, j
1,...,J,
 1/2 1/2
Tj j 1  j wj

and the overall calibration statistic,

R0  e0
Z0
,
1/2
w0
J J J
where R0
R j, e0
ej and w0
wj . Each of these statistics is the square root of
j  1 j  1 j  1

a goodness-of-fit statistic for a binomial event with Rj observed outcomes and ej expected
2

outcomes. Under the null hypothesis and weak conditions on the distribution of the probability

forecasts, these statistics are asymptotically normally distributed. If the null hypothesis for the

forecasts in subset j is rejected, then these forecasts do not adequately represent the observed

frequency of occurrence in that subset. The calibration tests and this flexible specification of the

J subsets of interest provide forecast users with an additional statistical tool with which to tailor

their evaluations of the transformed volatility forecasts.

In summary, a statistical framework for volatility forecast evaluation that permits the user

to tailor the evaluation to their economic interests is proposed. As discussed by Murphy (1973),

purely statistical and purely economic evaluations of forecasts represent the end points of a

continuum. By permitting the user to specify the event to be forecast, a probability scoring rule

and a partitioning of the unit interval for calibration tests, this framework permits evaluations

that incorporate elements of both types. In the following section, an empirical application

13
illustrating this framework for one-step-ahead volatility forecasts is presented. Evaluations of k-

step-ahead forecasts using this framework are reserved for future research.

IV. Evaluating Exchange Rate Volatility Forecasts

To illustrate the proposed evaluation framework, one-step-ahead volatility forecasts of

several foreign exchange rates are evaluated with respect to four different economic events of

interest. Within this empirical exercise, no one set of forecasts is found to minimize all of the

probability-based loss functions. In fact, not only does the minimizing set of forecasts vary

across exchange rates and events, it also varies across the in-sample and out-of-sample periods.

These results highlight two key features of volatility forecast evaluation (and forecast evaluation,

in general): the need for careful selection of the loss function employed and the usefulness of out-

of-sample forecasting results for model comparison and specification.

4a. Exchange Rate Data

The four exchange rates examined are the logged daily Canadian dollar (CD),

Deutschemark (DM) and Japanese yen (YN) spot exchange rates with respect to the U.S. dollar

and the U.S. dollar exchange rate with respect to the British pound (BP), as recorded by the

Federal Reserve Bank of New York, from 1980 to 1995. The in-sample period used for

parameter estimation is 1980-1993 (3516 observations), and the out-of-sample period is 1994-

1995 (499 observations).

Given established unit root results for exchange rates (as per Baillie and Bollerslev (1989)

and others) and after standard unit root testing, the log exchange rates are modeled as I(1)

14
processes; that is,

yt
 t
1  t, t t
1 ~ D 0, ht ,

where yt L 100 log s t and st is one of the spot exchange rates. The moving-average term is

included due to the presence of a small, but significant, first-order serial correlation in yt.

Table 1 presents the in-sample, least squares estimates of these conditional mean parameters,

higher-order moments of the in-sample and out-of-sample t series and the portmanteau statistics
2
for up to twentieth-order serial correlation in the t and t series. The kurtosis estimates and the

Q2(20) statistics indicate the presence of conditional heteroskedasticity in all four series.

4b. Volatility Models and Forecasts

To address the presence of variance dynamics in these series, three categories of volatility

models are used: simple models, GARCH models and stochastic volatility models. These

models are estimated over the in-sample period, and the parameter estimates are used to generate

volatility forecasts over the out-of-sample period. (The parameter estimates and relevant

standard errors are reported in the Appendix.)

Three simple volatility models are examined. The first model is the Gaussian

homoskedastic model (homo.), which ignores variance dynamics and assumes ht
2.

Although, as shown in Table 2, the squared standardized residuals from this model still display

significant conditional heteroskedasticity, it is included in the evaluation exercise as a

benchmark. As in West and Cho (1994), two autoregressive models of t are estimated using
2
ordinary least squares and an assumed conditionally normal distribution. The autoregressive t
10
2
model assuming a lag-order of 10 specifies ht
 i t
i and is denoted as AR10sq.
i1

15
(Note that volatility forecasts from this model can be negative, and in such cases, the forecasts

are set to a small positive number.) The second autoregressive model (AR10ab), as developed by

Davidian and Carroll (1987) and Schwert (1989), is based on t and is specified as

12 2

ht
 i t
i .
2 i1

From the second category, five models are specified and estimated via maximum

likelihood methods. The first three models share the same GARCH(1,1) specification (i.e.,
2
ht
 t  ht
1 ), but use different distributional assumptions. Specifically, the three

assumed, symmetric forms of D are the normal distribution, the t-distribution and the generalized

exponential distribution (GED); all of which are special cases of the generalized t-distribution

discussed in McDonald and Newey (1988). Both the t- and GED-distributions permit fatter tails

than the normal distribution, an empirical regularity often found in financial time series; see

Figure 1 for a comparison of these distributions based on the estimated shape parameters for the

BP series. These models are denoted as GARCH, GARCH-t and GARCH-g, respectively.

The fourth and fifth models are derived from the calibrated, exponential-smoothing

model currently used in the volatility forecasting methodology proposed in the J.P. Morgans

RiskmetricsTM system; see J.P. Morgan (1996). The assumed variance dynamics are

Q
2 2
h t
( 1  ) i t
i
( 1  ) t
1  ht
1;
i1

note that this calibrated model is equivalent to the IGARCH model of Engle and Bollerslev

(1986) with the parameter set to zero. The parameter is calibrated according to the degree of

persistence desired. For this exercise, the commonly-used value of =0.94 is specified, and the

16
conditional mean parameters are those estimated using the homoskedastic model. The two

exponential-smoothing models used in this exercise differ only in their assumed distribution D;

the e.smooth-N model uses the standard normal, and the e.smooth-t model uses the t-distribution

with the same shape parameter as estimated for the GARCH-t model.

The final model is the AR(1) stochastic volatility model (s.v.) used by Harvey, Ruiz and

Shephard (1994); that is, after removing the conditional mean based on least squares parameter

estimates,

t
exp t / 2 t, t ~ N( 0, 1 )

2
t
t
1  t, t ~ N( 0, ) ,

where t t. Thus, the innovation term is subject to two independent shocks. Estimation of

the model is conducted using the Kalman filter on the measurement equation
2 2
log t
t  logt and the transition equation t
t
1  t. To use quasi-maximum
2
likelihood methods, the measurement equation is rewritten as logt
 t  t, where
2

E log t and t is assumed to be distributed N 0, 2/2 . (Note that alternative

estimation techniques have been proposed; see Jacquier, Polson and Rossi, 1994). The log-

likelihood function to be maximized is

2 2
T T
T T logt   at
ln L ; 1, ..., T
 log(2)  log Gt  ,
2 2 t1 t1 Gt

where at and Gt are the Kalman filter estimates of t and its conditional variance, respectively,
2
and
, .

One-step-ahead volatility forecasts are easily obtained from the simple and GARCH(1,1)

models, and the corresponding probability forecasts are derived via integration. One-step-ahead

17
volatility forecasts from the s.v. model are obtained using the Kalman filter, but converting them

into probability forecasts is not as straightforward. The standardized innovation term

t /2 2/8
zt
e t is the product of a standard normal and a lognormal distribution with mean e
2/4 2/4
and variance e e  1 . Integration over the distribution of this random variable is both

difficult and computationally intensive. To alleviate this problem, this distribution is empirically

approximated by simulating 10,000 draws from it and creating 0.5% quantiles. The probability

that zt M l t, ut is approximated as Pt
F ut  F lt , where F is the empirical cumulative

distribution function of zt. (The magnitude of this approximation error is small and can be made

arbitrarily small by increasing the number of simulated draws.)

Standard diagnostics of the in-sample and out-of-sample standardized residuals are

presented in Table 2. In-sample, almost all of the volatility models, except for the homo. and s.v.

models, consistently remove the serial correlation in the squared standardized residuals in all four

series. Out-of-sample, only the homo. model fails to consistently do so. These results indicate

that all of the volatility models, with the possible exception of the s.v. model, may be appropriate

specifications of the variance dynamics. In fact, as shown in Table 3 for the BP series, the in-

sample and out-of-sample volatility forecasts are highly correlated, especially within the GARCH

class of models. These results are similar for the other three series.

Before evaluating the generated probability forecasts, Tables 4 and 5 report the in-sample

and out-of-sample forecast evaluation results using the statistical loss functions from Section II.

The in-sample results clearly indicate that the loss function chosen directly impacts the forecast

evaluation results. MSE in minimized for all four series by the AR10sq model, as expected.

However, as indicated by the S test results, it does not do so at the 5% level of statistical

18
significance. The MAE and LL loss functions are minimized by the s.v. model significantly; i.e.,

most of the alternative models forecasts fail the one-on-one S test at the 5% level. The results

for HMSE indicate no clear pattern, and the GMLE loss function is minimized, but not

significantly, in all cases by the GARCH model.

The out-of-sample forecast evaluation results vary considerably in comparison to the in-

sample results. No one set of forecasts minimize a specific loss function across all four series,

and few of the forecasts that minimize a loss function do so significantly. (West and Cho (1995)

find qualitatively similar results for MSE using different statistical tests and various forecast

horizons.) Of the twenty forecast evaluations reported in Table 5, in only eight cases is the

minimizing set of forecasts the same across the in-sample and out-of-sample periods, and six of

these cases are the MAE and LL with respect to the s.v. model. These results indicate that

forecast evaluation based on such loss functions varies considerably; hence, the forecast user

should do so only after careful consideration of how these loss functions reflect their particular

economic interests.

4c. Specifying Economic Events

In the next subsection, the four economic events used for evaluating the volatility

forecasts from the nine models are described. In order to simplify this sample evaluation, only

the event to be forecast is varied. The QPS is the probability scoring rule used throughout, and

the subsets of interest are the deciles [0, 0.1), [0.1, 0.2),...,[0.9, 1].

The first two events are based on the behavior of yt, the one-day return. The first event

is yt M 0.5 , 1 , which is a relatively high probability event, as shown in Table 6. (Note that,

19
after adjusting for the specified conditional mean, this event is t M 0.5 t
1, 1 t
1 . )

Such events are of interest, for example, to spot traders deciding how to structure their market

positions or to options traders holding out-of-the-money positions. The second event is related to

the value-at-risk estimates now commonly used in managing the market risk of trading

portfolios; see Lopez (1997) for further discussion. In such cases, traders generally wish to know

a specified lower-tail quantile of their portfolio returns distribution in order to manage their

market risk exposure. For this exercise, the lower five percent quantile of the empirical, in-

sample distribution of yt is specified as the threshold loss (TL) of interest to the forecast users,

and the event to be forecast is yt < TL. As shown in Table 6, the observed in-sample

frequency of occurrence of this event is constructed to be five percent, and the out-of-sample

frequency is relatively close to five percent for all four currencies.

The last two events focus on the behavior of the level term yt. The third event is a %

move in yt; that is, yt M (1/100) yt
1 , (1/100) yt
1 . Such an event would be of interest, for

example, to a portfolio manager deciding when to transact in a foreign-currency denominated

asset. The parameter is set at 2, except for the case of YN in which it is set to 0.2 (due to the

high level of yt). The last event is based on the idea of a central bank forecasting whether an

exchange rate will remain within a specified target zone; i.e., the event of interest is

yt M L , U . In this evaluation exercise, L and U are set as 5% of yT, the last in-sample

observation. The observed frequencies of occurrence for event 3 are relatively high, while event

4 occurs only about 15 percent of the time, except for the YN series which is much higher. For

the purpose of illustration, Figure 2 plots the out-of-sample probability forecasts from the

GARCH-g model for the BP series for all four events of interest.

20
 4d. Forecast Evaluation Results

Table 7 presents the in-sample QPS results for the four exchange rates and the four

specified events. The minimum QPS value for each column in each panel is underlined. An

interesting result is that the GARCH-g model minimizes the QPS value in half of the 16 cases,

although the significance of this result is low given the poor S test results. That is, many of

models generate QPS values that are greater than that of the GARCH-g model, but not

significantly so. Overall, these results suggest that the GARCH-g model seems to characterize

well the variance dynamics and the distribution of zt for these series. The fitted GED

distributions fit these series well both in the center of the distribution, as per events 1 and 3, and

in the tails, as per event 2.

Turning to the out-of-sample QPS results presented in Table 8, the results are less

conclusive. Overall, only three matches with the in-sample results occur. The GARCH-g model

generates the minimum QPS values in only four cases, and this result (as well as all other

minimum QPS values) is not generally significant at the 5% level. (Although not shown, sub-

sample results with the out-of-sample period divided in half provide similar results.) Thus, the

out-of-sample results indicate, much like the results in Table 5, that the chosen volatility models

provide forecasts that are practically indistinguishable under these event-based, loss functions.

However, by focusing on probability forecasts, calibration tests can be used as an

additional tool for evaluating the competing sets of out-of-sample volatility forecasts. The

calibration test results are presented in the companion columns of Table 8. Probability forecasts

that are well calibrated at the five percent significance level are denoted with + symbol.

Clearly, in many cases, the calibration test reduces the number of models under consideration by

21
rejecting the probability forecasts that are not calibrated over the user-specified subsets of the

unit interval. (Once again, subsample results are consistent with the full-sample results.) For

example, for BP with respect to event 1, forecasts from the GARCH-g model generate the lowest

QPS value; the S test eliminates three competing models since they are significantly greater than

this value; and the calibration test eliminates three additional model. Hence, in this case,

although the forecasts from the GARCH-g, e.smooth-N and s.v. models are indistinguishable

using this statistical evaluation framework, the number of models to consider has been reduced

from nine to three. Although, as shown in Table 8, this evaluation procedure does not

distinguish this well between competing volatility forecasts in all cases, it can be used to assist

the forecast user in limiting the choice of volatility forecasts to consider.

In conclusion, this sample forecast evaluation provides three clear results. First, the

different model rankings presented in Tables 4, 5, 7 and 8 clearly indicate that forecast

evaluations are directly affected by the loss function selected. In the absence of economic loss

functions, forecast users should carefully select the statistical loss functions that best match their

interests. Second, the disparity between the in-sample and out-of-sample results indicate that

tests of model fit and model specification should not be considered sufficient analysis. Out-of-

sample forecast evaluation provides additional, useful information for volatility model

specification and model selection. Third, the proposed framework for evaluating generated

probability forecasts can generally reduce the number of competing volatility forecasts under

consideration, a result that should be of particular interest to forecast users who must select

among volatility models.

22
V. Conclusions

The forecast evaluations most meaningful to agents facing the issue of volatility model

selection are those conducted under economically meaningful loss functions. In the general

absence of such loss functions, researchers rely on statistical loss functions, such as MSE.

Although reasonable for point forecasts, MSE and related statistical loss functions are

problematic for volatility forecasts. In addition, since model rankings are highly dependent on

the loss function chosen and these loss functions are generally devoid of economic content, the

results from such analysis is of questionable worth.

In this paper, a framework for volatility forecast evaluation based on statistical tools that

can be tailored to the economic interests of the forecast user is proposed. Specifically, the

forecasts from a volatility model are transformed (via integration or simulation) into probability

forecasts of events of interest. These probability forecasts are then evaluated using a proper

scoring rule selected by the user and calibration tests over the subsets of the unit interval that are

of interest to the user.

The empirical exercise in Section IV clearly indicates that the loss function directly

influences the forecast evaluation results. Thus, the use of MSE as the standard loss function for

volatility forecast evaluation must be replaced with a more thoughtful selection. Yet, even under

the appropriate loss function, forecast evaluations may provide comparative results that are not

significantly different, as shown by the S test results. (In fact, such results for one-step-ahead

volatility forecasts may not be surprising in light of Nelson (1992) as well as Nelson and Foster

(1994), who find that short-term volatility forecasting is robust to a variety of misspecifications.)

An advantage of the proposed framework is that calibration tests provide further analysis of the

23
transformed volatility forecasts using criteria tailored to the users interests.

This evaluation framework introduces a number of questions that require further analysis.

Most immediately, the properties of the transformed volatility forecasts must be more clearly

delineated. For example, further research is required to determine whether optimal forecasts

under the probability-based loss functions are available. Secondly, the properties of the QPS and

other scoring rules must be examined in light of West (1996); that is, uncertainty due to

parameter estimation could be incorporated into this analysis. Other avenues for research are

evaluating volatility forecasts from other models (i.e., different specifications for ht and D) and

for other financial time series under different loss functions. A systematic exploration of the

theoretical and empirical aspects of the proposed evaluation framework should permit the

formulation of more useful volatility model specifications and volatility forecasts.

24
 References

Akgiray, V., 1989. Conditional Heteroskedasticity in Time Series of Stock Returns: Evidence
and Forecasts, Journal of Business, 62, 55-80.

Andersen, T.G. and Bollerslev, T., 1997. Answering the Critics: Yes, ARCH Models Do
Provide Good Volatility Forecasts, Working Paper #227, Department of Finance,
Kellogg Graduate School of Management, Northwestern University.

Baillie, R.T. and Bollerslev, T., 1989. The Message in Daily Exchange Rates: A Conditional
Variance Tale, Journal of Business and Economic Statistics, 7, 297-305.

Baillie, R.T, and Bollerslev, T., 1992. Prediction in Dynamic Models with Time-Dependent
Conditional Variances, Journal of Econometrics, 52, 91-113.

Bollerslev, T., 1986. Generalized Autoregressive Conditional Heteroskedasticity, Journal of

Econometrics, 31, 307-327.

Bollerslev, T., Engle, R.F. and Nelson, D.B., 1994. ARCH Models, in Engle, R.F. and
McFadden, D., eds. The Handbook of Econometrics, Volume 4, 2959-3038. Amsterdam:
North-Holland.

Bollerslev, T. and Ghysels, E., 1996. Periodic Autoregressive Conditional Heteroskedasticity,

Journal of Business and Economic Statistics, 14, 139-157.

Bollerslev, T. and Mikkelsen, H.O., 1996. Modeling and Pricing Long Memory in Stock
Market Volatility, Journal of Econometrics, 73, 151-184.

Bollerslev, T. and Wooldridge, J.M. (1992), Quasi-Maximum Likelihood Estimation and

Inference in Dynamic Models with Time-Varying Covariances, Econometric Reviews,
11, 143-179.

Brailsford, T.J. and Faff, R.W., 1996. An Evaluation of Volatility Forecasting Techniques,
Journal of Banking and Finance, 20, 419-438.

Brier, G.W., 1950. Verification of Forecasts Expressed in Terms of Probability, Monthly

Weather Review, 75, 1-3.

Christoffersen, P.F., 1998. Evaluating Interval Forecasts, International Economic Review, 39,
841-862.

Davidian, M. and Carroll, R.J., 1989. Variance Function Estimation, Journal of the American
Statistical Association, 82, 1079-1091.

25
Diebold, F.X. and Lopez, J.A., 1995. Modeling Volatility Dynamics, in Hoover, K., ed.,
Macroeconometrics: Developments, Tensions and Prospects, 427-466. Boston: Kluwer
Academic Press.

Diebold, F.X. and Lopez, J.A., 1996. Forecast Evaluation and Combination, in Maddala, G.S.
and Rao, C.R., eds., Handbook of Statistics, Volume 14: Statistical Methods in Finance,
241-268. Amsterdam: North-Holland.

Diebold, F.X. and Mariano, R., 1995. Comparing Predictive Accuracy, Journal of Business
and Economic Statistics, 13, 253-264.

Diebold, F.X. and Rudebusch, G.D., 1989. Scoring the Leading Indicators, Journal of
Business, 62, 369-391.

Dimson, E. and Marsh, P., 1990. Volatility Forecasting without Data-Snooping, Journal of
Banking and Finance, 14, 399-421.

Dunsmuir, W., 1979. A Central Limit Theorem for Parameter Estimation in Stationary Vector
Time Series and its Application to Models for a Signal Observed with Noise, The Annals
of Statistics, 7, 490-506.

Engle, R.F., 1982. Autoregressive Conditional Heteroskedasticity with Estimates of the

Variance of U.K. Inflation, Econometrica, 50, 987-1008.

Engle, R.F., 1993. Statistical Models for Financial Volatility, Financial Analysts Journal, 49,
72-78.

Engle, R.F. and Bollerslev, T., 1986. Modeling the Persistence of Conditional Variances,
Econometric Reviews, 5, 1-50.

Engle, R.F. and Gonzalez-Rivera, G., 1991. Semiparametric ARCH Models, Journal of
Business and Economic Statistics, 9, 345-359.

Engle, R.F., Hong, C.-H., Kane, A. and Noh, J., 1993. Arbitrage Valuation of Variance
Forecasts with Simulated Options, in Chance, D.M. and Tripp, R.R., eds., Advances in
Futures and Options Research, 393-415. Greenwich, CT: JIA Press.

Engle, R.F., Kane, A. and Noh, J., 1996. Index-Option Pricing with Stochastic Volatility and
the Value of Accurate Variance Forecasts, Review of Derivatives Research, 1, 139-158.

Engle, R.F. and Ng, V., 1993. Measuring and Testing the Impact of News on Volatility,
Journal of Finance, 48, 1749-1778.

26
Geweke, J., 1989. Exact Predictive Densities in Linear Models with ARCH Disturbances,
Journal of Econometrics, 44, 307-325.

Ghysels, E., 1993. On Scoring Asymmetric Periodic Probability Models of Turning-Point

Forecasts, Journal of Forecasting, 12, 227-238.

Granger, C.W.J. and Pesaran, M.H., 1996. A Decision Theoretic Approach to Forecast
Evaluation, Working Paper #96-23, University of California, San Diego.

Granger, C.W.J., White, H. and Kamstra, M., 1989. Interval Forecasting: An Analysis Based
Upon ARCH-Quantile Estimators, Journal of Econometrics, 40, 87-96.

Harvey, A.C., Ruiz, E. and Shepard, N., 1994. Multivariate Stochastic Variance Models,
Review of Economic Studies, 61, 247-264.

Jacquier, E., Polson, N.G. and Rossi, P.E., 1994. Bayesian Analysis of Stochastic Volatility
Models, Journal of Business and Economic Statistics, 12, 371-389.

J.P. Morgan, 1996. J.P. Morgan/Reuters RiskmetricsTM - Technical Document. New York: J.P.
Morgan. (http://www.jpmorgan.com/RiskManagement/RiskMetrics/pubs.html).

Kroner, K.F., Kneafsey, K.P. and Claessens, S., 1995. Forecasting Volatility in Commodity
Markets, Journal of Forecasting, 14, 77-96.

Lahiri, K. and Wang, J.G., 1994. Predicting Cyclical Turning Points with Leading Index in a
Markov Switching Model, Journal of Forecasting, 13, 245-263.

Lee, K.Y., 1991. Are the GARCH Models Best in Out-of-Sample Performance?, Economic
Letters, 37, 305-308.

Lopez, J.A., 1997. Regulatory Evaluation of Value-at-Risk Models, Federal Reserve Bank of
New York Staff Report #33.

Mandelbrot, B., 1963. The Variation of Certain Speculative Prices, Journal of Business, 36,
394-419.

McDonald, J.B. and Newey, W.K., 1988. Partially Adaptive Estimation of Regression Models
via the Generalized T-Distribution, Econometric Theory, 4, 428-457.

Murphy, A.H., 1973. A New Vector Partition of the Probability Score, Journal of Applied
Meteorology, 12, 595-600.

Murphy, A.H. and Daan, H., 1985. Forecast Evaluation in Murphy, A.H. and Katz, R.W., eds.,

27
 Probability, Statistics and Decision Making in the Atmospheric Sciences, 379-437.
Boulder, Colorado: Westview Press.

Nelson, D.B., 1992. Filtering and Forecasting with Misspecified ARCH Models: I, Journal of
Econometrics, 52, 61-90.

Nelson, D.B., and Foster, D.P., 1994. Asymptotic Filtering Theory for Univariate ARCH
Models, Econometrica, 62, 1-41.

Noh, J., Engle, R.F., and Kane, A., 1994. Forecasting Volatility and Option Prices of the S&P
500 Index, Journal of Derivatives, 2, 17-30.

Pagan, A.R. and Schwert, G.W., 1990. Alternative Models for Conditional Stock Volatility,
Journal of Econometrics , 45, 267-290.

Schwert, G.W., 1989. Why Does Stock Market Volatility Change over Time?, Journal of
Finance, 44, 1115-1154.

Seillier-Moiseiwitsch, F., and Dawid, A.P., 1993. On Testing the Validity of Sequential
Probability Forecasts, Journal of the American Statistical Association, 88, 355-359.

Taylor, S.J., 1987. Forecasting the Volatility of Currency Exchange Rates, International
Journal of Forecasting, 3, 159-170.

West, K.D., 1996. Asymptotic Inference about Predictive Ability, Econometrica, 64, 1067-
1084.

West, K.D. and Cho, D., 1994. The Predictive Accuracy of Several Models of Exchange Rate
Volatility, Journal of Econometrics, 69, 367-391.

West, K.D., Edison, H.J. and Cho, D., 1993. A Utility-Based Comparison of Some Models of
Exchange Rate Volatility, Journal of International Economics, 35, 23-45.

Winkler, R.L., 1993. Evaluating Probabilities: Asymmetric Scoring Rules, Management

Science, 4, 1395-1405.

Wooldridge, J.M., 1990. A Unified Approach to Robust Regression-Based Specification

Tests, Econometric Theory, 6, 17-43.

Zellner, A. and Hong, C., 1991. "Bayesian Methods for Forecasting Turning Points in Economic
Time Series: Sensitivity of Forecasts to Asymmetry of Loss Structures," in Lahiri, H. and
Moore, G., eds., Leading Economic Indicators: New Approaches and Forecasting
Records, 129-140. Cambridge, U.K.: Cambridge University Press.

28
Table 1. Conditional Mean Estimates and Residual Analysis for the First-Differenced
Exchange Rate Series

yt
 t
1  t

t G t
1 ~ N( 0, )

BP CD DM YN
Conditional Mean Parameters
0.0115 0.0004 0.0002 -0.0218
(0.0131) (0.0047) (0.0126) (0.0114)

0.0814* 0.0496* 0.0365* 0.0331*

(0.0168) (0.0166) (0.0169) (0.0169)

In-sample Moments of t
variance 0.5170 0.0701 0.5203 0.4256

kurtosis 5.3181* 6.3905* 4.7579* 5.4200*

(0.0826) (0.0826) (0.0826) (0.0826)

Q(20) 26.550 14.158 27.925 28.220

Q2(20) 577.26* 522.10* 391.10* 225.45*

Out-of-sample Moments of t
variance 0.2377 0.0967 0.4493 0.5390

kurtosis 5.0975* 5.9554* 4.8715* 6.3198*

(0.2193) (0.2193) (0.2193) (0.2193)

Q(20) 61.776* 25.474 22.573 23.958

Q2(20) 63.408* 37.965* 106.534* 45.200*

Note: The four exchange rate series are the logged daily spot rates from 1980 to 1995. The in-sample period is
1980-1993 (3516 observations), and the out-of-sample period is 1994-1995 (499 observations). The conditional
mean parameters of the in-sample, first-differenced series are estimated using least squares. Asymptotic standard
errors are listed in parentheses. The test for excess kurtosis is relative to the standard normal distribution; this
statistic is asymptotically distributed N(0,24/T). The portmanteau statistics for up to 20th-order serial correlation,
2
Q(20) and Q2(20), are for the t series and the t series, respectively. * indicates significance at the 5% level.

29
Table 2. Standard Diagnostic Tests of the In-Sample and Out-of-Sample
Standardized Residuals

In-Sample Out-of-Sample

Kurtosis Q(20) Q2(20) Kurtosis Q(20) Q2(20)

Panel A. BP
homo. 5.32 26.62 581.68* 5.13 62.34* 63.42*
AR10sq 4.58 22.70 30.42 4.75 38.36* 20.77
AR10ab 4.51 21.14 21.78 5.13 31.51* 15.66
GARCH 4.43 22.94 17.75 5.00 30.50 20.40
GARCH-t 4.45 27.03 16.87 5.05 26.03 20.30
GARCH-g 4.45 28.98 16.87 5.05 25.21 20.45
e.smooth 4.50 23.91 16.38 5.63 21.57 15.62
s.v. 4.47 22.47 35.69* 5.58 30.19 28.49

Panel B. CD
homo. 6.39 14.29 524.34* 5.96 25.47 37.96*
AR10sq 5.58 14.81 36.20* 5.52 22.95 20.82
AR10ab 5.83 14.52 21.85 6.00 23.89 20.27
GARCH 6.11 16.07 10.79 5.93 23.48 18.39
GARCH-t 6.30 17.25 13.58 5.89 23.73 17.69
GARCH-g 6.20 17.09 12.20 5.91 23.80 18.00
e.smooth 6.77 18.12 44.59* 5.79 22.12 16.38
s.v. 6.77 16.28 176.98* 6.16 21.85 21.53

Panel C. DM
homo. 4.76 28.08 393.50* 4.86 22.83 106.54*
AR10sq 4.21 30.79 34.85* 4.51 18.56 37.97*
AR10ab 4.21 31.08 24.57 4.59 17.83 33.22*
GARCH 4.12 35.65* 14.85 4.38 16.21 28.74
GARCH-t 4.13 36.23* 13.88 4.40 16.20 28.38
GARCH-g 4.12 37.08* 14.17 4.41 16.41 28.74
e.smooth 4.34 42.19* 14.41 4.55 15.90 24.63
s.v. 4.30 35.24 52.01* 4.43 16.38 35.64*

Panel D. YN
homo. 5.42 28.09 226.12* 6.32 23.93 45.19*
AR10sq 5.27 28.64 35.09* 6.35 22.67 21.83
AR10ab 5.93 29.48 39.37* 6.09 24.41 21.86
GARCH 5.51 29.84 33.00* 6.21 23.35 13.45
GARCH-t 5.87 33.34* 31.30 6.40 25.12 12.90
GARCH-g 5.75 35.97* 31.18 6.35 25.21 13.30
e.smooth 7.26 30.44 27.84 6.81 25.40 12.54
s.v. 6.84 34.75* 51.38* 5.99 29.15 29.49

Note: The first column in each panel presents the kurtosis of the in-sample and out-of-sample standardized
residuals. The columns labeled Q(20) and Q2(20) are the portmanteau statistics for these standardized residuals and
their squares. * denotes significance at the five percent level for a 2(20) distribution.

30
Table 3. Correlations between the Volatility Forecasts for BP

In-sample
AR10sq AR10ab GARCH GARCH-t GARCH-g e.smooth s.v.

AR10sq 1.00

AR10ab 0.94 1.00

GARCH 0.86 0.84 1.00

GARCH-t 0.87 0.85 0.99 1.00

GARCH-g 0.87 0.85 0.99 0.99 1.00

e.smooth 0.84 0.83 0.99 0.99 0.99 1.00

s.v. 0.69 0.78 0.84 0.83 0.83 0.85 1.00

Out-of-sample
AR10sq AR10ab GARCH GARCH-t GARCH-g e.smooth s.v.

AR10sq 1.00

AR10ab 0.96 1.00

GARCH 0.84 0.84 1.00

GARCH-t 0.86 0.85 0.99 1.00

GARCH-g 0.85 0.85 0.99 1.00 1.00

e.smooth 0.82 0.82 0.99 0.99 0.99 1.00

s.v. 0.62 0.71 0.82 0.81 0.81 0.84 1.00

31
Table 4. In-Sample Forecast Evaluation Results for Statistical Loss Functions

MSE MAE LL HMSE GMLE

Panel A. BP
homo. 1.1539 0.5665* 8.6182* 4.3204 0.3396*
AR10sq 1.0935 0.5473* 8.2178* 3.4809 0.2542
AR10ab 1.0996 0.5249* 7.8512* 4.3371* 0.2571
GARCH 1.0944 0.5464* 7.9304 3.4159* 0.2416
GARCH-t 1.1106 0.5604* 7.8821 3.1699 0.2437
GARCH-g 1.1127 0.5495* 7.7035 3.4653* 0.2426
e.smooth 1.0973 0.5458* 7.9064* 4.1656* 0.2572*
s.v. 1.1188 0.5047 7.4904 5.9245* 0.3004*

Panel B. CD
homo. 0.0264 0.0794* 8.5695* 5.3887 -1.6586
AR10sq 0.0248 0.0749* 7.9613* 4.2311 -1.7953
AR10ab 0.0249 0.0713* 7.3964* 5.9917 -1.8100
GARCH 0.0255 0.0770* 7.4381 5.1159 -1.8246
GARCH-t 0.0261 0.0790* 7.4158 5.1547 -1.8184
GARCH-g 0.0256 0.0769* 7.6094 5.3606 -1.8236
e.smooth 0.0254 0.0747 7.4049 7.3534 -1.7894
s.v. 0.0256 0.0696 7.1285 10.0064 -1.7339

Panel C. DM
homo. 1.0168 0.5706 7.4755* 3.7563 0.3464*
AR10sq 0.9750 0.5536* 7.1299* 3.0720 0.2683
AR10ab 0.9761 0.5330 6.7509 3.8699* 0.2644
GARCH 0.9772 0.5570* 7.0919* 3.0957* 0.2582
GARCH-t 0.9824 0.5694* 7.3830* 2.8496 0.2604
GARCH-g 0.9813 0.5574* 7.3371 3.1214* 0.2585
e.smooth 0.9839 0.5552* 6.8452 3.9607* 0.2807
s.v. 0.9861 0.5212 6.5459 4.7767* 0.2857*

Panel D. YN
homo. 0.8004 0.4852* 8.8365* 4.4207 0.1452*
AR10sq 0.7746 0.4756* 8.5198* 4.2530 0.0885
AR10ab 0.7823 0.4451* 7.9554* 7.1143* 0.1140
GARCH 0.7841 0.4763* 8.3220* 4.5366 0.0861
GARCH-t 0.7994 0.4741* 8.3996 5.4410 0.0923
GARCH-g 0.8020 0.4801* 8.4988 4.9428 0.0896
e.smooth 0.7892 0.4771* 8.2154* 8.0179 0.1279*
s.v. 0.8045 0.4228 7.4526 13.5137* 0.2282*

Note: The columns represent the values of the statistical loss functions from Section II for the models in-sample,
fitted conditional variances. The minimum value within each column of each panel is underlined. * indicates that
this particular set of in-sample forecasts rejects the null hypothesis of equality with the minimum value using the S
test at the five percent significance level.

32
Table 5. Out-of-Sample Forecast Evaluation Results for Statistical Loss Functions

MSE MAE LL HMSE GMLE

Panel A. BP
homo. 0.3089* 0.4553* 11.8837* 1.1567 -0.1994*
* *
AR10sq 0.2423 0.3466 10.1319 1.5552 -0.3888
AR10ab 0.2318 0.3034* 9.1532* 2.5910 -0.4523
GARCH 0.2275 0.2999* 8.9282 2.4515 -0.4611
GARCH-t 0.2290 0.3067* 8.8628 2.3033 -0.4582*
*
GARCH-g 0.2263 0.2981 8.4016 2.4919 -0.4701
e.smooth 0.2257 0.2650 8.0144 5.8356 -0.4424
s.v. 0.2299 0.2672 8.2337 5.8601 -0.4099

Panel B. CD
homo. 0.0470 0.0953 7.7033 9.5930 -1.2779
AR10sq 0.0465 0.1010 7.8258 7.3429 -1.3103
AR10ab 0.0469 0.0979 7.5533 11.2557 -1.2133
GARCH 0.0478 0.1123 8.1170 6.9315 -1.2686
GARCH-t 0.0482 0.1146 8.9795 6.1234 -1.2806
GARCH-g 0.0480 0.1125 8.2514 6.7614 -1.2715
e.smooth 0.0467 0.1085 8.3924 6.0445 -1.3083
s.v. 0.0478 0.0943 7.3589 15.9774 -1.0905

Panel C. DM
homo. 0.7853 0.5343 8.0663 2.9008 0.2130
AR10sq 0.7430 0.4950* 7.4763* 2.8007 0.1266
*
AR10ab 0.7470 0.4707 7.0276 3.9370 0.1347
GARCH 0.7402 0.4862* 7.2766 3.0493 0.1151
GARCH-t 0.7451 0.4981* 7.1669 2.8330 0.1198
GARCH-g 0.7452 0.4867* 7.0650 3.0989 0.1174
e.smooth 0.7453 0.4778 6.9420 4.3004 0.1378
s.v. 0.7527 0.4516 6.8011 4.9428 0.1581

Panel D. YN
homo. 1.5456 0.5595 8.2659 8.5368 0.4073
AR10sq 1.5322 0.5701* 8.2178* 7.5746 0.3488
AR10ab 1.5265 0.5373 7.7350* 9.6895 0.3726
GARCH 1.4995 0.5750* 7.8797 6.3216 0.2922
GARCH-t 1.5127 0.5806* 7.7722 6.8564 0.2938
GARCH-g 1.5192 0.5847* 7.8844 6.4474 0.2921
e.smooth 1.5034 0.6009 8.2284 7.7505 0.3138
s.v. 1.5571 0.5175 7.3176 14.9116 0.5133

Note: The columns represent the values of the statistical loss functions from Section II for the models out-of-
sample volatility forecasts. The minimum value within each column of each panel is underlined. * indicates that this
particular set of in-sample forecasts rejects the null hypothesis of equality with the minimum value using the S test
at the five percent significance level.

33
Table 6. In-Sample and Out-of-Sample Observed Event Frequencies

Number of in-sample observations: 3515

Number of out-of-sample observations: 499

Event 1 Event 2 Event 3 Event 4

Panel A. BP
In-sample 72.4% 5.0% 83.1% 10.6%

Out-of-sample 84.4% 1.6% 92.8% 19.0%

Panel B. CD
In-sample 96.8% 5.0% 89.0% 12.5%

Out-of-sample 94.4% 6.6% 94.8% 12.4%

Panel C. DM
In-sample 72.7% 5.0% 91.8% 13.2%

Out-of-sample 72.9% 4.2% 79.4% 12.6%

Panel D. YN
In-sample 77.4% 5.0% 89.4% a 36.5%

Out-of-sample 72.1% 4.8% 84.6% a 83.8%

a
The parameter E is set to 0.2, instead of 2 as for the series, due to the high level of the Japanese yen.

Note: Event 1 refers to the innovation event yt M
0.5 , 1 or, equivalently, t M
0.5

t1, 1

t1 .
Event 2 refers to the innovation event yt < TL or, equivalently, t < TL

t1. TL is the lower five percent
quantile of the in-sample, empirical distribution of yt. Event 3 refers to the level event
yt M (1
/100) yt1 , (1/100) yt1 , where E is set at 2 for BP, CD and DM and at 0.2 for YN. Event 4 refers to
the level event yt M L , U , where L and U are 5% of the last in-sample observation yT.

34
Table 7. In-sample QPS Results

Event 1 Event 2 Event 3 Event 4

Panel A. BP
homo. 0.2024* 0.0474 0.1183 0.0178
AR10sq 0.1945 0.0465 0.1150 0.0177
AR10ab 0.1934 0.0465 0.1146 0.0177
GARCH 0.1946 0.0463 0.1146 0.0177
GARCH-t 0.1998* 0.0477* 0.1194* 0.0180
GARCH-g 0.1934 0.0463 0.1140 0.0177
e.smooth-N 0.1953 0.0464 0.1154 0.0178
e.smooth-t 0.1978* 0.0473 0.1184* 0.0179
s.v. 0.1961 0.0464 0.1150 0.0177

Panel B. CD
homo. 0.0314 0.0476* 0.0982 0.0089
AR10sq 0.0302 0.0458 0.0910 0.0090
AR10ab 0.0303 0.0455 0.0902 0.0091
GARCH 0.0305 0.0460 0.0909 0.0090
GARCH-t 0.0324* 0.0478* 0.0958* 0.0090
GARCH-g 0.0304 0.0457 0.0899 0.0090
e.smooth-N 0.0305 0.0460 0.0917 0.0090
e.smooth-t 0.0315* 0.0471* 0.0944* 0.0089
s.v. 0.0307 0.0457 0.0915 0.0090

Panel C. DM
homo. 0.2014* 0.0473 0.0701 0.0212
AR10sq 0.1938 0.0465 0.0682 0.0210
AR10ab 0.1921 0.0467 0.0683 0.0209
GARCH 0.1926 0.0466 0.0686 0.0209
GARCH-t 0.1982* 0.0473* 0.0719* 0.0212
GARCH-g 0.1912 0.0465 0.0683 0.0208
e.smooth-N 0.1929 0.0468* 0.0692 0.0209
e.smooth-t 0.1962* 0.0473* 0.0715* 0.0211
s.v. 0.1927 0.0469 0.0689 0.0209

Panel D. YN
homo. 0.1788* 0.0473 0.0949* 0.0055
AR10sq 0.1739 0.0466 0.0932 0.0056
AR10ab 0.1711 0.0470 0.0933 0.0056
GARCH 0.1724 0.0468 0.0931 0.0056
GARCH-t 0.1762* 0.0469 0.0961* 0.0057
GARCH-g 0.1697 0.0467 0.0926 0.0055
e.smooth-N 0.1724 0.0470 0.0942 0.0057
e.smooth-t 0.1772* 0.0473 0.0971* 0.0058
*
s.v. 0.1720 0.0475 0.0947 0.0055

Note: The columns contain the QPS values for the models in-sample probability forecasts for the four specified
events. The minimum QPS value within each column of each panel is underlined. * indicates that this particular set
of in-sample forecasts rejects the null hypothesis of equality with the minimum QPS value using the S test at the
five percent significance level; i.e., the models thus marked have QPS values significantly greater than the minimum
value.

35
Table 8. Out-of-sample QPS Results

Event 1 Event 2 Event 3 Event 4

Calib. Calib. Calib. Calib.

Panel A. BP
homo. 0.1600* 0.0178 0.0919 0.0135
AR10sq 0.1392* 0.0163 0.0734 0.0121
AR10ab 0.1338 0.0161 0.0696 0.0115
GARCH 0.1330 0.0160 0.0682 0.0114
GARCH-t 0.1411* 0.0171 0.0757* 0.0119
GARCH-g 0.1312 + 0.0160 0.0677 0.0112
e.smooth-N 0.1318 + 0.0161 0.0672 + 0.0109
e.smooth-t 0.1338 0.0165 0.0672 + 0.0111
s.v. 0.1332 + 0.0159 0.0683 + 0.0113 +

Panel B. CD
homo. 0.0538 + 0.0615 + 0.0504 + 0.0290
AR10sq 0.0542 + 0.0625 + 0.0503 + 0.0278
AR10ab 0.0543 + 0.0624 + 0.0505 + 0.0283
GARCH 0.0551 + 0.0641 0.0518 + 0.0279
GARCH-t 0.0571 0.0666 0.0551* 0.0279
GARCH-g 0.0545 + 0.0632 + 0.0511 + 0.0279
e.smooth-N 0.0544 + 0.0634 0.0510 + 0.0280
e.smooth-t 0.0557 0.0652 0.0508 + 0.0280
s.v. 0.0540 + 0.0620 + 0.0508 + 0.0296 +

Panel C. DM
homo. 0.2005 0.0402 + 0.1509 0.0104
AR10sq 0.1939 + 0.0400 + 0.1429 + 0.0104
AR10ab 0.1928 + 0.0406 + 0.1428 + 0.0104
GARCH 0.1923 + 0.0402 + 0.1416 + 0.0105
GARCH-t 0.1961 0.0410 0.1462* 0.0105
GARCH-g 0.1918 + 0.0401 + 0.1417 + 0.0105
e.smooth-N 0.1929 + 0.0407 + 0.1424 + 0.0104
e.smooth-t 0.1942 + 0.0412 0.1424 + 0.0104
s.v. 0.1929 + 0.0408 + 0.1442* 0.0103 +

Panel D. YN
homo. 0.2008 + 0.0457 + 0.1295 + 0.0091
AR10sq 0.1993 + 0.0454 + 0.1272 + 0.0091
AR10ab 0.1977 + 0.0453 + 0.1274 + 0.0092
GARCH 0.1976 + 0.0451 + 0.1256 + 0.0092
GARCH-t 0.2007 0.0461 0.1304 0.0097
GARCH-g 0.1976 + 0.0451 + 0.1248 + 0.0090
e.smooth-N 0.2016 + 0.0459 0.1288 + 0.0093
e.smooth-t 0.2042 0.0470 0.1289 + 0.0094
s.v. 0.2012 0.0456 + 0.1327 0.0093

Note: The event columns contain the QPS values for the models out-of-sample probability forecasts for the four
specified events. The minimum QPS value within each column of each panel is underlined. * indicates that this
particular set of out-of-sample forecasts rejects the null hypothesis of equality with the minimum QPS value using
the S test at the five percent significance level; i.e., the models thus marked have QPS values significantly greater
than the minimum value. The companion columns labeled calib. present the results from the Z0 calibration test
using the deciles of the unit interval (i.e., [0,0.1); [0.1,0.2); ...; [0.9,1]) as the subsets of interest. The + symbol
indicates that the forecasts do not reject the null hypothesis and are thus considered to be well calibrated over the
selected subsets.

36
 Figure 1. Alternative Distributions of the BP Standardized Residuals

Figure 1a.

Figure 1b.

Note: The standard normal distribution is plotted with a solid line. The estimated t- and GED-distributions are
plotted with dotted and dashed lines, respectively. The estimated shape parameter (i.e., degrees of freedom) for the
t-distribution is 5.7, and the estimated shape parameter for the GED-distribution is 1.4.

37
 Figure 2. Out-of-Sample Probability Forecasts from the GARCH-g Model for BP

Note: The out-of-sample probability forecasts from the GARCH-g model for events 1 through 4 are plotted with a
dot-dash line, a dotted line, a dashed line and a solid line, respectively.

38
 APPENDIX: Volatility Model Parameter Estimates

Table A.1. Volatility Model Parameters for the Gaussian Homoskedastic Model

yt
 t
1  t

t G t
1 ~ N 0, h t

ht
2

Parameter BP CD DM YN

0.0118 0.0035 0.0005 -0.0216

(0.0236) (0.0201) (0.0228) (0.0224)

O 0.0815* 0.0499* 0.0362* 0.0334*

(0.0257) (0.0206) (0.0143) (0.0153)

c2 0.5168* 0.0700* 0.5203* 0.4255*

(0.0244) (0.0028) (0.0237) (0.0191)

Note: These parameters values are maximum likelihood estimates. The standard errors in parentheses are the
robust standard errors of Bollerslev and Wooldridge (1992).* indicates significance at the five percent level.

39
Table A.2. Volatility Model Parameters for the AR10sq Model
yt
 t
1  t

t G t
1 ~ N 0, h t
10
2
ht
 i t
i .
i1
Parameter BP CD DM YN
0.0118 0.0035 0.0005 -0.0216
(0.0236) (0.0201) (0.0228) (0.0224)

O 0.0815* 0.0499* 0.0362* 0.0334*

(0.0257) (0.0206) (0.0143) (0.0153)

q 0.2600* 0.0355* 0.2781* 0.2578*

(0.0169) (0.0037) (0.0263) (0.0230)

?1 0.0726* 0.0244 0.0515* 0.0414*

(0.0169) (0.0169) (0.0169) (0.0168)

?2 0.0183 0.0363* 0.0160 0.0232

(0.0169) (0.0170) (0.0169) (0.0169)

?3 0.0180 0.0406* 0.0325 0.0363*

(0.0169) (0.0171) (0.0169) (0.0169)

?4 0.0379* 0.0093 0.0353* -0.0163

(0.0169) (0.0171) (0.0169) (0.0169)

?5 0.0437* 0.0224 0.0596* 0.0282

(0.0169) (0.0171) (0.0169) (0.0169)

?6 0.0924* 0.0560* 0.0709* 0.0869

(0.0169) (0.0171) (0.0169) (0.0169)

?7 0.0773* 0.0247 0.0406* 0.0152

(0.0169) (0.0171) (0.0169) (0.0169)

?8 0.0492* 0.0908* 0.0606* 0.0779

(0.0169) (0.0170) (0.0169) (0.0169)

?9 0.0494* 0.0593* 0.0462* 0.0541

(0.0169) (0.0170) (0.0169) (0.0169)

?10 0.0391* 0.1314* 0.0539* 0.0443

(0.0168) (0.0169) (0.0169) (0.0169)

Note: The conditional mean parameters are estimated using ordinary least squares. The conditional variance
parameters are estimated using ordinary least squares on the squared residuals. Standard errors are presented in
parentheses. * indicates significance at the five percent level.

40
Table A.3. Volatility Model Parameters for the AR10ab Model

yt
 t
1  t t G t
1 ~ N 0, ht
12 2

ht
 i t
i
2 i1

Parameter BP CD DM YN
0.0118 0.0035 0.0005 -0.0216
(0.0236) (0.0201) (0.0228) (0.0224)

O 0.0815* 0.0499* 0.0362* 0.0334*

(0.0257) (0.0206) (0.0143) (0.0153)

q 0.2447* 0.0794* 0.2487* 0.2611*

(0.0168) (0.0069) (0.0169) (0.0201)

?1 0.0835* 0.0291 0.0652* 0.0450*

(0.0169) (0.0169) (0.0169) (0.0168)

?2 0.0462* 0.0027 0.0312 0.0399*

(0.0169) (0.0170) (0.0169) (0.0168)

?3 0.0270 0.0498* 0.0499* 0.0335*

(0.0169) (0.0171) (0.0169) (0.0169)

?4 0.0356* 0.0167 0.0498* 0.0055

(0.0169) (0.0171) (0.0169) (0.0169)

?5 0.0390* 0.0605* 0.0457* 0.0358*

(0.0169) (0.0171) (0.0169) (0.0169)

?6 0.1044* 0.0895* 0.0877* 0.0727*

(0.0169) (0.0171) (0.0169) (0.0169)

?7 0.0453* 0.0491* 0.0457* 0.0289

(0.0169) (0.0171) (0.0169) (0.0169)

?8 0.0548* 0.0963* 0.0731* 0.0727*

(0.0169) (0.0170) (0.0169) (0.0169)

?9 0.0476* 0.0744* 0.0442* 0.0583*

(0.0168) (0.0170) (0.0169) (0.0169)

?10 0.0575* 0.1198* 0.0453* 0.0536*

(0.0168) (0.0169) (0.0169) (0.0169)

Note: The conditional mean parameters are estimated using ordinary least squares. The conditional variance
parameters are estimated using ordinary least squares on the absolute value of the residuals. Standard errors are
presented in parentheses. * indicates significance at the ten percent level.

41
Table A.4. Volatility Model Parameters for the Gaussian GARCH(1,1) Model
yt
 t
1  t

t G t
1 ~ N 0, h t
2
ht
 t
1  ht
1

Parameter BP CD DM YN
0.0067 0.0010 0.0060 -0.0135
(0.0111) (0.0037) (0.0111) (0.0104)

O 0.0704* 0.0425* 0.0349 0.0231

(0.0178) (0.0196) (0.0181) (0.0185)

q 0.0083* 0.0018* 0.0153* 0.0180*

(0.0033) (0.0002) (0.0029) (0.0048)

? 0.0510* 0.1432* 0.0810* 0.0587*

(0.0203) (0.0320) (0.0259) (0.0242)

A 0.9332* 0.8430* 0.8919* 0.8993*

(0.0810) (0.0588) (0.0733) (0.1225)

Note: These parameters values are maximum likelihood estimates. The standard errors in parentheses are the
robust standard errors of Bollerslev and Wooldridge (1992). * indicates significance at the five percent level.

Table A.5. Volatility Model Parameters for the GARCH(1,1)-t Model

yt
 t
1  t

t G t
1 ~ t 0, ht, v
2
ht
 t
1  ht
1
Parameter BP CD DM YN
-0.0002 -0.0033 0.0127 0.0054
(0.0116) (0.0035) (0.0101) (0.0095)

O 0.0457* 0.0320 0.0319 0.0290

(0.0205) (0.0187) (0.0167) (0.0168)

q 0.0095* 0.0011* 0.0134* 0.0083*

(0.0033) (0.0005) (0.0037) (0.0030)

? 0.0588* 0.1290* 0.0792* 0.0546*

(0.0105) (0.0182) (0.0114) (0.0101)

A 0.9266* 0.8691* 0.9018* 0.9247*

(0.0137) (0.0190) (0.0138) (0.0149)

v 5.7176* 5.7167* 5.7162* 5.6939*

(0.0549) (0.0518) (0.0425) (0.0567)

Note: These parameters values are maximum likelihood estimates. Asymptotic standard errors, presented in
parentheses, are derived from the value of the numerical Hessian at the estimated parameter values. * represents
significance at the five percent level.

42
Table A.6. Volatility Model Parameters for the GARCH(1,1)-GED Model

yt
 t
1  t

t G t
1 ~ GED 0, h t,
2
ht
 t
1  ht
1

Parameter BP CD DM YN
0.0009 -0.0021 0.0145 0.0073
(0.0100) (0.0033) (0.0096) (0.0088)

O 0.0389* 0.0327 0.0265 -0.0068

(0.0173) (0.0167) (0.0168) (0.0146)

q 0.0089* 0.0014* 0.0138* 0.0115*

(0.0030) (0.0004) (0.0035) (0.0040)

? 0.0548* 0.1333* 0.0789* 0.0581*

(0.0095) (0.0166) (0.0106) (0.0113)

A 0.9284* 0.8572* 0.8966* 0.9156*

(0.0127) (0.0170) (0.0131) (0.0186)

M 1.3838* 1.3663* 1.4071* 1.1665*

(0.0467) (0.0417) (0.0463) (0.0363)

Note: These parameters values are maximum likelihood estimates. Asymptotic standard errors, presented in
parentheses, are derived from the value of the numerical Hessian at the estimated parameter values. * represents
significance at the 5% level.

43
Table A.7. Volatility Model Parameters for the Stochastic Volatility Model

yt
 t
1  t

t
exp t / 2 t, t ~ N( 0, 1 )
2
t
t
1  t, t ~ N( 0, )
t t

2
Measurement equation: log t
1.27  t  t

Transition equation: t
 t
1  t

Parameter BP CD DM YN
0.0118 0.0035 0.0005 -0.0216
(0.0236) (0.0201) (0.0228) (0.0224)

O 0.0815* 0.0499* 0.0362* 0.0334*

(0.0257) (0.0206) (0.0143) (0.0153)

k 0.9789* 0.9648* 0.9778* 0.9378*

(0.0069) (0.0100) (0.0068) (0.0303)
2
0.0173* 0.0484* 0.0186* 0.0609*
(0.0071) (0.0107) (0.0066) (0.0298)

Note: The conditional mean parameters are estimated using ordinary least squares, and the standard errors are
presented in parentheses. * indicates significance at the 5% level. The conditional variance parameters are
estimated using the Kalman filter and quasi-maximum likelihood methods as proposed in Harvey et al. (1994).
These standard errors are estimated using the method proposed by Dunsmuir (1979).

44