Conditional Heteroskedasticity in Asset Returns: A New Approach

Author(s): Daniel B. Nelson

Source: Econometrica , Mar., 1991, Vol. 59, No. 2 (Mar., 1991), pp. 347-370
Published by: The Econometric Society

Stable URL: https://www.jstor.org/stable/2938260

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

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Econometrica

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 Econometrica, Vol. 59, No. 2 (March, 1991), 347-370

CONDITIONAL HETEROSKEDASTICITY IN ASSET RETURNS:

A NEW APPROACH

BY DANIEL B. NELSON1

GARCH models have been applied in modelling the relation between conditional
variance and asset risk premia. These models, however, have at least three major
drawbacks in asset pricing applications: (i) Researchers beginning with Black (1976) have
found a negative correlation between current returns and future returns volatility.
GARCH models rule this out by assumption. (ii) GARCH models impose parameter
restrictions that are often violated by estimated coefficients and that may unduly restrict
the dynamics of the conditional variance process. (iii) Interpreting whether shocks to
conditional variance "persist" or not is difficult in GARCH models, because the usual
norms measuring persistence often do not agree. A new form of ARCH is proposed that
meets these objections. The method is used to estimate a model of the risk premium on
the CRSP Value-Weighted Market Index from 1962 to 1987.

KEYWORDS: Autoregressive conditional heteroskedasticity (ARCH), generalized au-

toregressive conditional heteroskedasticity (GARCH), exponential ARCH, market volatil-
ity, risk premium, nonlinear time series.

1. INTRODUCTION

AFTER THE EVENTS of October 1987, few would argue with the proposition that
stock market volatility changes randomly over time. Understanding the way in
which it changes is crucial to our understanding of many areas in macroeco-
nomics and finance, for example the term structure of interest rates (e.g., Barsky
(1989), Abel (1988)), irreversible investment (e.g., Bernanke (1983), McDonald
and Siegel (1986)), options pricing (e.g., Wiggins (1987)), and dynamic capital
asset pricing theory (e.g., Merton (1973), Cox, Ingersoll, and Ross (1985)).
Recent years have also seen a surge of interest in econometric models of
changing conditional variance. Probably the most widely used, but by no means
the only such models,2 are the family of ARCH (autoregressive conditionally
heteroskedastic) models introduced by Engle (1982). ARCH models make the
conditional variance of the time t prediction error a function of time, system
parameters, exogenous and lagged endogenous variables, and past prediction
errors. For each integer t, let (t be a model's (scalar) prediction error,
vector of parameters, xt a vector of predetermined variables, and ot2
variance of (t given information at time t. A univariate ARCH model bas

1 This paper is a revision of part of Chapter III of my Ph.D. dissertation (Nelson (198
Department of Education and the University of Chicago Graduate School of Business provided
financial support. I am indebted to George Constantinides, John Cox, Nancy Hammond, Daniel
McFadden, Mark Nelson, Adrian Pagan, James Poterba, G. William Schwert, Stephen Taylor,
Jeffrey Wooldridge, Arnold Zellner, two referees, and a co-editor for helpful comments. Seminar
participants at the 1987 Econometric Society Winter meetings, at the 1989 A.S.A. meetings, at
Berkeley, Chicago, M.I.T., Northwestern, Princeton, the Research Triangle Econometrics Work-
shop, Rochester, Stanford, Wharton, and Yale also made helpful suggestions. Jiahong Shi provided
able research assistance. Remaining errors are mine alone.
2 See, e.g., Poterba and Summers (1986), French, Schwert, and Stambaugh (1987), and Nelson
(1988, Chapter 1).

347

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
348 DANIEL B. NELSON

Engle (1982) equations 1-5 sets

(1.1) {tSt Zt,

(1.2) zt - i.i.d. with

E(zt) =0, Var(zt) = 1, and

(1.3) ?"t 2=(T 2((t_j, ft-2 ........ ,t,xt,b)

= (02(ot_lZt_1 O't-2Zt-2 I * * , t, xt, b)-
The system (1.1)-(1.3) can easily be given a multivariate interpretation, in
which case zt is an n by one vector and ort2 is an n by n matrix. We refer to any
model of the form (1.1)-(1.3), whether univariate or multivariate, as an ARCH
model.
The most widely used specifications for o.2(,*,..., ) are the linear ARCH
and GARCH models introduced by Engle (1982) and Bollerslev (1986) respec-
tively, which make ,t2 linear in lagged values of 62 = o,t2z2 by defining
p

(1.4) c,t2 =w0+ E, aj Z 2 j t2 , and

j=1
q p

(1.5) a2 = w +
i=l j=1

respectively, where w, the

case of (1.5), we refer to both (1.4) and (1.5) as GARCH models, to distinguish
them as special cases of (1.3).
The GARCH-M model of Engle and Bollerslev (1986a), adds another equa-
tion

(1.6) Rt =a +bcrt2+ t,
in which o7t2, the conditional variance of Rt, enters the conditional m
as well. For example if Rt is the return on a portfolio at time t, its requir
of return may be linear in its risk as measured by o,t2.
Researchers have fruitfully applied the new ARCH methodology in asset
pricing models: for example, Engle and Bollerslev (1986a) used GARCH(1, 1) to
model the risk premium on the foreign exchange market, and Bollerslev, Engle,
and Wooldridge (1988) extended GARCH(1, 1) to a multivariate context to test
a conditional CAPM with time varying covariances of asset returns.
Substituting recursively for the p13-t2 terms lets us rewrite (1.5) as3
00

(1.7) ?2 =(D* + E z2urt2

k=1

It is readily verified that if w, the aj, and the f3i are nonnegative, w* a
are also nonnegative. By setting conditional variance equal to a constant plus a

3The representation (1.7) assumes that {o-2} is strictly stationary, so that the recursion can be
carried into the infinite past.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 349

weighted average (with positive weights) of past squared residuals, GARCH

models elegantly capture the volatility clustering in asset returns first noted by
Mandelbrot (1963): " .. . large changes tend to be followed by large changes-of
either sign-and small changes by small changes... " This feature of GARCH
models accounts for both their theoretical appeal and their empirical success.
On the other hand, the simple structure of (1.7) imposes important limitations
on GARCH models: For example, researchers beginning with Black (1976) have
found evidence that stock returns are negatively correlated with changes in
returns volatility-i.e., volatility tends to rise in response to "bad news" (excess
returns lower than expected) and to fall in response to "good news" (excess
returns higher than expected).4 GARCH models, however, assume that only the
magnitude and not the positivity or negativity of unanticipated excess returns
determines feature o-,2. If the distribution of zt is symmetric, the chang
variance tomorrow is conditionally uncorrelated with excess returns today.5 In
(1.4)-(1.5), o-7 is a function of lagged ot- and lagged Z2, and so is invaria
changes in the algebraic sign of the zr's-i.e., only the size, not the sign, of
lagged residuals determines conditional variance. This suggests that a model in
which o-t2 responds asymmetrically to positive and negative residuals m
preferable for asset pricing applications.
Another limitation of GARCH models results from the nonnegativity con-
straints on w* and the k in (1.7), which are imposed to ensure that o-t2 re
nonnegative for all t with probability one. These constraints imply that increas-
ing Z2 in any period increases t2 for all m > 1, ruling out random oscillatory
behavior in the ot2 process. Furthermore, these nonnegativity constraints c
create difficulties in estimating GARCH models. For example, Engle, Lilien,
and Robins (1987), had to impose a linearly declining structure on the a
coefficients in (1.4) to prevent some of them from becoming negative.
A third drawback of GARCH modelling concerns the interpretation of the
"persistence" of shocks to conditional variance. In many studies of the time
series behavior of asset volatility (e.g., Poterba and Summers (1986), French,
Schwert, and Stambaugh (1987), and Engle and Bollerslev (1986a)), the central
question has been how long shocks to conditional variance persist. If volatility
shocks persist indefinitely, they may move the whole term structure of risk
premia, and are therefore likely to have a significant impact on investment in
long-lived capital goods (Poterba and Summers (1986)).
There are many different notions of convergence in the probability literature
(almost sure, in probability, in LP), so whether a shock is transitory or

4The economic reasons for this are unclear. As Black (1976) and Christie (1982) no
financial and operating leverage play a role, but are not able to explain the extent of the asymmetric
response of volatility to positive and negative returns shocks. Schwert (1989a,b) presents evidence
that stock volatility is higher during recessions and financial crises, but finds only weak relations
between stock market volatility and measures of macroeconomic uncertainty. See also Nelson
(1988), and Pagan and Hong (1988).
sAdrian Pagan pointed out, however, that in a GARCH-M model o-,2 may rise (fall) on
when returns are negative (positive), even though [,21 - ot2 and Rt are conditionally unco
since in (1.6) E[Iztl IRt < O]>E[Iztl IRt>O ] if a and b are positive.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
350 DANIEL B. NELSON

persistent may depend on our definition of convergence. In linear models it

typically makes no difference which of the standard definitions we use, since the
definitions usually agree. In GARCH models, the situation is more complicated.
For example, the IGARCH(1, 1) model of Engle and Bollerslev (1986a) sets

(1.8) '72 = (t' + (7t2Jl(1-a) +a AZ2-]a 0<<

When w = 0, o-t2 is a martingale. Based on the nature of persistence

models, it seems that IGARCH(1, 1) with w > 0 and w = 0 are analogous to
random walks with and without drift, respectively, and are therefore natural
models of "persistent" shocks. This turns out to be misleading, however: in
IGARCH(1, 1) with X = 0, ot2 collapses to zero almost surely, and in
IGARCH(1, 1) with X > 0, ot2 is strictly stationary and ergodic (Geweke (1986
Nelson (1990a)) and therefore does not behave like a random walk, since
random walks diverge almost surely.
The reason for this paradox is that in GARCH(1, 1) models, shocks may
persist in one norm and die out in another, so the conditional moments of
GARCH(1, 1) may explode even when the process itself is strictly stationary and
ergodic (Nelson (1990a)).
The object of this paper is to present an alternative to GARCH that meets
these objections, and so may be more suitable for modelling conditional vari-
ances in asset returns. In Section 2, we describe this ot2 process, and de
some of its properties. In Section 3 we estimate a simple model of stock market
volatility and the risk premium. Section 4 concludes. In Appendix I, we provide
formulas for the moments of ot- and (t in the model presented in Section 2
Proofs are in Appendix II.

2. EXPONENTIAL ARCH

If -t2 is to be the conditional variance of (t given info

clearly must be nonnegative with probability one. GARCH models ensure this
by making ot2 a linear combination (with positive weights) of positive ran
variables. We adopt another natural device for ensuring that ot2 remains
nonnegative, by making ln (-t2) linear in some function of time and lagged zt's.
That is, for some suitable function g:

00

(2.1) ln (St2) = at + E fk9g(Zt-k), 1- 1,

k=1

where {tt}t=-_oo and {Pk}k= , X are real, nonstochastic, sc

Pantula (1986) and Geweke (1986) have previously proposed ARCH models of
this form; in their log-GARCH models, g(zt) = lnlztlb for some b > 0.6

6 See also Engle and Bollerslev (1986b).

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 351

To accommodate the asymmetric relation between stock returns and volatility

changes noted in Section 1, the value of g(z,) must be a function of both
magnitude and the sign of zt. One choice, that in certain important cases tur
out to give o-7 well-behaved moments, is to make g(zt) a linear combination
zt and lzt :

(2.2) g(zt) ozt + y[IztI -EIztI].

By construction, {g(zt)}t= is a zero-mean, i.i.d. random sequence. The two

components of g(zt) are Ozt and -y[zt I - ElztI], each with mean zero. If the
distribution of zt is symmetric, the two components are orthogonal, though of
course they are not independent. Over the range 0 < zt < ?, g(zt) is linear in zt
with slope 0 + y, and over the range - c <zt < 0, g(zt) is linear with slope
0 - y. Thus, g(zt) allows the conditional variance process {ot2} to respond
asymmetrically to rises and falls in stock price.
To see that the term y4 I zt I - E I]z I] represents a magnitude effect in the spir
of the GARCH models discussed in Section 1, assume for the moment that
y > 0 and 0 = 0. The innovation in ln (-2t+ ) is then positive (negative) w
magnitude of zt is larger (smaller) than its expected value. Suppose now
y = 0 and 0 < 0. The innovation in conditional variance is now positive (nega-
tive) when returns innovations are negative (positive). Thus the exponential
form of ARCH in (2.1)-(2.2) meets the first objection raised to the GARCH
models in Section 1.
In Section 1 we also argued that the dynamics of GARCH models were
unduly restrictive (i.e., oscillatory behavior is excluded) and that they impose
inequality constraints that were frequently violated by estimated coefficients.
But note that in (2.1)-(2.2) there are no inequality constraints whatever, and
that cycling is permitted, since the /3k terms can be negative or positive.
Our final criticism of GARCH models was that it is difficult to evaluate
whether shocks to variance "persist" or not. In exponential ARCH, however,
ln (o-t2) is a linear process, and its stationarity (covariance or strict) and ergodic
ity are easily checked. If the shocks to {ln (ot2)} die out quickly enough, an
we remove the deterministic, possibly time-varying component {ta}, then {ln (o-
is strictly stationary and ergodic. Theorem 2.1 below states conditions for the
ergodicity and strict stationarity of {exp(-at)-t2} and {exp(-at/2){t}, which
are {to} and {ft} with the influence of {ta} removed.

THEOREM 2.1: Define {o-7t2}, {t}, and {zt} by (1.1)-(1.2) and (2.1)-(2.2), and
assume that y and 0 do not both equal zero. Then {exp ( - a t)0t-}, {exp ( - at/2)(t},
and {ln (o-t2) - at} are strictly stationary and ergodic and {ln (o-2) - at} is covar
ance stationary if and only if Ek=1f2 < c. If Ek1f,82 = cc, then Iln (t2) - at] = X
almost surely. If k=E1f2 < , then for k > O, CovI{Ztkln (ot2)}=fk[6+
yE(ztlztl)], and Cov[ln(t2), ln(ot2-k)] = Var [g(zt)]E>ljj+k.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
352 DANIEL B. NELSON

The stationarity and ergodicity criterion in Theorem 2.1 is exactly the same as
for a general linear process with finite innovations variance,7 so if, for example,
In (o-2) follows an AR(1) with AR coefficient A, ln (o-2) is strictly stationary and
ergodic if and only if JA I < 1.
There is often a simpler expression for ln (o-2) than the infinite moving
average representation in (2.1). In many applications, an ARMA process pro-
vides a parsimonious parameterization:

(2.3) ln(o-r2)=a ( + 1P,L+ +qLq)

We assume that [1 - Ei=I p A i] and [1 + Ei=i qtliiy'] have no common roots

By Theorem 2.1, {exp( a)o-t2} and {exp(-at/2)(t} are then strictly stationary
and ergodic if and only if all the roots of [1 - Ei= Py i] lie outside the unit
circle. While an ARMA representation may be suitable for many modelling
purposes, Theorem 2.1 and the representation (2.1) also allow "long memory"
(Hosking (1981)) processes for {ln (o-t2)}.
Next consider the covariance stationarity of {o0t2} and {ft}. According
Theorem 2.1, E,82 < oo implies that {exp(-a )o2} and {exp(-at/2){t} are
strictly stationary and ergodic. This strict stationarity, however, need not imply
covariance stationarity, since {exp(-at))o-t2} and {exp(-at/2){t} may fail to
have finite unconditional means and variances. For some distributions of {zt}
(e.g., the Student t with finite degrees of freedom), {exp (- at)o-t)2 and
{exp(-at/2)(t} typically have no finite unconditional moments. The results for
another commonly used family of distributions, the GED (Generalized Error
Distribution (Harvey (1981), Box and Tiao (1973)8), are more encouraging. The
GED includes the normal as a special case, along with many other distributions,
some more fat tailed than the normal (e.g., the double exponential), some more
thin tailed (e.g., the uniform). If the distribution of {zt} is a member of this
family and is not too fat tailed, and if Ei=1 .,B 2 < ?, then {o7t2} and {ft} h
finite unconditional moments of arbitrary order.
The density of a GED random variable normalized to have a mean of zero
and a variance of one is given by

(2.4) (2 4) f()
f(z)= = P [ 2)zA ]c<<c
A2+/v)F(1v x < z < cO O < v < x,
<~c

where F(*) is the gamma function, and

(2.5) A- [2(12/v)(1/V)/F(3/V)]1/2

7The assumption that zt has a finite variance can be relaxed. For examp
Cauchy, and g(zt) - Ozt + y IZt |, with y and 0 not both equal to zero. By the T
(Billingsley (1986, Theorem 22.8)), o-t2 is finite almost surely if and only if I E,
p 3-2)1 <00 and E,> 11,1 I <0X. Then {exp(- at)o-t21 and {exp(-a-t/2)t) are stric
ergodic.
8 Box and Tiao call the GED the exponential power distribution.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 353

v is a tail-thickness parameter. When v = 2, z has a standard normal distribu-

tion. For v < 2, the distribution of z has thicker tails than the normal (e.g.,
when v = 1, z has a double exponential distribution) and for v > 2, the distribu-
tion of z has thinner tails than the normal (e.g., for v = oo, z is uniformly
distributed on the interval [3 1/2, 31/2]).

THEOREM 2.2: Define {o7, t}= by (1.1)-(1.2) and (2.1)-(2.2), and as-
sume that y and 0 do not both equal zero. Let {zt}t= -??? be i.i.d. GED with
mean zero, uariance one, and tail thickness parameter u > 1, and let Ei=l12< ??
Then {exp ( - at)o-t2} and {exp ( - at/2)ft} possess finite, time-invariant moments
of arbitrary order. Further, if 0 < p < oo, conditioning information at time 0 drops
out of the forecast pth moments of exp (- t)o-t2 and exp(-at/2)(t as t -3 x:

(2.6) plimE[exp (-pat)ot2P] -0, Z-2and.

t -->00

- E [exp ( -Pat ) 0-t2p ] 0, and

(2.7) plim E[Iexp ( -p a t2)1ft IP I zo, Z-1, Z-21 ...*

t -->00

- E[exp(-pat)jftjI] =0,
where plim denotes the limit in probability.

That is, if the distribution of the zt is GED and is thinner-tailed than the
double exponential, and if Ep32 < o, then exp (-at)2 and exp (at/2)(t are not
only strictly stationary and ergodic, but have arbitrary finite moments, which in
turn implies that they are covariance stationary.
Since the moments of {exp(-at)o-t} and {exp(-at/2)(t} are of interest for
forecasting, Appendix I derives the conditional and unconditional moments
(including covariances) of {exp ( -at)o-t2} and {exp ( -at/2)(t} under a vari
distributional assumptions for {zt), including Normal, GED, and Student

3. A SIMPLE MODEL OF MARKET VOLATILITY

In this Section, we estimate and test a simple model of market risk, asset
returns, and changing conditional volatility. We use this model to examine
several issues previously investigated in the economics and finance literature,
namely (i) the relation between the level of market risk and required return,
(ii) the asymmetry between positive and negative returns in their effect on
conditional variance, (iii) the persistence of shocks to volatility, (iv) "fat tails" in
the conditional distribution of returns, and (v) the contribution of nontrading
days to volatility.
We use the model developed in Section 2 for the conditional variance
process, assuming an ARMA representation for ln (o,27). To allow
possibility of nonnormality in the conditional distribution of returns, we a
that the {zt} are i.i.d. draws from the GED density (2.4). To account for the

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 354 DANIEL B. NELSON

contribution of nontrading periods to market variance, we assume that each

nontrading day contributes as much to variance as some fixed fraction of a
trading day, so if, for example, this fraction is one tenth, than (72 on a typical
Monday would be 20% higher than on a typical Tuesday. Other researchers
(e.g., Fama (1965), and French and Roll (1986)) have found that nontrading
periods contribute much less than do trading periods to market variance, so we
expect that 0 < a << 1.9
Specifically, we model the log of conditional variance as

(3 .1) In (-27) =a O+ (1 + -1IL + + +qLq)

where zt is i.i.d. GED with mean zero, variance one, and tail thickness
parameter v > 0, and {ta) is given by

(3.2) at =+ln(1+Nt>),

where Nt is the number of nontrading days between tra

and a and a are parameters. If the unconditional expectation of ln (o 2) exists,
then it equals a + ln(1 + Nt ). Together, (3.1)-(3.2) and Theorem 2.1 imply
that ((1 + NS)- 1/2ft} is strictly stationary and ergodic if and only if all the roots
of (1 - A Y - PYP) lie outside the unit circle.10
We model excess returns Rt as

(3.3) Rt =a +bRt1 +cot2 +,t

where the conditional mean and variance of (t at time t are 0 and -t2
respectively, and where a, b, and c are parameters. The bRt_1 term allows for
the autocorrelation induced by discontinuous trading in the stocks making up an
index (Scholes and Williams (1977), Lo and MacKinlay (1988)). The Scholes and
Williams model suggests an MA(1) form for index returns, while the Lo and
MacKinlay model suggests an AR(1) form, which we adopt. As a practical
matter, there is little difference between an AR(1) and an MA(1) when the AR
and MA coefficients are small and the autocorrelations at lag one are equal,
since the higher-order autocorrelations die out very quickly in the AR model.
As Lo and MacKinlay note, however, such simple models do not adequately
explain the short-term autocorrelation behavior of the market indices, and no
fully satisfactory model yet exists.
The theoretical justification for including the c?t2 term in (3.3) is meager,
since the required excess return on a portfolio is linear in its conditional
variance only under very special circumstances. In Merton's (1973) intertempo-
ral CAPM model, for example, the instantaneous expected excess return on the
market portfolio is linear in its conditional variance if there is a representative
agent with log utility. Merton's conditions (e.g., continuous time, continuous

9French and Roll (1986) and Barclay, Litzenberger, and Warner (1990) offer economic interpre-
tations.
loAgain, we assume that [1 - Y1 DA1y'] and [1 + Y1 q"lqY'] have no common roots.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 355

trading, and a true "market" portfolio) do not apply in our model even
under the log utility assumption. Backus and Gregory (1989), and Glosten,
Jagannathan, and Runkle (1989) give examples of equilibrium models in which a
regression of returns on o-,2 yields a negative coefficient." There is, ther
no strong theoretical reason to believe that c is positive. Rather, the justifica-
tion for including co-2 is pragmatic: a number of researchers using GARCH
models (e.g., French, Schwert, and Stambaugh (1987) and Chou (1987)) have
found a statistically significant positive relation between conditional variance
and excess returns on stock market indices, and we therefore adopt the form
(3.3).
For a given ARMA(p, q) exponential ARCH model, the {Zt}t.1,T
and {Ot2}It=1, T sequences can be easily derived recursively given the data
{Rt}iT' and the initial values . 2 To close the model,
ln (o 2),..., ln (O+max{p2q+ ) were set equal to their un
(a + ln(1 + ?SN1)), ... ,(a + ln(1 + aSNi+max{p,q+ })). Th
log likelihood LT as

(3.4) LT = ln (v/A) - 2 #Rt - a - bR - ct2 )/o-tA

t= 1

-l+v-u)In (2) - In [ F( 1/ u 21in (ct2)

where A is defined in (2.5). Given the parameters and initial states, we can
easily compute the likelihood (3.4) recursively, using (3.1) and setting

(3.5) zt = l-7(Rt - a - bRt- ct2).

In light of (3.5), however, it may be that setting the out-of-sample values of
ln (Ut2) equal to their unconditional expectation is not innocent: using the tru
parameter values but the wrong -t2 in (3.5) leads to an incorrect fitted value of
Zt, in turn leading to an incorrect fitted value for ut2 1, and so on. In simulations
using parameter estimates similar to those reported below, fitted o-t2 generated
by (3.4)-(3.5) with an incorrect starting value converge very rapidly to the o'2
generated by (3.4)-(3.5) and a correct starting value. In a continuous time limit,
it is possible to prove that this convergence takes place instantaneously (Nelson
(1990b)).
Under sufficient regularity conditions, the maximum likelihood estimator is
consistent and asymptotically normal. Unfortunately, verifying that these condi-
tions hold in ARCH models has proved extremely difficult in both GARCH
models and in the Exponential ARCH model introduced in this paper. Weiss
(1986) developed a set of sufficient conditions for consistency and asymptotic
normality in a variant of the linear ARCH formulation of Engle (1982). These
conditions are quite restrictive, and are not satisfied by the coefficient estimates
obtained in most studies using this form of ARCH. In the GARCH-M model, in
which conditional variance annears in the conditional mean the asvmntotics are

11 See also Gennotte and Marsh (1987).

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
356 DANIEL B. NELSON

even more uncertain, and no sufficient conditions for consistency and asymp-
totic normality are yet known. The asymptotics of exponential ARCH models
are equally difficult, and as with other ARCH models, a satisfactory asymptotic
theory for exponential ARCH is as yet unavailable. In the remainder of this
paper we assume (as is the usual practice of researchers using GARCH models)
that the maximum likelihood estimator is consistent and asymptotically normal.
For our empirical analysis, we use the daily returns for the value-weighted
market index from the CRSP tapes for July 1962-December 1987. An immedi-
ate problem in using this data is that we wish to model the excess returns
process but do not have access to any adequate daily riskless returns series. As
an initial approximation to the riskless rate, we extracted the monthly Treasury
bill returns from the CRSP tapes, assumed that this return was constant for
each calendar day within a given month, and computed daily excess returns
using this riskless rate series and value-weighted CRSP daily market return
As a check on whether measurement errors in the riskless rate series are likely
to bias the results seriously, we also fit the model using the capital gains series,
ignoring both dividends and the riskless interest rate. As shown below, it made
virtually no difference in either the estimated parameters or the fitted variances.
To select the order of the ARMA process for ln (o-2), we used the Schwarz
Criterion (Schwarz (1978)), which provides consistent order-estimation in the
context of linear ARMA models (Hannan (1980)). The asymptotic properties of
the Schwarz criterion in the context of ARCH models are unknown.
The maximum likelihood parameter estimates were computed on VAX 8650
and 8550 computers using the IMSL subroutine DUMING. Table I lists
likelihood values for ARMA models of various orders on the CRSP excess
returns series. For both the excess returns and capital gains series, the Schwarz
Criterion selected an ARMA(2, 1) model for ln(o.7).13 Table II gives the
parameter estimates and estimated standard errors for both ARMA(2, 1) mod-
els. The estimated correlation matrix of the parameter estimates in the excess
returns model is in Table III. The asymptotic covariance matrix was computed
using the score.
First note that except for the parameter c (the risk premium term in (3.3)),
the two sets of coefficient estimates are nearly identical. The fitted values of
ln (o!2) from the two models are even more closely related: their means and
variances for the 1962-1987 period are nearly equal (- 9.9731 and 0.6441, vs.
- 9.9766 and 0.6415, respectively) and the two series have a sample correlation
of 0.9996. In other words, the series are practically identical, so ignoring
dividends and interest payments appears likely to introduce no important errors
in forecasting the volatility of broad market indices.

12 Logarithmic returns are used throughout: i.e., if S, is the level of the value-weighted market
index at time t and d, are the dividends paid at t, then the value-weighted market return, capital
gain, and excess return are computed as In[(S, + d,)/St1], In[SI/S,_1], and In[(S, + d,)/S,1] -
RR, respectively, where RR, is our proxy for the riskless rate.
13 The AIC (Akaike (1973)) chose the highest-order model estimated.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 357

TABLE I

LIKELIHOOD VALUES FOR ARMA MODELS FOR CRSP VALUE-WEIGHTED EXCESS RETURNS

Observations = 6408.

Deterministic Conditional Variance Model (i.e., y = H = A = V = 0), Likelihood = 22273.313.

ARMA Exponential ARCH Model Likelihoods:

AR Order
0 1 2 3 4

MA Order 0 22384.898 22888.052 22891.937 22894.237 22894.902

1 22429.942 22893.687 22915.454 22916.799 22916.894
(SC)
2 22473.728 22894.167 22916.762 22917.035 22918.708
3 22532.768 22894.385 22916.941 22918.853 22922.439
4 22590.536 22900.990 22917.670 22918.857 22923.752
(AIC)

(SC = Model selected by the criterion of Schwarz (1978).)

(AIC = Model selected by the information criterion of Akaike (1973).)

TABLE II
PARAMETER ESTIMATES FOR THE CRSP EXCESS
RETURNS AND CAPITAL GAINS MODELS
(Standard Errors in Parentheses)

Parameter CRSP Excess Returns CRSP Capital Gains

a -10.0593 - 10.0746
(0.3462) (0.3361)
8 0.1831 0.1676
(0.0277) (0.0271)
y 0.1559 0.1575
(0.0125) (0.0126)
1.92938 1.92914
(0.0145) (0.0146)
A2 - 0.92941 - 0.92917
(0.0145) (0.0146)
V - 0.9782 - 0.9781
(0.0062) (0.0063)
H -0.1178 -0.1161
(0.0090) (0.0090)
a 3.488 *10-4 3.416 *10-4
(9.850 10-5) (9.842 10-5)
b 0.2053 0.2082
(0.0123) (0.0123)
c - 3.3608 - 1.9992
(2.0261) (2.0347)
v 1.5763 1.5760
(0.0320) (0.0320)

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
358 DANIEL B. NELSON

TABLE III

ESTIMATED CORRELATION MATRIX FOR PARAMETER ESTIMATES

EXCESS RETURNS MODEL
(Only lower triangle reported)

Parameter

a1
a - 0.0087 1
y - 0.0023 - 0.1639 1
,A 1 0.0139 0.0715 -0.2996 1
A2 - 0.0138 - 0.0719 0.2994 - 0.999992 1
4' - 0.0094 - 0.0392 0.1020 - 0.8356 0.8352
1
H -0.0472 0.1909 -0.0591 0.4342 -0.4346
-0.2211 1
a - 0.0674 - 0.0342 - 0.0222 0.0734 - 0.0715
0.0064 0.0251 1
b - 0.0144 - 0.0395 - 0.0109 0.0439 -0.0443
- 0.0256 - 0.0772 - 0.1979 1
c - 0.0702 0.1111 0.0815 - 0.0891 0.0871
- 0.0151 - 0.0064 - 0.8287 0.1198 1
0.0591 0.1763 0.2853 - 0.0182 0.0179
- 0.0414 - 0.1554 0.0042 0.0402 0.1240
1
a 8 y A1 2
IF H a b c

Next we examine the empirical issues raised earlier in the Section:

(i) Market Risk and Expected Return: The estimated risk premium is nega-
tively (though weakly) correlated with conditional variance, with c - 3.361
with a large standard error of about 2.026. This contrasts with the significant
positive relation between returns and conditional variance found by researchers
using GARCH-M models (e.g., Chou (1987), and French, Schwert, and
Stambaugh (1987)), but agrees with the findings of other researchers not using
GARCH models (e.g., Pagan and Hong (1988)). Given the results of Gennotte
and Marsh (1987), Glosten, Jagannathan, and Runkle (1989), and Backus and
Gregory (1989), our findings of a negative (albeit insignificant) coefficient should
not be too surprising.
(ii) The asymmetric relation between returns and changes in volatility, as
represented by 0, is highly significant. Recall that a negative value of 0 indicates
that volatility tends to rise (fall) when returns surprises are negative (positive).
The estimated value for 0 is about - 0.118 (with a standard error of about
0.008) which is significantly below zero at any standard level. Figure 1 plots the
estimated g(z) function.
Figures 2 and 3 plot o-S (the daily conditional standard deviation of returns)
and the log value of the CRSP value-weighted market index, respectively. The
oS series is extremely variable, with lows of less than one-half of a percent and
highs over five percent. All the major episodes of high volatility are associated
with market drops.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 359

g(Z)
0

... \ .......... ...... .......... ........--- ---.............................

00

* ~~~0

...\\.....~
0 ~ ~~~...... ........

4 -3 -2 -1 0 1=3
z

FIGURE 1

Daily Volatility

- ... ............................................................... ..

Q e ~ ~ ~. . . . . . . . . . . . . . . . . . . . . . . . . . . .
'Dates: JUly 1 962 to December 1987

62 66 70 74 78 82 86 90
FIGURE 2

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 360 DANIEL B. NELSON

Log of Market Index

0

0 0 ......

o . , . . . . .. . . . , .. . , . . ., . . i..

'Dates: July 1962: through: December 1987,

? 62 66 70 74 78 82 86 90
FIGURE 3

(iii) Persistence of Shocks: The largest estimated AR root is approximately

0.99962 with a standard error of about 0.00086, so the t statistic for a unit root
is only about - 0.448. To gain intuition about the degree of persistence implied
by the largest AR root p, it is useful to think of the half-life h of a shock
associated with this root, i.e., the number h such that

(3.6) ph= 1/2

p = 0.99962 implies a half-life h of over 1820 trading days, about 7.3 years. In
contrast, the half-life implied by the smaller AR root is less than two weeks.
While this indicates substantial persistence and perhaps nonstationarity, it is
hard to know how seriously to take the point estimates, since we have only
about 25 years of data, about four times our estimated half-life for the larger
AR root. The usual cautions about interpreting an estimated AR root near the
unit circle as evidence of truly infinite persistence also apply.14
(iv) Fat Tails: It is well known that the distribution of stock returns has more
weight in the tails than the normal distribution (e.g., Mandelbrot (1963), Fama
(1965)), and that a stochastic process is thick tailed if it is conditionally normal
with a randomly changing conditional variance (Clark (1973)). Our estimated
model generates thick tails with both a randomly changing conditional variance
oy2and a thick-tailed conditional distribution for f. Recall from the discussi

14 See, e.g., Cochrane (1988). It is also unclear what the effect of a unit root in ln (o,2) is on
asymptotic properties of the parameter estimates. The (unverified) regularity conditions for asymp-
totic normality require that the scoring function and hessian obey a central limit theorem and
uniform weak law of large numbers respectively, which may or may not require {t2, (,} to have f
moments. It may be that the standard asymptotics are valid even in the presence of a unit root in
In (GT2).

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 361

of the GED(v) distribution in Section 2 that if v < 2, the distribution of zt (and

therefore the conditional distribution of {t) has thicker tails than the normal
distribution. The estimated v is approximately 1.58 with a standard error of
about 0.03, so the distribution of the zt is significantly thicker-tailed than the
normal.
(v) The estimated contribution of nontrading days to conditional variance is
roughly consistent with the results of French and Roll (1986). The estimated
value of 5 is about 0.183, with a standard error of about 0.028, so a nontrading
day contributes less than a fifth as much to volatility as a trading day.
The general results just discussed are quite robust to which ARMA model is
selected, though of course the parameter estimates change somewhat. The
results also appear to be quite robust with respect to the sample period: another
paper, Nelson (1989), reports strikingly similar parameter estimates in an
exponential ARCH model fit to daily capital gains on the Standard 90 stock
index from 1928 to 1956.

Specification Tests

To test the fit of the model, several conditional moment tests (Newey (1985))
were fit using orthogonality conditions implied by correct specification. Correct
specification of the model has implications for the distribution of {zt. For
example, E[zt] = 0, E[z72] = 1, and E[g(zt)] = 0. Since the GED distribution is
symmetric, we also require that E[zt lzt I]= 0. The first four orthogonality
conditions test these basic properties. Correct specification also requires that
{[2 - ayt2} and {(t} (or equivalently {Z2 _ 1) and {zt}) are serially uncorrelat
Accordingly, we test for serial correlation in zt and Z2 at lags one through five.
Table IV reports first the sample averages for the fourteen selected orthogo-
nality conditions and their associated t statistics, and then chi-square statistics
and probability values for two combinations of the orthogonality conditions, the
first including only conditions relating to correct specification of the conditional
variance process 7t and the second also testing for correct specification of the
conditional mean process.
In the first chi-square test, the CRSP model does extremely well, with a
probability value of 0.94. Considered individually, none of the first nine orthogo-
nality conditions are significantly different from zero at any standard signifi-
cance level. When the last five conditions, which test for serial correlation in
{Zt}, are included, the probability value drops to 0.16, which still does not r
at any standard significance level, although statistically significant serial correla-
tion is found at lag two. Overall, the fit of the CRSP model of the conditional
variance process {o,2} seems remarkably good.5' 16

15 Engle, Lilien, and Robbins (1987) and Pagan and Sabau (1987) based conditional moment tests
on {(0 rather than on {z,}. Basing tests on {zj} is analogous to a GLS correction and seems likely to
increase the power of the specification tests. As a check, chi-square tests were recomputed using {(,)
instead of {zt). The test statistics were drastically lower in each instance.
16 The specification tests for the 1928-1956 Standard 90 capital gains data reported in Nelson
(1989) were not as favorable: the tests found evidence of negatively skewed returns and serially
correlated residuals, rejecting the model at any standard level.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
362 DANIEL B. NELSON

TABLE IV

SPECIFICATION TEST RESULTS FOR CRSP VALUE-WEIGHTED EXCESS RETURNS,

ARMA(2, 1) MODELa

Orthogonality Sample
Conditions Averages t Statistics

1 E(zt)=0 -0.007 -0.349

2 E(Z2) _ 1 = 0 -0.001 -0.037
3 ElzI -A21/'T(2/V)/F(l/V)=0 2.583 10-5 0.002
4 E[zt (Iztl -EIzI)]=0 -0.027 -1.254
5 E[( Z2 _ 1)(Z72 1-1)] = 0 0.102 1.023
6 E[(72 - 1)(z2-2-1)] = 0 0.028 0.529
7 E[( Z2 _ 1)(Z23 - 1)] = 0 0.079 0.864
8 -E[Z 2 _ 1)(z 4 - 1)] = 0 0.042 0.947
9 E[(72 - 1)(z25- 1)]= 0 0.019 0.469
10 E(zt 0zt-1) = . 0.022 0.949
11 E(zt Zt-2) = 0 0.034 -2.521
12 E(zt Zt-3)= 0 0.018 1.377
13 E(zt Zt-4)= 0 0.015 1.174
14 E(Zt -Z_5)= 0 0.020 1.563

dx2 statistic for conditional moment test us

degrees of freedom, the probability value = 0.94. x2 statistic for conditional moment test using all
14 orthogonality conditions = 19.14. With 14 degrees of freedom, the probability value = 0.16.

Our conditional moment tests leave many potential sources of misspecifica-

tion unchecked. It therefore seems desirable to check the forecasting perfor-
mance of the model during periods of rapidly changing volatility, and to check
for large outliers in the data. From Figure 2, two periods stand out as times of
high and rapidly changing volatility: the market break of September, 1973-De-
cember, 1974, and the last five months of 1987. Figures 4 and 5 plot returns and
the one-day-ahead ex ante 99% prediction intervals implied by the estimated
model for these periods. The 99% prediction intervals have a width of approxi-
mately 5.56 (Tt.
In 1973-1974, the model seems to track the change in volatility quite closely,
with no serious outliers (Figure 4). The model also succeeds quite well in
picking up the volatility of the period after October 19, 1987, with no serious
outliers (Figure 5). On October 19, 1987, however, the model's performance is
mixed at best: the ex ante prediction intervals for the day are approximately
+ 7%, the widest in the data set up to that time, brought on by the sharp drops
in the market during the preceding week. Unfortunately, the drop in the index
that day was approximately 20.25%, about 7.78 -at, a serious outlier. If the
estimated model were literally true, the expected number of outliers of this size
or greater in a 252 year period (the length of the data set) is only about 1. 10'
-i.e., the probability of observing such an outlier in a data set of this length is
extremely low.
Although October 19, 1987 is the most extreme outlier in the sample, it is not
the only large outlier. Table V lists the largest, ranked in order of the implied

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 363

The Crash of 1973-1 974

LOl

Lnfl
0

0
Lnf

* Dates: September 19, 1973

through December 31, 1974

0, ,

0 0 25 50 75 1 00 1 50 200 250 300

FIGURE 4

The Crash of 1 987

0~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

L()
0

-7 Daites: August 1 1 through

0 0 25 50 ~~~~~75 100
FIGURE 5

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 364 DANIEL B. NELSON

TABLE V

LARGEST OUTLIERS IN THE SAMPLE

Date R,(%) ,(%) z, Expected Frequencya

10/19/87 -20.25 2.44 -7.78 1.11 10-5

9/11/87 - 4.58 0.84 - 5.38 0.05
8/17/82 4.36 0.87 4.95 0.20
1/8/86 - 2.42 0.57 - 4.81 0.31
8/2/78 2.23 0.48 4.53 0.67
8/16/71 3.50 0.82 4.33 1.21
10/9/79 - 3.37 0.73 -4.29 1.36
7/7/86 - 3.02 0.72 - 4.12 2.20

dThe expected number of Iz, I values of this size or g

year sample (6408 observations).

value for Iz, I. The standardized GED has only one parameter, v, to control
shape of the conditional distribution, and this may well not be flexible
enough-i.e., there are too many "large" IZ,J values. Nonparametric met
(as in Engle and Gonzailez-Rivera (1989)), or more flexible parametric families
of distributions, would probably improve the model.

4. CONCLUSION

This paper has presented a new class of ARCH models that do not suffer
from some of the drawbacks of GARCH models. Ideally, we would like ARCH
models that allow the same degree of simplicity and flexibility in representing
conditional variances as ARIMA and related models have allowed in represent-
ing conditional means. While this paper has made a contribution to this end, the
goal is far from accomplished: it remains to develop a multivariate version of
exponential ARCH, and a satisfactory asymptotic theory for the maximum
likelihood parameter estimates. These tasks await further research.

Graduate School of Business, University of Chicago, 1101 E. 58th St., Chicag

IL 60637, U.S.A.

Manuscript received February, 1989; final revision received February, 1990.

APPENDIX I: THE MOMENTS OF Ot2 AND 6t

By (2.1) and the independence of the {zt), the joint moments and conditional moments of
and {(t) take either the form

(A1.1) E[exp (at + big(zt,))] =exp(at)11E[exp(b1g(zt,,))] or

(A1.2) E[ZPkZ'J exp (at + EbEzte)) ] [ >l, e (bg(zt,))]

exp (at) * E[ztp-k exp (bkg(zt-k))]

E E[ Ztq, exp (b,g(zt_,))]

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 365

for nonnegative integers p, q, j, and k /=j. For exampl

we set {a,} = {a,} and {b,} = {f},l in (Al.1). For t
Zt-, Zt-k-1 .. ., set

a,=a,+ E,lIg(z1_1), b1=,/3 for 1sis<k-1,and b=Ofori>k.

i=k

To obtain the moments of 6, and 6t-k rt , we proceed similarly, using (A1.2).

To evaluate the expectations in (A1.1)-(A1.2), we must make a further distributional assumption
about {z1}. When {z1} i.i.d. N(O, 1), the following result, combined with (A1.1)-(A1.2), gives the
joint conditional and unconditional moments of the o(r2 and 6, processes:

THEOREM Al.1: Let z N(O, 1). For any finite, real scalar b and positive integer p,

(A1.3) E[exp (g(z)b)] = {P(yb + Ob) exp [ b2(0 + y)2/2]

+'P(yb - Ob) exp [b2(y - 0)2/2] }exp [ -by(2/7T)1/2] oo

and

(Al.4) E[zP exp (g(z)b)] =exp [-by(2/7 )1/12] - (p +1) (27r) -1/12
{exp [b2(0 + y)2/4] D-(P+1)[-b(y + 0)]

+(- 1)pexp [b2(y _ 0)2/4] D-(P+ )[-b(y - 0)]) <oo,

where P() is the standard normal cumulative distribution function, F[ ] is the Gamma function, and
Dq[ ] is the parabolic cylinder function (Gradshteyn and Ryzhik (1980)).

PROOF OF THEOREM Al.l: (A1.3) follows by straightforward but tedious calculus. (A1.4) is easily
proven with the help of Gradshteyn and Ryzhik (1980) Formula 3.462 #1. The finiteness of the
expression in (A1.4) follows as a special case in the proof of Theorem A1.2 below. Q.E.D.

Theorem A1.2 deals with the more general GED case.

THEOREM A1.2: Let p be a nonnegative integer, and let z GED(v) with E(z) = 0, Var(z) = 1,
and v > 1. Then

(Al.5) E[zP exp (g(z)b)]

=exp [-byF(2/v)A21/JIF(1/v)] . 2P/u AP

E (21/vAb)k[(y + 1)k + (-1( p +) I 2F(llv)F(k + 1)

If z GED (v) with v < 1, or z Student t with d degrees of freedom (d > 2), and z is normalized to
satisfy E(z) = 0, Var(z) = 1, then E[exp(g(z)b)] and E[zP exp(g(z)b)] are finite if and only if

(Al.6) by+ lbO I 0.

If z GED(1), then E[exp(g(z)b)] and E[zP exp(g(z)b)] are finite if and only if

(A1.7) by+ IbOl <21/2

The restriction (A1.6) is rarely satisfied in practice. In computing the unconditional expect
of o-t2, b is one of the moving average coefficients {/,3J, at least some of which are positi
,80 - 1. If b > 0, (A1.6) implies either y < 0 or y = 0 = 0. If y < 0, residuals larger than
decrease conditional variance, which goes against the intuition developed in Section 1. In the
author's experience in fitting exponential ARCH models, the estimated value of y is always positiv

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 366 DANIEL B. NELSON

PROOF OF THEOREM A1.2: The density of z given in (2.4) and Gradshteyn and Ryzhik (1980)
Formula 3.381 #4 yield

(A1.8) E(z I) = A21/'F(2/ )/1F(1/V).

Straightforward calculus then yields

(A1.9) E[ ZP exp (g(z)b)]

= 2' /; exp[-byA21/'F(2/c)/F(1/L)]
2F(1/u')

l (- +(P+ 1)/I [eby(9+y)(2)b/ + ( _ bA(-0(2 ] dy

Expanding the part of the integrand in square brackets in a Taylo

(A1.10) E[ zP . exp (g(z)b)]

2P/"AP
) exp [ -bA21/'F(2/c)/F(1/tu)]

f = [[( +y)21/'bA Ok+ (-l )P[(y y- )21/'bA ] F(k+ 1) dy.

If we can interchange the order of summation and integration in (A1.10), then Gradshteyn and
Ryzhik (1980) Formula 3.381 #4 yields (A1.5). First, consider the related expression

00 k e- Yy- 1+(p+ l+k)lt

(A111) fJ E 2 F(k+1) dy, where
o k=O Fk1

(A1.12) A max { |(0 + y)bA2'/" 1, 1(y - O)bA21/" |} .

The terms in (A.11) are nonnegative, so by monotone convergence (Rudin (197
the order of integration and summation in (A1.11) can be reversed. If the int
finite, then by dominated convergence (Rudin (1976, Theorem 11.32)), we can inte
of summation and integration in (A1.10) and the integral in (A1.10) is finit
therefore, it remains only to show that (A1.11) is finite if v > 1. By Gradshtey
Formula 3.381 #4,

00 0 e-y- 1(p+lk)lt 002,AkF[(p + 1 + k)/v]

(A1.13) o
fok=0
E 2AF(?1
k (k dy
kO= Fk?l
E F(k + 1)
By construction, A > 0. When A = 0, the converg
root test (Rudin (1976, Theorem 3.33)), the sum

(A1.14) limsup ln (A) + k-1 ln (F[(k + 1 +p)/v]) - k-1 ln (F(k + 1)) < 0.
k -oo

By the asymptotic expansion for ln(F(x)) in Davis (1965, Equation 6.1.41),

(A1.15) limsup ln (A) + k-1 In (F[(k + 1 +p)/v]) - k1 n (F(k + 1))

k- oo

=limsup ln (A) + k-'[(k + 1 +p)/v - 1/2] ln (k + 1 + p)

k -oo

+(k+ 1)/k -k-'[(k+ 1 +p)/v- 1/2] ln(v)

- k- l[k + 1/2] ln (k + 1) - (k +p + 1)7kv + O(k-1).

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 367

Expanding ln(k + 1 +p) in a Taylor series around p = 0 and substituting into (A.l15),

(A1.16) lim sup In (A) + k -in (F[(k + 1 +p)/v]) - k-' In (F(k + 1))
k -oo

= lim sup ln (A) -k-[(k + 1 +p)/v - 1/2] ln (v) - (k +p + 1)7kv

k--oo

+ 1 + O(k-1) - [(k+1 +p)/kv- 1 -k-1] ln (k + 1).

The last term on the right-hand side of (A1.16) asymptotically dominates every other term if v # 1,
diverging to - oo if v > 1, so the series converges when v > 1 and A < oo.
If z Student t, or GED with v < 1, it is easy to verify that E[exp(b g(z))] = oo unless
(y + O)b < 0 and (y - H)b < 0, which holds if and only if by + I Ob I < 0. If this inequality is satisfi
E[exp(g(z)b] < oo. The GED case with u = 1 is similar. Q.E.D.

APPENDIX II: PROOFS

PROOF OF THEOREM 2.1: That I In (r7 - a, I is finite almost surely when Ek <cX fo
immediately from the independence and finite variance of the g(z,) terms in (2.1) and
Billingsley (1986, Theorem 22.6). Since Iln((ro2) - a,I is finite almost surely, so are exp(-a,)o-
exp( -a,/2)g,. This, combined with Stout (1974, Theorem 3.5.8) and the representation in (1.
and (2.1)-(2.2) implies that these series are strictly stationary and ergodic. For all t, the expectat
of (ln(or72)-a,)=0, and the variance of (ln(o-r2)-a,) is Var(g(z1))E7'k 32. Since Var(g
finite and the distribution of In ((rt) - a, is independent of t, the first two moments of (In (crt
are finite and time invariant, so {ln (cr2) - a,} is covariance stationary.
If E7k= l = cc, then I In (cr2) - a, = X almost surely by Billingsley (1986, Theorems 22.3
22.8). Q.E.D.

PROOF OF T
and any existing moments are time invariant. We will show that exp(-a1)o-12 and exp(-a1/2)g1
have finite moments of arbitrary positive order. As shown in Appendix I, the conditional, uncondi-
tional, and cross moments of exp(-a1)o-r2 and exp(-a1/2)g1 have the form (A1.1)-(A1.2). By
H6lder's Inequality, if (rt2 and gr have arbitrary finite moments, the cross moments are also finite.
By the independence of the z1, E[H(exp(-at/2)grt)Id] =E IzIdE[(exp(-at/2)ort)d]. Since z has
arbitrary finite moments, we need only show that E[(exp(-a1/2)or1)d] is finite for all d > 0 if {,8i} is
square-summable. This expectation is given by
00

(A2.1) E[(exp(at/2)ort)d] = fEexp[idP1g(zt_,)],

where the individual expectation terms in (A2.1) are obtained by setting b= 11d/2 in Theorem
Al.1.
A sufficient condition for an infinite product H1 ,-a, to converge to a finite, nonzero number is
that the series ,=. 1jIal - 1 I converge (Gradshteyn and Ryzhik (1980, Section 0.25)). Let a, equal
the ith term in (A2.1). Define

(A2.2) S(,lB) -exp ['d,8yAF(21v)2'1"'1F(l1v)] * E(exp [g(z),B' d]) .

We then have

(A2.3) S(0) = 1, S'(0) = 'dyAF(2/v)2l/'/F(1/v), and

S"(,/3)=0(1) as 3--l0.

Expanding S(,/) and exp[- 'd,8yAF(2/v)2'17/F(1/v)] in Taylor series around

tuting into (A1.5), we have

(A2.4) a 1 =E(exp [ g(z),I12d]) - 1

= [1 - 'd/31yAF(2/v)2//F(l1/v) + o(p2)]

1 + 'd/31yAF(2/v)21//F(l/v) + O(82)] -1

= ( ,2) as /8 0.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 368 DANIEL B. NELSON

0(132) in (A2.4) means that for so

(A2.5) sup 7-21o(137)I<M.

i13,i <E,13,=#O

By (A2.4)-(A2.5), E/32 < X implies Ei= a1 - 1 < cx and thus H11 =a, < c.
Finally, we must prove (2.6)-(2.7). The proofs of (2.6) and (2.7) are substantially identical, so we
prove only (2.6). By Theorem A1.1,

(A2.6) E[exp (-pa, )or12Pjz0, Z_ , Z-2 *... ] -E[exp(-pa, )o12P]

= [exp (p 1 jg(zt)) - I1E[exp(p38g(zt,_))]]

t- 1

HjE[exp (pf31g(z_1))].

The last term on the right-hand side of (A2.6) is finite by Theorem A1.2. (2.6) will therefore be
proven if we can show that

(A2.7) plim [exp (pE3jg(z,,J) - H E[exp(p,8jg(zt_i))] =0.

t ??- 0 J=t j=t

First, consider the unco

We have

00 "O~~~~~~~~~~~~~0

(A2.8) Var[EpPjg(zt_) =p2Var(g(z,)) E p y2 O as t oo.

Since convergence in L2 implies convergence in probability,

(A2.9) exp [PpJ g ( zt] t - 1

Finally,
00 0

(A2.10)

= ex

=exp [ 0(B)] as t-oo

by (A2.4)-(A2.5) and the square summability of {,3j}. Q.E.D.

REFERENCES

ABEL, A. B. (1988): "Stock Prices Under Time-Varying Dividend Risk: An Exact Solution in an
Infinite-Horizon General Equilibrium Model," Journal of Monetary Economics, 22, 375-393.
AKAiKE, H. (1973): "Information Theory and an Extension of the Maximum Likelihood Principle,"
in Second International Symposium on Information Theory, ed. by B. N. Petrov and F. Csaki.
Budapest: Akademiai Kiado, 267-281.
BACKUS, D., AND A. GREGORY (1989): "Theoretical Relations Between Risk Premiums and
Ccnditional Variances," Mimeo, Federal Reserve Bank of Minneapolis and Queen's University.
BARCLAY, M. J., R. H. LITZENBERGER, AND J. WARNER (1990): "Private Information, Trading
Volume, and Stock Return Variances," Review of Financial Studies, 3, 233-254.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 CONDITIONAL HETEROSKEDASTICITY 369

BARSKY, R. B. (1989): "Why Don't the Prices of Stocks and Bonds Move Together?" American
Economic Review, 79, 1132-1145.
BERNANKE, B. S. (1983): "Irreversibility, Uncertainty and Cyclical Investment," Quarterly Journal of
Economics, 98, 85-106.
BILLINGSLEY, P. (1986): Probability and Measure, Second Edition. New York: John Wiley and Sons.
BLACK, F. (1976): "Studies of Stock Market Volatility Changes," 1976 Proceedings of the American
Statistical Association, Business and Economic Statistics Section, 177-181.
BOLLERSLEV, T. (1986): "Generalized Autoregressive Conditional Heteroskedasticity," Journal of
Econometrics, 31, 307-327.
BOLLERSLEV, T., R. F. ENGLE, AND J. M. WOOLDRIDGE (1988): "A Capital Asset Pricing Model with
Time Varying Covariances," Journal of Political Economy, 96, 116-131.
Box, G. E. P., AND G. C. TIAO (1973): Bayesian Inference in Statistical Analysis. Reading, MA:
Addison-Wesley Publishing Co.
CHou, R. Y. (1987): "Volatility Persistence and Stock Returns-Some Empirical Evidence Using
GARCH," Journal of Applied Econometrics, 3, 279-294.
CHRISTIE, A. A. (1982): "The Stochastic Behavior of Common Stock Variances: Value, Leverage
and Interest Rate Effects," Journal of Financial Economics, 10, 407-432.
CLARK, P. K. (1973): "A Subordinated Stochastic Process Model with Finite Variance for Specula-
tive Prices," Econometrica, 41, 135-155.
COCHRANE, J. (1988): "How Big is the Random Walk in GNP?" Journal of Political Economy, 96,
893-920.
Cox, J., J. INGERSOLL, AND S. Ross (1985): "An Intertemporal General Equilibrium Model of Asset
Prices," Econometrica, 53, 363-384.
DAVIS, P. J. (1965): "The Gamma Function and Related Functions," in Handbook of Mathematical
Functions, ed. by M. Abramowitz and I. A. Stegun. New York: Dover Publications.
ENGLE, R. F. (1982): "Autoregressive Conditional Heteroskedasticity With Estimates of the Vari-
ance of United Kingdom Inflation," Econometrica, 50, 987-1008.
ENGLE, R. F., AND T. BOLLERSLEV (1986a): "Modelling the Persistence of Conditional Variances,"
Econometric Reviews, 5, 1-50.
(1986b): "Modelling the Persistence of Conditional Variances: A Reply," Econometric
Reviews, 5, 81-87.
ENGLE, R. F., D. M. LILIEN, AND R. P. ROBINS (1987): "Estimating Time Varying Risk Premia in
the Term Structure: The ARCH-M Model," Econometrica, 55, 391-408.
ENGLE, R. F., AND G. GONZA'LEz-RIVERA (1989): "Semiparametric ARCH Models," U.C.S.D.
Economics Department Discussion Paper 89-17.
FAMA, E. F. (1965): "The Behavior of Stock Market Prices," Journal of Business, 38, 34-105.
FRENCH, K. R., AND R. ROLL (1986): "Stock Return Variances: The Arrival of Information and the
Reaction of Traders," Journal of Financial Economics, 17, 5-26.
FRENCH, K. R., G. W. SCHWERT, AND R. F. STAMBAUGH (1987): "Expected Stock Returns and
Volatility," Journal of Financial Economics, 19, 3-29.
GENNOTTE, G., AND T. A. MARSH (1987): "Variations in Economic Uncertainty and Risk Premiums
on Capital Assets," Mimeo, U.C. Berkeley Graduate School of Business.
GEWEKE, J. (1986): "Modelling the Persistence of Conditional Variances: A Comment," Economet-
ric Reviews, 5, 57-61.
GLOSTEN, L. R., R. JAGANNATHAN, AND D. RUNKLE (1989): "Relationship Between the Expected
Value and the Volatility of the Nominal Excess Return on Stocks," Banking Research Center
Working Paper #166, Northwestern University.
GRADSHTEYN, I. S., AND I. M. RYZHIK (1980): Table of Integrals, Series and Products. New York:
Academic Press.
HANNAN, E. J. (1980): "The Estimation of the Order of an ARMA Process," The Annals of
Statistics, 8, 1071-1081.
HARVEY, A. C. (1981): The Econometric Analysis of Time Series. Oxford: Philip Allan.
HOSKING, J. R. M. (1981): "Fractional Differencing," Biometrika, 68, 165-176.
Lo, A., AND C. MACKINLAY (1988): "Stock Market Prices Do Not Follow Random Walks: Evidence
from a Simple Specification Test," Review of Financial Studies, 1, 41-66.
MANDELBROT, B. (1963): "The Variation of Certain Speculative Prices," Journal of Business, 36,
394-419.
McDONALD, R. M., AND D. SIEGEL (1986): "The Value of Waiting to Invest," Quarterly Journal of
Economics, 101, 707-728.
MERTON, R. C. (1973): "An Intertemporal Capital Asset Pricing Model," Econometrica, 41,
867-888.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms
 370 DANIEL B. NELSON

NELSON, D. B. (1988): "The Time Series Behavior of Stock Market Volatility and Returns,"
Unpublished Doctoral Dissertation, M.I.T. Economics Department.
(1989): "Modelling Stock Market Volatility Changes," 1989 Proceedings of the American
Statistical Association, Business and Economic Statistics Section, 93-98.
(1990a): "Stationarity and Persistence in the GARCH(1, 1) Model," Econometric Theory, 6,
318-334.
(199Gb): "Filtering and Forecasting with Misspecified ARCH Models I: Getting the Right
Variance With the Wrong Model," forthcoming, Journal of Econometrics.
NEWEY, W. K. (1985): "Generalized Method of Moments Specification Testing," Journal of
Econometrics, 29, 229-256.
PAGAN, A. R., AND Y. HONG (1988): "Non-Parametric Estimation and the Risk Premium,
forthcoming in Nonparametric and Semiparametric Methods in Econometrics, ed. by W. Barnett,
J. Powell, and G. Tauchen. Cambridge: Cambridge University Press.
PAGAN, A. R., AND H. C. L. SABAU (1987): "Consistency Tests for Heteroskedastic and Risk
Models," Mimeo, University of Rochester Economics Department, October, 1987.
PANTULA, S. G. (1986): "Modelling the Persistence of Conditional Variances: A Comment,"
Econometric Reviews, 5, 71-73.
POTERBA, J. M., AND L. H. SUMMERS (1986): "The Persistence of Volatility and Stock Market
Fluctuations," American Economic Review, 76, 1142-1151.
RUDIN, W. (1976): Principles of Mathematical Analysis. New York: McGraw-Hill Book Company.
SCHOLES, M., AND J. WILLIAMS (1977): "Estimating Betas from Nonsynchronous Data," Journal of
Financial Economics, 5, 309-327.
SCHWARZ, G. (1978): "Estimating the Dimension of a Model," The Annals of Statistics, 6, 461-464.
SCHWERT, G. W. (1989a): "Business Cycles, Financial Crises and Stock Volatility," Carnegie-Roches-
ter Conference Series on Public Policy, 31, 83-126.
(1989b): "Why Does Stock Market Volatility Change over Time?" Journal of Finance, 44,
1115-1154.
STOUT, W. F. (1974): Almost Sure Convergence. London: Academic Press.
WEISS, A. A. (1986): "Asymptotic Theory for ARCH Models: Estimation and Testing," Econometric
Theory, 2, 107-131.
WIGGINS, J. B. (1987): "Option Values Under Stochastic Volatility: Theory and Empirical Esti-
mates," Journal of Financial Economics, 19, 351-372.

This content downloaded from

93.180.40.214 on Tue, 11 Mar 2025 19:21:51 UTC
All use subject to https://about.jstor.org/terms