JOURNALOF

Econometrics
ELSEVIER Journal of Econometrics 64 (1994) 307-333

Autoregressive conditional heteroskedasticity

and changes in regime
James D. Hamilton*-a, Raul Susmelb

(Received August 1992; final version received September 1993)

Abstract

ARCH models often impute a lot of persistence to stock volatility and yet give
relatively poor forecasts. One explanation is that extremely large shocks, such as the
October 1987 crash. arise from quite different causes and have different consequences for
subsequent volatility than do small shocks. We explore this possibility with U.S. weekly
stock returns, allowing the parameters of an ARCH process to come from one of several
different regimes, with transitions between regimes governed by an unobserved Markov
chain. We estimate models with two to four regimes in which the latent innovations come
from Gaussian and Student f distributions.

Key ~lords: ARCH models; Stock prices; Regime-switching models; Volatility

JEf. class+ztion: C22; Cl2

1. Introduction

Financial markets sometimes appear quite calm and at other times highly
volatile. Describing how this volatility changes over time is important for two
reasons. First, the riskiness of an asset is an important determinant of its price.
Indeed, empirical estimation of the conditional variance of asset returns forms

* Correspondmg author.
We are grateful to the NSF for support under grant number SES-8920752. Data and software used
in this study can be obtained at no charge by writing James D. Hamilton; alternatively, data and
software can be obtained by writing ICPSR, Institute for Social Research, P.O. Box 1248, Ann
Arbor, MI 48106.

0304-4076/94/$07.00 (0 1994 Elsevier Science S.A. All rtghts reserved

SSDI 030440769301587 C
the core of several hundred studies of financial markets surveyed by Bollerslev,
Chou, and Kroner (1992). Second, efficient econometric inference about the
conditional mean of a variable requires correct specification of its conditional
variance. Studies suggesting the practical relevance of this issue include Morgan
and Morgan (1987), Bera, Bubnys and Park (1988) Pagan and Ullah (1988)
Connolly (1989) Diebold, Lim, and Lee (1993) and Schwert and Seguin (1990).
One popular approach to modeling volatility is the autoregressive condi-
tional heteroskedasticity (ARCH) specification introduced by Engle (1982).
Bollerslev, Chou, and Kroner (1992) characterized this general class of models as
follows. Suppose that a variable u, is governed by

u, = 6,’ o,, (1.1)

where (II,) is an i.i.d. sequence with zero mean and unit variance. The conditional
variance of u, is specified to be a function of its past realizations:
cr: = y(u,_ 1, u,-2, ... ). (1.2)
Often it is assumed that 11,_ N(0, 1) and that g( .) depends linearly on the past
squared realizations of u:

~: = Uo + ~ UiU:_i + ~ hia:_i. (1.3)

i=l i=l

This is a Gaussian GARCH(p, 4) specification introduced by Bollerslev (1986);

when p = 0 it becomes the ARCH(q) specification of Engle (1982).
Section 2 discusses some of the shortcomings of such models as a description
of the volatility in stock return data. We argue that a promising alternative is to
allow for the possibility of sudden, discrete changes in the values of the param-
eters of an ARCH(q) process as in the Markov-switching model in Hamilton
(1989). Markov-switching ARCH models have previously been employed in
Brunner’s (1991) study of inflation and Cai’s (forthcoming) analysis of Treasury
bill yields. Section 3 of this paper motivates a parsimonious parameterization of
a Markov-switching ARCH model which differs from that used by these re-
searchers, and develops optimal forecasts for this specification. Section 4 pre-
sents an empirical application, which also differs from previous studies in
allowing up to four different regimes. Another innovation is that stock returns
within any given regime are modeled with a Student t distribution rather than
a Normal distribution. We argue that this class of models better describes
a number of key features of the data.

2. Stock market volatility and traditional ARCH specifications

The stock price series used in this analysis is the value-weighted portfolio of
stocks traded on the New York Stock Exchange contained in the CRISP data
 J.D. Hamilton, R. SusmeliJournal of Econometrics 64 (1994) 307-333 309

tapes. The raw data are returns from Wednesday of one week to Tuesday of the
following week. The original data start with the week ended Tuesday, July 3,
1962 and end with the week ended Tuesday, December 29, 1987. All results
reported in this paper condition on the first four observations and begin
estimation with the week ended Tuesday, July 31,1962. Thus T 5 1327 observa-
tions were used to estimate each model.
Let y, denote the weekly stock return measured in percent; for example,
y, = - 2.0 means that stock prices fell 2% in week t. In all of the models we
investigated, the process u, that is described by the ARCH process is the residual
from a first-order autoregression for stock returns:

yt=X+4y,-,+u,. (2.1)
By far the most popular ARCH model that has been used to describe financial
market volatility is the GARCH (1, 1) specification.’ For our stock return data,
we make two modifications to the usual GARCH (1, 1) specification which have
received support in other studies and which clearly improve the model’s ability
to describe the weekly stock return series used in our study.
First, we follow a suggestion by Bollerslev (1987) and Baillie and DeGennaro
(1990) and treat u, in (1 .l) as drawn from a Student t distribution with v degrees
of freedom (normalized to have unit variance). The implied conditional density
for u, is then:

r((r + 1)/2)
.f(UtIUt-l,Ut~Z,...)= &r(,,,2) (1, - 2)- “2 a; ’

2 1)/2
ur

2)

These

’
a generally a wide
where d,_, is a dummy variable that is equal to zero if uI_, > 0 and equal to
unity if u,_ r d 0. The leverage effect predicts that c > 0.
Our maximum likelihood estimates for this specification based on weekly
stock returns are as follows:2

y1 = 0.3 1 + 0.27 !‘I , + II,, (2.4)

(0.05) (0.03)

C$ = 0.28 + 0.07 IA:_, + 0.23 d,_ 1 . uf_, + 0.78 rr:_,, (2.5)

(0.12) (0.03) (0.06) (0.06)

C = 5.6. (2.6)
(0.8)

The numbers in parentheses are the usual ‘asymptotic standard errors’ based on
the matrix of second derivatives of the log-likelihood function. Table 1 presents
some summary statistics for this model and compares it with others we have
estimated.
All of the parameters are highly statistically significant on the basis of either
Wald tests or likelihood ratio tests, suggesting that this model offers a clear
improvement over the null hypothesis of homoskedastic errors. However, earlier
researchers have raised several concerns about the adequacy of such a specifica-
tion, as we now discuss.

2.1. Forecasting pe~fi,rmance

If the specification were correct and the parameters were known with cer-
tainty, then a: would be the conditional expectation of u:. Hence a mean
squared error loss function,

Ejb: - R)21ut~1,u1-2,... ),

would be minimized with respect to 8, by choosing 0, = 0: for 0: given by (2.5).

The average value of (6: - c?:)~ turns out to be given by

MSE = T-l 5 (12; - 6:)’ = 815.6. (2.7)

I=,

’ Generating the sequence of variances ’(mt

* I in (2.5) requires knowledge of the starting value a;. For
the reported results, 06 was treated as a separate parameter which was estimated by maximum
likelihood along with the others. We also fit the models following Bollerslev’s (1986) suggestion of
setting r~i equal to the average value of ti:. with very similar results.
Table I
Summary statistics for various specifications

No. of Log- Degrees of Persistence

Model parameters likelihood AIC Schwarz freedom (j.)

Constant variance 2 3095.2 - 3097.2 - 3102.4

Gaussian GARCH (I, I) s* _ 2944.7 - 2949.1 - 2962.7 0.99

(2 X IO_‘)
Student t GARCH-L(I, I) 7* - 2822.0 _ 2829.0 - 2847.2 5.6 0.96
(5 x 10-Y (0.83)

GED GARCH-L( I. I) I’ 2839.8 - 2846.X 2865.0 1.2 0.96

(0.057)

Gaussian ARCH-L(2) 6 - 2934.4 - 2940.4 - 2956.0 0.74

Gaussian SWARCH-L(2,2) 9 - 2836.3 - 2845.3 - 2868.7 0.42

(2 x 10-0)

Student t ARCH-L(2) - 2846.0 - 2853.0 - 2871.2 4.7 0.71

(0.61)

Student t SWARCH-L(2.2) IO ~ 2818.0 - 2828.0 2X54.0 5.7 0.59

(2 X lo-“) (0.83)

Student t SWARCH-L(3, 2) 13 - 2802.7 - 2815.7 2849.4 7.2 0.48

(I X 10-h) (1.4)
Student t SWARCH-L(4,2) I5 - 2798.1 - 2813.1 _ 2852.0 8.7 0.50
(0.01) (2.0)

The count of the number of parameters attributed to the GARCH spectfications does not include the estimate of the initial variance 06. The count of the
number of parameters for the SWARCH-L(3, 2) and SWARCH-L(4, 2) specifications does not include the transition probabilities p,, imputed to be zero.
The second column reports Y*, the maximum value achieved for the log of the likelihood function. The number in parentheses below each entry reports
what the p-value for a likelihood ratio test of that model against the preceding specification would be under the assumption that twice the difference in
log-likelihoods is distributed x2 with degrees of freedom equal to the difference in number of parameters between the null and alternative.
AIC was calculated as Y* - k for k the number of parameters in column I.
Schwarz was calculated as Y* - (k, 2) *In(r) for T = 1327.
The degree of freedom parameter is the magnitude v in expression (2.2) for the Student t distribution or the parameter L for the GED distribution [see
Eq. (2.4) in Nelson, 19911. The standard error for this parameter is in parentheses.
The persistence parameter i IS characterized by expression (2.1 I) for a GARCH( I, I) and by the largest eigenvalue of the matrix in (3.16) for an
ARCH-L(Z, 2) or SWARCH-L(2. 2) model.

r- -,- __ ., _.
312 J.D. Hamilton, R. SusmeliJournd of Econommics 64 11994) 307 -333

For a standard of comparison, let j denote the sample average return and s2
the unconditional sample variance:

j= T-’ = T-’ i (yt - j)2.

t=1 1=1

If we simply forecast the variance of y, to be the constant s2 throughout the

sample, the corresponding value for the loss function is

MSE = T-’ i ;(yt - j)’ - s2j2 = 757.9. (2.8)

r=1

Thus the Student t GARCH-L( 1, 1) model actually yields poorer forecasts than
just using a constant variance. The ‘R”, or percent improvement of (2.7) over
(2.8), is - 0.08.
The average squared forecast error is probably an unfair standard for judging
this specification, since it is based on fourth moments of the actual data yr. The
unconditional fourth moment would fail to exist if (2.5) were the data-generating
process. The third row of Table 2 compares the forecast cr: with the uncondi-
tional sample variance s2 using three alternative loss functions.” The GARCH-
L (1, 1) does almost as poorly in terms of mean absolute error, but somewhat
better for other loss functions.
Tables 1 and 2 also report results for a number of other simple ARCH models.
The Student t does better than the generalized error distribution suggested by
Nelson (1991) and is vastly superior to a conditionally Gaussian model, both in
terms of forecasting performance and in terms of a classical test of the null
hypothesis of Gaussian residuals against the alternative of Student t. The leverage
effect [the parameter < in Eq. (2.3)] is likewise highly statistically significant and
helps improve the forecasts, and the GARCH(1, 1) yields much better forecasts
than low-order ARCH models. Thus, despite its weak forecasting performance,
the model in (2.4)-(2.6) is the best representative that we have found for describing
these stock return data out of the class of models given in (l.l)-(1.3).

2.2. Persistence

Our second concern with models such as that in (2.4))(2.6) is the high degree
of persistence they imply for stock volatility. To calculate a measure of this
persistence, replace ‘r’ with ‘t + m‘ in Eq. (2.3) and write the result as

E(u:+,lu t+mm1> 4+m-27 ... ) = a0 + a, -d+,n-1 + 5-4+m-1’~~:+m-l

+ hl -Eb:+,-1 Iut+,n-2, ut+m-3, . . . 1. (2.9)

3 West, Edison, and Cho (1993) suggested that an interesting alternative basis for comparing
forecasts is to calculate the utility of an investor with a particular utility function investing on the
basis of different variance forecasts.
 J.D. Hamilton, R. Susmrl~Journal CI~ Econometrics 64 (1994) 307-333 313

Table 2
Comparison of one-period-ahead forecasts of different models

Loss function (percent improvement)

Model MS.5 MAE lLEIZ ILEl

Constant variance 757.9 7.30 8.63 2.13

Gaussian GARCH (1, I) 866.8 7.69 8.27 2.04

( - 0.14) ( - 0.05) (0.04) (0.04)

Student t GARCH-L(1, I) 815.6 7.05 7.34 1.92

( - 0.08) (0.03) (0.15) (0. IO)

GED GARCH-L(I, I) 840.71 7.18 7.73 1.95

(~ 0.11) (0.02) (0.10) (0.08)

Gaussian ARCH-L(2) 1094.7 7.34 1.63 I .99

( ~ 0.44) ( - 0.01) (0.12) (0.06)

Gaussian SWARCH-L(2, 2) 822. I 6.66 7.65 I .95

( - 0.08) (0.09) (0.1 I) (0.09)

Student t ARCH-L(2) 904.1 7.02 7.93 2.00

( - 0.19) (0.04) (0.08) (0.06)

Student t SWARCH-L(2,2) 850.4 6.67 7.60 1.94

( ~ 0.12) (0.09) (0.12) (0.09)

Student f SWARCH-L(3.2) 814.3 6.53 7.14 1.92

( - 0.07) (0.1 I) (0.10) (0.10)
Student t SWARCH-L(4,2) 709.3 6.38 8.02 I .94
(0.06) (0.13) (0.07) (0.09)

The following loss functions are reported:

T

MSE= T-’ i ifi: -qJ)‘, MAE = T-’ 1 Ifi: - a:l,

,=I I= I

[LE]‘= T I i (ln(t?:) - In(rr:))‘. ILEl = T-’ i Iln(ti:) - In(

1= I ,=I
For the constant variance model, fi, represents y, - j and u,J represents s2. The percent improvement
compares each model to the constant variance specification.

Take conditional expectations of both sides of (2.9), noting that the symmetry of
the distribution of c,+~_ I implies that Ii,+,_ 1 is uncorrelated with u:+,,_ 1:

E(u:+,Iu,,u,;I ,... 1 = uo + a, -E(u:+,-,lu,, u,- ,,...I

+ <-E(d t+m-,I~lrur~,r...).E(~f+,~,Iul,ut~I,...)

+ h, *E(u:+,-,Iu,, u,- ,I... ). (2.10)

314 J.D. Hamilton, R. Susmel/Journal of’ Econometrics 64 (1994) 307-333

Table 3
Comparison of four-period-ahead and eight-period-ahead forecasts

Percent improvement in loss function

Model MSE MAE lW* ILEl

Four-week-ahead forecusts
Gaussian GARCH(1, I) ~ 0.22 ~ 0.24 - 0.11 - 0.06
Student t GARCH-L(l, 1) ~ 0.12 0.03 0.08 0.05
Student t SWARCH-L(2,2) - 0.00 0.05 0.02 0.02
Student t SWARCH-L(3,2) 0.00 0.07 0.02 0.04
Student f SWARCH-L(4,2 ) 0.01 0.07 - 0.01 0.03

Eight-week-&ad fbrecusrs
Gaussian GARCH( 1, I) ~ 0.24 - 0.41 - 0.19 - 0.15
Student t GARCH-L(1, 1) ~ 0.09 - 0.02 0.09 0.02
Student t SWARCH-L(2,2) 0.00 0.05 0.05 0.0 1
Student t SWARCH-L(3,2) 0.00 0.07 0.05 0.03
Student r SWARCH-L(4.2 ) 0.00 0.06 0.03 0.03

Table entries report the percent improvement of forecasts of each model compared to the con-
stant-variance forecast based on each of the four loss functions described in Table 2.

Let e+ rn,l denote the m-period-ahead forecast of the variance:

2
Ol + ,,I 1I s E(u:+,&,,u,_~ ,... ).

Thus for example gf+ , ,, = CJ:+1. Using this notation, (2.10) can be written

Thus the m-period-ahead forecast follows a simple first-order difference equa-

tion in the forecast horizon m, with decay parameter i given by

A = (a, + b, + </2). (2.11)

For the values in (2.5), i, is estimated to be 0.96. Thus the model implies
extremely persistent movements in stock price volatility. For example, any
change in the stock market this week will continue to have nonnegligible
consequences a full year later: (0.96)52 = 0.12.
Such persistence is difficult to reconcile with the poor forecasting performance
~ if the variance were this persistent, the model should do a much better job at
forecasting it. However, the model’s forecasts continue to worsen relative to
a constant-variance specification as the forecast horizon increases (see Table 3).
The consequence of the big value for i can be seen dramatically in the effects
of the stock market crash in October 1987. The top panel of Fig. 1 plots weekly
stock returns for the last six months of 1987, in which the October crash is an
 J.D. Hamilton, R. Susmel/Journal of‘ Econometrics 64 (1994) 307-333 315

I
-40 ’
87

40 , I

:;: h
________-__-----_______
-10 ! ____--
___--
-20 __--
\_,--
-30
.I
-40 ’
87

i; _jz*____
0 _______--------____________----__ ___
-10 - /------------
-20 - 1 ,’
-30 - 1 /’
(,
-40
87

Fig. I. Top panel: Weekly returns on the New York Stock Exchange during the second half of 1987.
Middle punt+ Widths of 95% confidence intervals for one-week-ahead forecasts of stock returns
implied by the Student t GARCH-L( I, 1) model. Bottom panel: + 2.(i,,, ~, where 63, , is cal-
culated from expression (3.15) for m = I using the parameters for the Student t SWARCH-L(3. 2)
model.

extreme outlier. The middle panel plots the corresponding 95% confidence
interval for future forecasts of U, based on the specification of the variance in
(2.5).4 The model anticipates possible weekly stock price movements in excess of
10% through the end of 1987 as a consequence of the October 1987 crash. This
perception was evidently not shared by stock market participants, whose beliefs

4 Recall that a standard f variable with 1’degrees of freedom has variance \~G(v~ 2). Thus if u, has
a Student r distribution with variance 0: and v degrees of freedom, then u; \:I’ - [u: .(!’ - 2)]
has the standard f distribution with 1’degrees of freedom. The 95% critical values for a t variable
with 5.6 degrees of freedom are f 2.5. Hence the 95% confidence intervals for u, would be calculated
as & 2.5 ~ ,~5.6- [ir;.(3.6)] = + 2*ir,.
about stock

specifications with changes in regime

A number of researchers have suggested that the poor forecasting perfor-

mance and spuriously high persistence of ARCH models might both be related
to structural change in the ARCH process. The high estimate for the persistence
parameter i. is known to be nonrobust across subsamples. Diebold (1986) and
Lamoureux and Lastrapes (1990) argued that the high estimated value for i. may
reflect structural changes that occurred during the sample in the variance
process. This is related to Perron’s (1989) observation that changes in regime
may give the spurious impression of unit roots in characterizations of the level of
a series. Cai (forthcoming) in particular noted that the volatility in Treasury bill
yields appears to be much less persistent when one models changes in the
parameter through a Markov-switching process, and it seems promising to
investigate whether a similar result might characterize stock returns.
For these reasons we explore a specification in which the parameters of an
ARCH process can occasionally change. Let J; be a vector of observed variables
and let s, denote an unobserved random variable that can take on the values
1,2, . . . , or K. Suppose that s, can be described by a Markov chain,

Prob(.s,=jl.s,_r = i,smz = k ,..., y1_,,ytm2 ,...)

= Prob(s, =jI.s,_, = i) = pij,

for i, j = I, 2, . . . , K. It is sometimes convenient to collect the transition prob-

abilities in a (K x K) matrix:

(3.1)

Note that each column of P sums to unity.

The variable s, is regarded as the ‘state’ or ‘regime’ that the process is in at date
t. By this we mean that s, governs that parameters of the conditional distribution
of y,. If the density of y, conditional on its own lagged values as well as on the
 J.D. Hamilton, R. Susmel/Journal of Econometrics 64 (1994) 307-333 317

current and previous q values for the state is of a known form,

f(Y,l S,,&l,..., s,~4,YI-l,Yt~2,..‘,Yo), (3.2)

then the methods developed in Hamilton (1989) can be used to evaluate the
likelihood function for the observed data and make inferences about the un-
observed regimes. For example, y, could follow an ARCH(q) process whose
parameters depend on the unobserved realization of s,, s,_ 1, . . . , stmq. Such
models have been fit to inflation data by Brunner (1991) and to Treasury bill
yields by Cai (forthcoming).
Although (3.2) provides a fairly general framework for describing structural
change, it has the limitation that the density of yI can only depend on a finite
number of lags of s represented by the parameter q. Thus, for example, one can
allow the parameters of an ARCH(q) process to change, but changes in
a GARCH(p, q) process with p > 0 are not allowed as a special case of (3.2).
The objective is to select a parsimonious representation for the different
possible regimes. A specification in which all of the parameters change with each
regime would likely be numerically unwieldy and overparameterized. Hamilton
(1989) suggested the following regime-switching model for the conditional mean:

Here pS, denotes the parameter p1 when the process is in the regime represented
by st = 1, while /lSt indicates pZ when s, = 2, and so on. The variable jjr was
assumed to follow a zero-mean qth-order autoregression:

j, = $Jrj,-r + 42j,-2 + ... + g&j,-, + 1,.

The idea behind this specification was that occasional, abrupt shifts in the
average level of yt would be captured by the values of pL,,.
A natural extension of this approach to the conditional variance would be to
model the residual u, in (2.1) as

ui = &&. (3.3)

Here 6, is assumed to follow a standard ARCH-L(q) process,

2, = h, - II, )

with L’,a zero mean, unit variance i.i.d. sequence, while 11,obeys

h: = a,+u,~:_, +u2u”:_, + “. + uqul& + t*d,_,*E:_,, (3.4)

where d,_ 1 = 1 if c,_ I < 0 and d,_ r = 0 for 6, 1 > 0. The underlying ARCH-
L(q) variable r?, is then multiplied by the constant &when the process is in the
regime represented by s, = 1, multiplied by & when s, = 2, and so on. The
factor for the first state y, is normalized at unity with q, > 1 forj = 2, 3, . . . , K.
The idea is thus to model changes in regime as changes in the scale of the
process. Conditional on knowing the current and past regimes, the variance
implied for the residual u, is

E(u:)s,.s,_ ,,..., s,_~,u,_~,u,_~. . . . . u,_,)

= Ys, (00 + ~I.(&-I&\,_,) + ~2-(4-2/$ ,) + ." + U&LIYs,J

+ 5.d,~,.(U:_IIHs,~,))

= c&2(s,,s,_,,.. .,yq). (3.5)

where&, = 1 fort.+, <Oandd,_, =Oforu,_, >O.

In the absence of a leverage effect (< = 0), we will say that u, in (3.3) follows
a K-state, qth-order Markov-switching ARCH process, denoted
4 - SWARCH(K, 4). In the presence of leverage effects (c # 0), we will call it
a SWARCH-L(K, q) specification. We investigated both Gaussian (c, - N(0, 1))
and Student t(ct distributed t with v degrees of freedom and unit variance)
versions of the model.
The appendix describes the algorithm used to evaluate the sample log-hkeli-
hood function,

z= 1 In.f(4’tIL’r~,,4’r-2,...,?‘~3),
,=,
(3.6)

which can be maximized numerically with respect to the population parameters

%4,QO?Ql, "2,....a,,p,,,p,,,... >pkkr~l~$,Z~~~~..Ykr~~ and v subject to the
constraints that y1 = 1, I,!=, plj = 1 for i = 1, 2. , K, and 0 < pi, < 1 for
i,j = 1, 2 3 . . . , K.5 The appendix also describes the inference about the particular
state the process was in at date t. When this inference is based on information
observed through date r it is called the ‘filter probability’:

Pbr~.~t~l~~~~, s,m,lyl,yrm L,..., y-3). (3.7)

Expression (3.7) denotes the conditional probability that the date t state was the
value s,, the date t - 1 state was the value s,_ I,..., and the date r - q state was
the value stmq. These probabilities condition on the values of Jj observed through
date t. Since there are Kqtl possible configurations for (.st, s, I, . . , s, 4), there

’ In practice these inequahtics were ensured by parameterking

P,,=H~,‘(I+f~~,+fI~,+.~~+f~,~6~1) for j=l.2 ._,,. K-l.

= l/(1 + o:, + rr,2* + “’ + fq&,) for j = K,

and estimating H,, for i = I. 2, ___, K and j = I, 2. ,,, . K - I wthout restrictions

 J.D. Humilton, R. Susmel~Journal of Ewnomerrrc.s 64 f IYY4) X17-333 319

are Kqtl separate numbers of the form of (3.7); these K qtl values sum to unity
by construction.
Alternatively, the full sample of observations can be used to construct the
‘smoothed probability’:

P(‘t14’TrVT~1,...,!:~3). (3.8)

Expression (3.8) denotes K separate numbers for each date t in the sample; again
these K numbers sum to unity.

3.1. Forecasts

To calculate m-period-ahead forecasts of I.$+,,,, consider first a hypothetical

situation in which we knew the values of s,, s_ ,, . . . , smq+ 1 with certainty,
meaning we would also know with certainty the values of Ci, = uJ& for
T = t, t - 1 1 ... 3 t - q + 1. For this information set the forecast of u:+,,, would be

E (u:+ mI s,, s, , . . . , s, - q+ 1, tit, 4 1, . . . , 6, q+ ,I

= E (~~s,_,~U1~+m~~,,~s,_
,...., s,_~+~,L~,, Ci_ ,,..., z?_~+,;

= EQ~s,+~l~,rst-
,,... ,.s,-,+,PEG:+,lfi,, Cc-,,...>&,+I), (3.9)

where the last equality follows from the fact that s, is independent of L’,and ii, for
all t and 5. Since s, follows a Markov chain, the first term in (3.9) is given by
K

E(s~,+,~s~,s~-~,.,.,.~~~~+,)
= C ~j*PrOb(st+, = jls,). (3.10)
j=l

The m-period-ahead transition probabilities can be calculated by multiplying

the matrix in (3.1) by itself m times:

1=
Prob(s,+, = 1 Is, = 1) Prob(s,+, = 1 Is, = 2) ... Prob(.s,+, = I Is, = K)
Prob(s, +m = 21.5, = 1) Prob(s,+, = 21s, = 2) ‘.. Prob(.s,+, = 2/s, = K)
p”

Prob(s,+, = Kls, = I) Prob(s,+, = Kls, = 2) ... Prob(s,+, = Kls, = K)

Thus, if the switching factors are collected in a (1 x K) vector b’,

E((~,,+~lsr = 4 =~‘P”ei, (3.11)

where ei denotes the ith column of the (K x K) identity matrix.

320 J.D. Hamilton, R. Susmel/Journnl qf Econometric.s 64 IIYY4) 307-333

The second term in (3.9) is equally simple to construct from the fact that
12,follows a standard ARCH-L(q) process:

E(~:+,I~I,~,-I,...,~,-q+l) (3.12)

= ao + a, *G;? + a,.&_, + ... + a;d:_,+, + <.d,.G: for m=l,

= a0 + (al + 5/2).&J+,- 1/f + u2*h;2t,-2,1 + ... + uq-h;z+,-q,t

for m=2,3,...,

where

IY;,, = a; for 7 ,< t,

= E(dflu”:,&_ r,...) for 5 > t.

The sequence (K:,t) for 7 = t + 2, t + 3, . . . is then calculated by iterating of

(3.12). Recalling that 6, = u,/&, the forecast in (3.12) is a function of
n,, U,P 1, ... 1urpq+ 1 and the particular set of values for s,, .s,_ 1, . . . , s,_~+ 1 that
were assumed:

= I;:+,~,t(.&,s,-I ?..... sr-4+lrUI. U,P I,..., f&+1).

The forecast in (3.9) can thus be written

E(u:+,Is,,~,~1,...,.~,-~+l,U1,,~,~1,...,~,~~+l)

= W%,)*h;z+ ,,,If(&..s-I >...1 s,~4+lrUt,U,~1 . . . . . u,-,+I). (3.13)

For given values of u,, LI,_1, . , umq+ 1, expression (3.13) describes a different
forecast of I,$+, for each possible configuration of s,, s,_ , , . . . , s,_~+ 1, which we
might denote

E (u:+ mI s,, s, - , , . , s, 4 + , > ti,, 6, - 1, , C, ,, + ,I

= h.(S,,.L1,..., S,m4+lrUf.U,-l . . . . . utq+,). (3.14)

In practice we do not know the value of s,, s,.. , , , s,_~+ 1. However, by the
law of iterated expectations,

$+,,,,, = E(u:+,~~~,,u,~I,...,u,~~+,)

=,,i_ .f ... i jk.h,s,-I,.... St-q+l,U,, U,-l,..., ut-,+I)

I ,=I .Smyil = 1

xp(.st,.s,-I...., s,-4+11J’t,JJ,-1 ,..., y-3)). (3.15)

That is, we simply weight each of the conditional forecasts in (3.13) by the filter
probability of that particular configuration to calculate an m-period-ahead
forecast of u:+, based on the actual data observed.
 Our model thus provides a framework that could generate the kind of
nonlinearity in stock return volatility documented by Friedman and Laibson
(1989) and Friedman (1992). These authors argued that conventional ARCH
models fail to forecast well because large and small shocks have different effects.
In the context of our model, suppose that the analyst is confident that the
market has been in state 1 for the past q periods and that p, , is close to unity. If
the date t residual is small, the analyst would continue to place a high probabil-
ity on the event that s, = 1, and the forecast would basically be

The marginal effect of u: on the forecast would then be given by (a, + <d,)/yi.
On the other hand, if the residual is sufficiently large that the analyst is
persuaded that the regime has shifted to s, = 2, the marginal effect of u: on the
forecast would be given by (a, + cd,)/y,. The specification thus allows for
nonlinearities of the type documented by Friedman and Laibson. In our model,
the nonlinearity arises as a result of the analyst’s inference about the current
volatility regime.

3.2. Persistence

Finally, we comment on the characterization of the persistence of the ARCH

component of a SWARCH process. From (3.12) the forecasts of fit+, obey the
following @h-order difference equation for m > 2:

h;2+nl,r = a0 + (a1 + yr/2)4?+,,,- Ill + dL,,~z,r + “’ + 4$?+,,,M,f.

It is well known that the solution to this difference equation takes the form

h-;+ ,n/r = (‘0 + (‘1jty + c11.1; + ... + C4iT’

where R,, jb2,... , i, are the eigenvalues of the matrix

(u, + 5/2) a, “’ u4-1 a4

(3.16)

We report the largest eigenvalue of this matrix as a measure of the persistence of

the ARCH component of a SWARCH-L(K, 4) process.
322 J.D. Hamilfon, R. Susmel~Journc~l of Economrrrics 64 I IN41 307-333

4. Empirical results

We fit a variety of different SWARCH specifications to the weekly stock

return data described in Section 2. We estimated models with q = 1 to 3 ARCH
terms and K = 2 to 4 states, with Normal and Student t innovations, and with
and without the leverage parameter 4. For each model, the negative
log-likelihood was minimized numerically using the optimization program
OPTIMUM from the GAUSS programming language, usually beginning with
steepest ascent and then switching to the BFGS algorithm. For models with
K = 2 we randomly generated over 200 different starting values; typically we
found a single local maximum for the likelihood function. For the K = 3
specification we used 25 different starting values. The K = 4 specification
proved extremely difficult to maximize, owing in part to a nearly singular
Hessian, as described below.
In every specification we looked at, the second ARCH parameter a, and
the leverage parameter 5 were strongly statistically significant on the basis of
both Wald tests and likelihood ratio tests. Further, both tests overwhelmingly
rejected the Normal in favor of the Student t formulation. We also investigated
a GED specification, which only slightly outperformed the Normal. The third
ARCH parameter a3 was not statistically significant or only marginally signifi-
cant in the specifications with K = 2, so we did not attempt to estimate this
parameter for the SWARCH models with K > 2. For these reasons, Tables
1 through 3 primarily report the results from SWARCH-L(K, 2) specifications
driven by Student t innovations.
Since the GARCH(1, 1) specifications are not strictly nested within the
SWARCH specifications, rows five and seven in Tables 1 and 2 also report
results for ARCH-L(2) models with Normal and Student t innovations. Al-
though an ARCH-L(2) process could be described as a special case of
SWARCH-L(K, 2) with K = 1. the usual regularity conditions justifying the 11’
approximation to the likelihood ratio test do not hold in this setting, since
the parameter q2 is unidentified under the null hypothesis that there is really
only one state.’ Table 1 nevertheless reports critical values for the likelihood
ratio tests as if the x2 approximation were valid. At a minimum we regard these
as a useful descriptive summary of the fit of alternative models. The p-values for
tests of the null hypothesis of only one or two states are so tiny that we have little
doubt that these hypotheses would be rejected by any more rigorous testing
procedures; indeed these events are so remote that the probabilities are not
reliably calculated by the numerical routines of the GAUSS programming
language on which the table entries are based. On the other hand, the suggestion

‘Hansen (1991, 1992) has proposed asymptotically valid tests, though their application here would
be quite difficult numerically.
 J.D. Hamilton, R. SusmeliJournal of Econometrics 64 11994) 307-333 323

in Table 1 that the three-state specification is rejected in favor of the four-

state specification at the 0.01 level may be sensitive to the distributional
assumptions.
In addition to conventional tests for statistical significance, Tables 2 and
3 compare models on the basis of forecasting performance. The SWARCH-
L(4,2) specification is the only model we investigated that has a better mean
squared error in forecasting u: than the constant variance specification. The
SWARCH-L(4,2) model is also clearly the best in terms of minimizing the mean
absolute error as well, and continues to give useful eight-week-ahead forecasts.
This model also performs reasonably when judged by the difference between
ln(u:) and ln(cr:), though it is not quite as good as the GARCH-L(l, 1)
specifications.
Table 1 also reports the model selection statistics proposed by Akaike
(1976) and Schwarz (1978), though the asymptotic justification for these
statistics again assumes the same regularity conditions referred to earlier,
which are not fulfilled for this application. Based on Akaike’s criterion, the
SWARCH-L(4, 2) is the best model among any we investigated, followed
by the SWARCH-L(3,2). Based on Schwarz’s criterion, the GARCH-L(1, 1)
is best, followed by the SWARCH-L(3, 2).
In estimating the SWARCH-L(3, 2) and SWARCH-L(4,2) specifications, we
initially imposed no constraints on any of the transition probabilities pij other
than the conditions that 0 < pij < 1 and c,!= 1 pij = 1. Several of these unrestric-
ted MLE’s fell on the boundary pij = 0, which is another violation of the
regularity conditions. To calculate standard errors we then imposed pij = 0 and
treated this parameter as a known constant for purposes of calculating the
second derivatives of the log-likelihood.
The estimated Student t SWARCH-L(3,2) specification is as follows, with
standard errors in parentheses:

y, = 0.35 + 0.25 y,_ 1 + u,,

(0.05) (0.03)

u, b i.i.d. Student t with unit variance and 7.2 d.f.,

(1.4)

h; = 0.57 + 0.03 a;_i + 0.12d;_2 + 0.42 d,_l 12;_~,

(0.11) (0.04) (0.05) (0.10)
 = 0 if r(,-, > 0,

81 = 1, .ilz = 4.4, 63 = 13.1,

( 1.O) (3.2)

0.9924 0 0.0026
(0.0069) (0.0029)
p = 0.0076 0.9914 0.0144 .
(0.0069) (0.0066) (0.0 144)
0 0.0086 0.983 1
(0.0066) (0.0119) _
The rowi, column i element of P represents the probability of going from state
i to state j.
Note that the estimated autoregressive coefficient $J is clearly nonzero the
new distributional assumptions and description of heteroskedasticity do not
alter the conclusion that weekly stock price returns exhibit positive serial
correlation.
The variance in the medium-volatility state (s, = 2) is four times as great as
that in the low-volatility state, while that in the high-volatility state (.st = 3) is
thirteen times as large as in the low-volatility state.
The top panel of Fig. 2 plots the weekly stock return series J’,, while the other
three panels plot the smoothed probabilities Prob(.s, = ilyT, J17. ,, . , J-~~).
The low-volatility state describes the long quiet period from January 1963
through the end of 1965. Most of the other observations come from the
medium-volatility state, with high-volatility episodes characterizing the last half
of 1962. May 1969 to January 1972, February 1973 to April 1976, the period
following the October 1987 crash, and probably for a short period around
November 1982 as well.
The U.S. economy experienced four economic recessions during this sample,
whose beginning and ending dates are marked with vertical lines in the bottom
panel of Fig. 2. The market is judged to have been in the high-volatility state
throughout the 1969970 and 1973-75 recessions and likely towards the end of
the 198 1~ 82 recession as well. Thus the episodes of high stock market volatility
appear to be related to general business downturns.’ The brief recession in 1980
does not appear to have coincided with unusually high volatility, however.
The estimated transition probabilities describe each state as highly persistent.
State 1 would be expected to last on average for (1 ~ fir 1)- ’ = 132 weeks, while

_ French and Sichel (1993) presented interesting related ewdencc that the volatility of economc
actiwty is higher during rcccssions than during expansions.
 J.D. Humilfon, R. Susmel;Journul c~f Econometrics 64 /I9941 307.-333 325

I
IO

-10

-20
-30
I /
62 65 68 71 74 77 80 83 86

62 65 68 71 74 77 80 83 06

I
62 65 68 71 74 77 80 83 86

1.00 -
0.75 -
0.50 -
0.25
0.00 -

62 65 68 71 74 77 El0 83 86

Fig. 2. Top punel: Weekly returns on the New York Stock Exchange from the week ended July 31,
1962 to the week ended December 29, 1987. Second panel: Smoothed probability that market was in
regime I for each indicated week [Prob(s, = 1 jar, )lr_ I, ._., ym3)],as calculated from the Student
t SWARCH-L(3, 2) specification. Third punel: Smoothed probability for regime 2. Fourfh
panel: Smoothed probability for regime 3. Vertical lines mark start and end of economic recessions
as determined by the National Bureau for Economic Research.

states 2 and 3 typically last for 116 weeks and 59 weeks, respectively. The market
was in the quiet state 1 for only a single episode in the sample, which episode was
preceded by state 3 and followed by state 2. Hence the maximum likelihood
estimate is that state 1 is never preceded by state 2 (flz, = 0) and state 1 is never
followed by state 3 (@r3 = 0).
Although the states are highly persistent, the underlying fundamental ARCH-
L(2) process for I?, is much less so, with decay parameter i estimated to be 0.48.
Note that i4 = 0.05, meaning that the volatility effects captured by G, die out
almost completely after a month. The bottom panel of Fig. 1 plots k 2ri,,( ~ , for
the second half of 1987. As with the GARCH specification appearing in the
middle panel, the crash initially widens these bands by an order of magnitude. In
326 J.D. Humilron. R. .Su.vnrl~Joumc~l CI/ E~~onomwtric~.s
64 (IYY4) 307 333

IO -
O-
-10 -
-20 -
-30 -
-40
62 65 68 71 74 77 80 83 86

.~~ -,,-3 ,,.,., .,,/-,.,.#---I ‘S,..,.,

65 71 77 83

Fig. paw/: Weekly returns on the New York Stock Exchange from the week ended July 31,
1962 to the week ended December 29, 1987. Middle panel: + 2 ‘6, where d: is calculated from (2.5).
the variance process for the Student I GARCH-L(I, I) model. Bottompanrl: k 2.6,,, ~, where
r?,:, , is calculated from expression (3.15) for m = I using the parameters for the Student t
SWARCH-L(3,2) model.

contrast to the subsequent slow decay implied by the GARCH specification,

however, this dramatic effect dies out relatively quickly, though a more modest
widening persists as long as the market remains in the high-volatility state 3.
Fig. 3 compares the f 2 * CJ,bands for the GARCH-L(1, 1) specification
(middle panel) with those for the SWARCH-L(3,2) model (bottom panel) for
the entire sample of observations. Both models infer prolonged episodes of high
variance in the early and middle 1970’s. For the SWARCH model this reflects
the long periods in which the market appeared to be in the high-volatility state
3, whereas for the GARCH model this results from the long moving average of
many large squared residuals. Although the broad patterns are similar, the
models differ significantly in the reaction to large outliers. The GARCH confi-
dence intervals shrink very gradually in response to these events, while the
SWARCH intervals quickly return to the trend level associated with a particular
regime. The improvement in forecasting of the SWARCH model over the
GARCH appears to be due to the ability to track the year-long shifts in
volatility without imputing a large degree of persistence to the effects of indi-
vidual outliers.
To measure the sensitivity of the forecasts of our model to uncertainty about
the population parameters, we ^^
conducted
^ ^the following
^ Monte Carlo experi-
ment. Let 0 = (2, $,S,, dl, dZ, l, U,,, 012, H3,, OjL, Q2, Q3, ?)’ denote the max-
imum likelihood estimate of the population parameters for the SWARCH-
L(3, 2) model described above with transition probabilities parametrized as in
Footnote 5. Let fi denote the asymptotic varianceecovariance matrix of e as
estimated from second derivatives of the log-likelihood. We generated 500
values for the vector 8 drawn from a N(6, d) distribution. For each fIi and each
date t in the sample we calculated what the historical one-period-ahead forecast
oftG+ 1 would have been for that value of &, based on the actual historical values
for y,, y, 1, . . . , y _ 3, with this forecast denoted CT:+, (ei). The sample variance of
a:+ 1(Si) across Monte Carlo draws,
500 i 500

(l/500) 1
i= 1
0~+l(si)-(1/500) C
j= I
af+l(ej)
I2> (4.1)

then gives an indication of how sensitive the forecast variance for date t + 1 is to
uncertainty about the true value of the parameter vector 0. The mean value for
the square root of (4.1) across the t = 1,2, . . . , 1327 observations was 1.1. This
standard error compares with a mean forecast given by

(l/500) ‘f g:+r(Bj)
j= 1

Hence the standard error arising from parameter uncertainty typically is modest
relative to the size of the forecast itself.
Our maximum likelihood estimates for a Student t SWARCH-L(4, 2) speci-
fication were as follows:

yt = 0.35 + 0.25 y,_ 1 + u,,

(0.05) (0.03)

u, = Js,%

6, = h, * c, ,

c, - i.i.d. Student t with unit variance and 8.7 d.f.,

(2.0)
h: = 0.55 + 0.02 a:_1 + 0.13 lI:-z + 0.41 d,_, ii:_r,
(0.11) (0.04) (0.05) (0.09)
328 J.D. Hamilton, R. SusmdiJournul of Econometrics 64 i 1994) 307- 333

d,_, = 1 if ur_r < 0:

Yl =o if ut_, >O,

91 = 1, i2 = 4.5, Q3 = 13.8, g4 = 169,

(0.9) (3.3) (168)

0.9931 0 0 0.2758
p =

L0.0069
0
0
0.9925
0.0034
0.0042
0.0153
0.9847
0
0
0.7242

I
0
’

It is interesting that the first three of these states are essentially the same as
states 1 through 3 of the SWARCH-L(3,2) specification, while the fourth state
corresponds to an order of magnitude increase in the variance even beyond that
predicted in the ‘high-volatility’ state 3. Fig. 4 plots the smoothed probabilities
implied by this model. Only two observations in the sample are clearly gener-
ated by state 4. One is the October 1987 crash. The other is more surprising, and
corresponds to the 5.8% surge in stock prices in the first week of January 1963.
This move followed two very quiet weeks for which the squared residuals were
essentially zero, forcing hf nearly to its lower limit of a0 = 2.5. The probability of
this occurring is the probability that a t(8.7) variable exceeds

5.37/J(2.5) e(8.7 - 2) t (8.7) = 3.87,

which probability is 0.002. Even if the process were imputed to have switched
from state 2 to state 3 for this date, the probability of generating so large a gain
would still only be 0.03. Since some sort of shift must have occurred at this date,
one is inclined to regard this observation as having come from the rare condi-
tions associated with state 4.
It is also interesting to note that p34 is estimated to be zero ~ the extreme state
4 appears never to have developed from a high-volatility state 3 but rather
always follows the moderate-volatility state 2. This is consistent with Bates’s
(1991) failure to find any indication in option prices that the market perceived an
increase in risk in the two months prior to the October 1987 crash.
Although these results are intriguing, one should not read too much into
them. It might appear that three parameters - pZ4, pa3, and .q, ~ are identified
solely on the basis of two observations. It is not quite this bad, since there is
a modest probability that some of the other observations also may have come
from regime 4. It turns out that c,‘=r Prob(s, = 4)41T,1.7._ r, . . . ,y_J) = 3.6.
These other observations also give some information about these parameters,
and of course all of the observations from regime 2 contain information about
P ~ one can say with considerable confidence that this probability must be
qt?te low. Even so, the Hessian for this SWARCH-L(4,2) specification is very
 62 65 68 71 74 77 80 83 86
_-- .___
I 00
0 75
050
025
000

62 65 68 71 74 77 ei- 83 6'6

I _____ .~~_. _~ J

62 65 68 71 74 77 80 83 86

Fig. 4. Top pmd: Weekly returns on the New York Stock Exchange from the week ended July 31.
1962 to the week ended December 29, 19X7. Src~ontl panel: Smoothed probability that market was in
regime I for each indicated week [Prob(s, = I I!‘.,. y7._!. _,., y_ 3)]. as calculated from the Student
I SWARCH-L(4, 2) specification. Third puwl: Smoothed probability for regime 2. Fourth
ptrnel: Smoothed probability for regime 3. Fifih pmd: Smoothed probability for regime 4.

nearly singular, and ‘asymptotic’ standard errors for these parameters are not
meaningful.
Although some of the individual parameters of the SWARCH-L(4,2) model
are not measured with much confidence, we nevertheless find the results of
interest. It is worth emphasizing that nothing about the model specification
forced the procedure to regard the October 1987 crash as an observation from
a single extreme regime. Indeed, the likelihood function suggests special treat-
ment of October 1987 only when a fourth state is allowed, the first three states
being reserved for broader patterns common to hundreds of observations.
Specifying a general probability law that allows a rich class of different possibili-
ties and letting the data speak for themselves in this way seems preferable
to imposing dummy variables or break points in an arbitrary fashion and
330 J.D. Hamilton, R. Susmel/Journal qf Econometrics 64 (1994) 307-333

regarding the break dates as the outcome of a deterministic rather than

a stochastic process. The improvement in the value of the likelihood function
achieved by broadening the class of dynamic models in this dimension seems
a valid and useful framework for deciding whether the quiet stock market of the
early 1960’s or the turbulence of October 1987 ought to be regarded as special
episodes.

5. Conclusion

This paper introduced a class of Markov-switching ARCH models which

were used to describe volatility of stock prices. Our SWARCH specification
offers a better statistical fit to the data and better forecasts. Our estimates
attribute most of the persistence in stock price volatility to the persistence of
low-, moderate-, and high-volatility regimes, which typically last for several
years. The high-volatility regime is to some degree associated with economic
recessions. Our analysis also confirms the findings of earlier researchers that
stock price decreases lead to a bigger increase in volatility than would a stock
price increase of the same magnitude, that the fundamental innovations are
much better described as coming from a Student t distribution with low degrees
of freedom than by a Normal distribution, and that weekly stock returns are
positively serially correlated.

Appendix: Procedure for evaluating the likelihood function

Step t of the iteration to calculate the likelihood function has as input

Ph, h-1, ... >.bq IYt?Yt-I>..‘, Y-3). (A.11

Each of the Kq + ’ numbers represented by (A. 1) is multiplied ps,, St+ , and by

~(J~,+~(s~+~,s~ ,..., ~,_~+r,y,,y~_r ,..., yf~q+l)toyieldtheKq+2separatenum-
bers

P(St+l,St,Sr-1,...,Sr-q,Yt+llYl,Yr-1,...,4’~3). (A.3

For the Gaussian specification the preceding calculation uses

- (Yt+ 1 - 2 - 4Y,Y
i 2a:+l(S,+l,S,,...,S,-q+l)I’
=&+l(s,,l~s,....,s,,,l)~exp
 J.D. Hamilton, R. Susmel/Journal of Economrtrics 64 (1994) 307-333 331

where a:(~,, s,_ r,...,.s_,)isgiven by(3.5)withu,=y,-x--_y,_r.Forthe

Student t version of the model we instead use

,f(Y*+ 1 Is I+l,~f,...,~f-q+l,YrrYl-l,...,Yr~q+l)

TC(v + 1)/21
= T(v/2).~.~.a,+.I(S,+1,St,...,S,~q+l)
m(v+ 1)!2
(Y t+1 - u - 4YfJ2
x
i l +(v-2).~,2_l(st+l,st,....s,-u+I) I
The numbers in (A.2) sum to the conditional density of yI+ 1,

f(Y,+llYr~Yf--lr...,Y~3)

= f $ ---~~~~~p(s,,,;r,.r,-,,....r,~,.~,+~iv,.i,-,.....li). (A.3)
s,+,=1s,=1 * 4

from which the sample log-likelihood (3.6) can be calculated. If for any given
s,+ 1, St>‘.. >Sr-qf 1 the numbers in (A.2) are summed over the K possible values
for s~_~ and the result is then divided by (A.3) one obtains

P(&C,> s t,...,.Lq+l yt+l>.Yt,..‘, Y-317

which is the input for step t + 1 of the iteration.

Theiterationwasstartedwithp(so,s_l,...,s~,Iy,,y_,,...,y~,)setequalto
the ergodic probabilities implied by the Markov chain as described in
Eq. (22.2.26) in Hamilton (1994). Kim’s (1994) algorithm for calculating the
smoothed probabilities p(s, (y,, y,_ 1, . . . , ym3) is described in Eq. (22.4.14) in
Hamilton (1994).

References

Akaike, Hirotugu, 1976, Canonical correlation analysis of time series and the use of an information
criterion, in: Raman K. Mehra and Dimitri G. Lainiotis, eds., System identification: Advances
and case studies (Academic Press, New York, NY).
Bates, David S., 1991, The crash of’87: Was it expected? The evidence from options markets, Journal
of Finance 46, 1009-1044.
Baillie, Richard T. and Ramon P. DeGennaro, 1990, Stock returns and volatility, Journal of
Financial and Quantitative Analysis 25, 203-214.
Bera, Anil K., Edward Bubnys, and Hun Park, 1988, Conditional heteroskedasticity in the market
model and efficient estimates of betas, Financial Review 23, 201-214.
Black, Fischer, 1976, Studies of stock market volatility changes, 1976 Proceedings of the American
Statistical Association, Business and Economic Statistics Section, 1777181.
Bollerslev, Tim, 1986, Generalized autoregressive conditional heteroskedasticity, Journal of Econo-
metrics 3 1, 3077327.
Bollerslev, Tim, 1987, A conditionally heteroskedastic time series model for speculative prices and
rates of return, Review of Economics and Stattstics 69, 542-547.
332 J.D. Hnmilton. R. Susmel~Journnl of Econonwtrics 64 (1994) 307~-333

Bollerslev, Tim, Ray Y. Chou, and Kenneth F. Kroner, 1992, ARCH modeling in finance: A review
of the theory and empirical evtdence, Journal of Econometrtcs 52. 5 59.
Brunner, Allan D., 1991. Testing for structural breaks m U.S. post-war inflation data, Mimeo.
(Board of Governors of the Federal Reserve System. Washington. DC).
Cai, Jun. forthcoming, A Markov model of unconditional vartance in ARCH. Journal of Business
and Economic Stattstics.
Connolly, Robert, A., 1989. An examination of the robustness of the weekend effect. Journal of
Financial and Quantitative Analysis 24, l33- 169.
Diebold. Francts X., 1986. Modeling the perststence of conditional v’artances: A comment. Econo-
metric Reviews 5, 51 56.
Diebnold, Francis X., Steve C. Lim. and C. Jevons Lee, 1993, A note on conditional heteroskedastic-
ity in the market model. Journal of Accounting, Auditing. and Finance 8, I41 150.
Engle. Robert F., 1982, Autoregressive condittonal heteroscedasticity with estimates of the variance
of United Kingdom inflation. Econometrica 50. 987-1007.
Engle, Robert F. 1991, Statistical models for financial vjolatility. Mtmeo. (liniversity of Californta,
San Diego. CA).
Engle, Robert F. and Chowdhury Mustafa 1992 Implied ARCH models from options prices,
Journal of Econometrics 52. 289 -31 I.
Engle. Robert F. and Victor K. Ng, 1991. Measuring and testing the impact of news on volatility,
Mimeo. (University of California, San Diego, CA).
French, Mark W. and Daniel F. Sichel, 1993, Cyclical patterns in the variance of economic activity,
Journal of Business and Economic Statistics I I, I I3 119.
Friedman, Benjamin M.. 1992. Big shocks and little shocks: Security returns with nonlinear
persistence of volatility, Mimeo. (Harvard University, Cambridge. MA).
Frtedman. Benjamin M. and David 1. Laibson. 1989, Economic implicattons of extraordinary
movements in stock prices, Brookinga Papers on Economic Activity 2. l37- 189.
Glosten, Lawrence R.. Ravi Jagannathan. and David Runkle, 1989, Relationship between the
expected value and the volatility of the nominal excess return on stocks, Mimeo. (Northwestern
University, Evanston. IL).
Gourieroux. Christian and Alain Monfort, 1992, Qualitative threshold ARCH models, Journal of
Econometrics 52, 159-- 199.
Hamilton, James D.. 1989, A new approach to the economic analysis of nonstationary time series
and the business cycle, Econometrica 57. 357-384.
Hamilton. James D., 1994. Time series analysis (Princeton University Press, Princeton. NJ).
Hansen, Bruce E.. 1991, Inference when a nuisance parameter is not identified under the null
hypotheses, Mtmeo. (University of Rochester, Rochester, NY).
Hansen, Bruce E.. 1992. The likelihood ratio test under non-standard condittons: Testtng the
Markov trend model of GNP, Journal of Applied Econometrics 7, S6l-~S82.
Ktm, Chang-Jin. 1994. Dynamic linear models with Markov-switching, Journal of Econometrtcs 60.
I 22.
Lamoureux. Christopher G. and William D. Lastrapes, 1990. Persistence in variance, structural
change and the GARCH model, Journal of Business and Economic Statisttcs 8. 2255234.
Lamoureux, Christopher G. and Wtlltam D. Lastrapes, 1993. Forecasting stock return variance:
Toward an understanding of stochastic implied volatilities. Review of Financial Studtes 5.
2933326.
Morgan, Alison and Ieuan Morgan, 1987. Measurement of abnormal returns from small firms.
Journal of Business and Economic Statistics 5, 121-129.
Nelson. Daniel, 1991. Conditional heteroskedastictty in asset returns: A new approach. Econo-
metrica 59, 3477370.
Pagan. Adrian and G. William Schwert, 1990, Alternative models for conditional stock volatility,
Journal of Econometrics 4.5. 2677290.
 J.D. Humilton, R. Su.wwl/Journul c~f Econometrics 64 (1994) 307-333 333

Pagan, Adrian and Aman Ullah, 1988, The econometric analysis of models with risk terms, Journal
of Applied Econometrics 3, 87-105.
Perron, Pierre, 1989, The great crash, the oil price shock, and the unit root hypothesis, Econometrica
57, 1361-1401.
Schwarz, Gideon, 1978, Estimating the dimension of a model, Annals of Statistics 6, 461-464.
Schwert, G William and Paul J. Seguin, 1990, Heteroskedasticity in stock returns, Journal of
Finance 45, I 12991155.
West, Kenneth D., Hali J. Edison, and Dongchul Cho, 1993, A utility based comparison of some
models of foreign exchange volatility, Journal of International Economics 35, 2346.