The Review of Economic Studies Ltd.

Quadratic ARCH Models

Author(s): Enrique Sentana
Source: The Review of Economic Studies, Vol. 62, No. 4 (Oct., 1995), pp. 639-661
Published by: The Review of Economic Studies Ltd.
Stable URL: http://www.jstor.org/stable/2298081
Accessed: 01/03/2010 21:40

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=resl.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

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 [email protected].

The Review of Economic Studies Ltd. is collaborating with JSTOR to digitize, preserve and extend access to
The Review of Economic Studies.

http://www.jstor.org
Reviewof EconomicStudies(1995) 62, 639-661 0034-6527/95/00290639$02.00
? 1995The Reviewof EconomicStudiesLimited

Quadratic ARCH Models

ENRIQUE SENTANA
CEMFI

First version received December 1991; final version accepted July 1995 (Eds.)

We introducea new modelfor time-varyingconditionalvariancesas the most generalquad-

ratic versionpossiblewithin the ARCH class. Hence, it encompassesall the existingrestricted
quadraticvariancefunctions.Its propertiesare very similarto those of GARCH models, but
avoids some of theircriticisms.In univariateapplicationsto daily U.S. and monthlyU.K. stock
marketreturns,QARCH adequatelyrepresentsvolatilityand risk premia.QARCH is easy to
incorporatein multivariatemodelsto capturedynamicasymmetriesthat GARCHrulesout. Such
asymmetriesare found in an empiricalapplicationof a conditionalfactor model to 26 U.K.
sectorialstock returns.

1. INTRODUCTION
The analysisof economic and financialtime-seriesdata usuallyinvolvesthe study of the
firstand possiblysecondconditionalmomentsof the series(givenpast behaviour)in order
to characterizethe dependenceof futureobservationson past values. Besides,these two
conditionalmomentsare often identifiedwith importanteconomicconcepts.For example,
considerthe univariatestochasticprocessfor stock marketexcessreturns,x,, t = 1, 2, . . ..
whose first two conditionalmomentsgiven the informationset (generatedby) X, 1, =
{Xt-I, Xt-2, .}are:

p t= E(xt IXt- OO)=(Xt - OO,(la)

a2= V(X,tl X,t_, oo.) 2(Xt_ 1,,o ( lb)

In this context, p, is usually associatedwith the risk premiumfor the stock market
as a whole, a 2 with its volatility, and p, /U2 with the market price of risk. From an
empiricalpoint of view, the first step often consists in the estimationof the conditional
meanand variance.In practice,though,this is not a simpletask as both p ( ) and a2( ) are
generallyunknownfunctionsof the informationset X,_, . The most commonapproach
employed is to assume a particularfunctional form for p( ) and &2( ) characterized
by certainunknownparameters(which have to be estimated),althoughnon-parametric
techniques(which obtain estimatesof p, and U2 directly)and mixed approaches(semi-
parametric)have gainedincreasedattentionrecently.
A fact that should be borne in mind when modellingconditionalvariancesis that
they are non-negative(or, in the multivariatecase, positive semi-definite).Just as it is
reasonableto use non-linearmodels for p( ) when x, is alwayspositive (e.g. Wooldridge
(1992)), it makes sense to use functionalforms for &2( ) that ensurepositivity.Here, an
additiveparametricfunctionalform for &2( ) is proposedas a naturalextensionof many
ARCH models alreadystudied. Since it is clear that a first-orderpolynomialin X,_1,
cannot be alwayspositive (unlessit is triviallyconstantor x, has boundedsupport),and
this is also true of any odd-orderone, an even-orderpolynomialis required.The purpose
of this paper is to discuss how to model the conditionalvariancea2( ) as a quadratic
functionof X,_ . To stressthis point we shallcall this formulationthe QuadraticARCH
639
640 REVIEWOF ECONOMICSTUDIES

(or QARCH) model. As we shall see, it combines many attractivefeatures of several

existing ARCH models while ensuring that estimated conditional variances are non-
negative.
In particular,the QARCH model nests the recentlyproposed AugmentedARCH
(AARCH) model of Bera and Lee (1990), which in turn encompassesEngle's (1982)
originalARCH model. It also nests the linear "standarddeviation"model discussedby
Robinson (1991), and the asymmetricARCH model in Engle (1990). This nestinghas at
least threenon-trivialadvantages.First, many theoreticalresultsderivedfor thesemodels
still hold with minor modificationsfor the QARCH model, including estimation and
testing,stationarityconditionsand persistenceproperties,autocorrelationstructure,tem-
poral and contemporaneousaggregation,forecasting,etc. Second, QARCH conditional
variancescan also be easilyintegratedin economicmodels,just as lineartime-seriesmodels
for the conditionalmean are (see Campbelland Hentschel(1992) for an applicationto a
stock returnsmodel). Third, the QARCH model is capable of improvingthe empirical
successof ARCH models, since it avoids some of its criticismswithout departingsignifi-
cantly from the standardspecification.In particular,it provides a very simple way of
calibratingand testing for dynamicasymmetriesin the conditionalvariancefunction of
the kind postulatedfor some financialseries (see e.g. Black (1976) and Nelson (1991)).
The QARCH formulationcan also be interpretedas a second-orderTaylorapproxi-
mation to the unknown function &2( ), or, alternatively,as the quadraticprojectionof
the squareinnovationon the informationset. Actually,this projectioncoincideswith the
projectionof the true conditionalvariance a2 on a "quadraticinformationset", so that
we can also understandthe QARCH model as producingsmooth filtered"estimates"of
or (see Nelson (1992) and Nelson and Foster (1994) for a relatedinterpretationof ARCH
models as filters).In addition,we can also relateQARCH to randomcoefficientsmodels
(cf. Tsay (1987) and Bera and Lee (1990)).
By and large, most of the existing literatureon ARCH models is concernedwith
univariatemodels(see e.g. Bollerslev,Engleand Nelson (1994) for a recentsurvey).Given
that many issues in finance, and in particular,asset pricing theories, are relatedto the
variancesand covariancesof many assets, it is of the utmost practicalimportanceto be
able to extendunivariatemodelsso as to capturetime-variationin the (conditional)mean
vector and covariancematrix. An additionaladvantageof the QARCH formulationis
that it is very easy to generalizeto multivariatemodels, either directly, or more con-
veniently,throughthe differentcovariancestructuressuggestedin the literature.In particu-
lar, QARCH formulationsare the only ones that can be easily adaptedto the context of
conditionallyheteroskedasticlatentfactormodels,and do not entaila largecomputational
burden.Hence, it can also be used to capturepotentialdynamicasymmetriesat a multiple
assets level.
The paperis organizedas follows. Section2 introducesthe QARCH(q)model, states
underwhatparameterrestrictionsall the otherproposedquadraticmodelscan be obtained
from it, discussesits positivity,and interpretsit both as a least squaresfilter of the true
conditionalvarianceu2 basedon a "quadratic"informationset, and as a quadraticTaylor
approximationto the unknown conditional variancefunction. It also discusses testing
proceduresand its generalizationto GQARCH(p,q) along the lines of Bollerslev(1986).
Section 3 obtainsthe stationarityconditionas well as an expressionfor the unconditional
variance.Fourth moments,the autocorrelationfunctionfor the squaresof the seriesand
the covariancesbetweenthese and lagged levels are also discussed.Univariateempirical
applicationsto daily U.S. and monthly U.K. stock returnsare carriedout in Section 4.
Potentialgeneralizationsto the multivariatecase areentertainedin Section5. In particular,
 SENTANA QUADRATIC ARCH MODELS 641

a latent factormodel with GQARCH factorsis discussedin some detail. In Section6, an

illustrativeempiricalapplicationof the model to monthly data on U.K. sectorialstock
returnsis carriedout. Specifically,a multivariatefactormodelis estimatedfor stockreturns
of 26 industrialsectors.Finally Section 7 concludes.

2. THE QARCH MODEL

2. 1. Definition
Assume initially that p(X,-,,,) )=O and that &2(X,_,,) is a function of the past q
values of x, only, i.e. &2(X,l) = &2(X,_I,q). The following conditional variance
parameterization:
orX,- I1,q) = + Eq V
I YtiXt-i+ Zq= I aijx2t-j+2
Y_j=i+I

= 0+ 'Xt,-,q+X;_i,qAXt -,q (2)

where ,i is a q x 1 vector and A a symmetricq x q matrix,is the most generalquadratic
version possible of the parametricARCH variancefunction h(Xt- ,q;rb) consideredin
Engle (1982), and for that reasonwe shall call it the QuadraticARCH (QARCH herein-
after) model. It thereforeencompassesall the examplesof quadraticvariancefunctions
proposed in the literature:the AugmentedARCH model of Bera and Lee (1990), the
standard'ARCH modelof Engle(1982), the linear"standarddeviation"modelconsidered
by Robinson (1991), and the asymmetricARCH model in Engle (1990). The AARCH
modelassumesthat fi = 0, whereasEngle'sARCH assumesthat, in addition,A is diagonal.
The linear standard deviation model assumes that ( p + g'X,t 1,q)2 which implies 0 =
2 =

p2, i =
2pq and A = oo', a rank 1 matrix.2 An obvious fourth restrictedparameterization,
of which the asymmetricARCH model is a special case, is also encompassed:a model
with A (band-)diagonalbut ,ip 0, which we shall term the (band-)diagonalQARCH
process.
The cross-productterms give an indicationof the extra effect of the interactionof
laggedvalues of x, on the conditionalvariance.Hence,by allowingthem to be non-zero,
we can accountfor the possibilitythat the occurrenceof, for example,two successivelarge
values of x, of the same sign affects the conditionalvarianceby more than an ARCH
model would allow. But the substantiveadvantageof the QARCH formulationvs. the
AARCH and ARCH models which it nests, is that by allowing yVto take any value, i.e.
by not centringthe quadraticpolynomialfor &2( ) at 0, a dynamicasymmetriceffectof
positive and negativelagged values of x, on u2 is allowed. As an example,let's take the
QARCH(I) model, i.e. o2 = 0+ f X, - + a IIxt- I . If vI is negative, the conditional variance
will be higher when x,tI is negative than when it is positive.3In the context of stock
market volatility, this could capture the leverageeffect noted by Black (1976). Hence,
the QARCH model provides an additive heteroskedasticalternativeto the asymmetric
multiplicativeheteroskedasticEGARCH model of Nelson (1991). On the other hand, in
1. Engle's(1982) originalmodel is sometimesreferredto as linearARCH. However,the term linearis
unfortunatesince as (2) shows it is only linear in the squaresof the past values, not in the informationset
X, - I. .
2. It does not encompass,though,the modelsof Taylor(1986) and Schwert(1989),in whichthe variance
is quadraticin the absolutevalue of innovations.
3. Besides, when I<0 the absolute value of the derivative of &2(x,-I) with respect to x, -_
(= , + 2a, ,x, , ) is also higher for negative than positive x, -I. Hence, the conditional variance function is not
only asymmetric,but also steeperfor x, -, negative.The rate of growthof this derivative,though,is assumed
to be the same (=a,,).
642 REVIEWOF ECONOMICSTUDIES

ARCH and AARCH models (i.e. V = 0), only the magnitude,not the sign, of x,_ affect
i the linear standarddeviation model symmetryis only achievableif
2 In contrast, in
ca,.
p = 0 (in which case we have AARCH with a2(0) = 0), or underhomoskedasticity.
As we mentionedbefore,one of the main reasonsfor using a quadraticpolynomial
impliesa non-negativevarianceeverywhere.To see
is to ensurethat our parameterization
underwhat conditionsthe right-handside of (2) will be non-negativefor any X,_ 1q, let's
re-write(2) as:

ip'/2 0 ][I

This quadraticfunctionwill be non-negativeif and only if the matrix[,, A t"E] and

hence A, are positive semidefinite.In that case, &2(X,_1,q) is a convex functionof X,_ l,q
which reaches its minimumat any X, I,q satisfyingthe first-ordercondition AXt l,q=
- V/2, with X ilq=argmin {&(Xt1q )}. In particular,if we choose Xt'L7I,q
where A+ is the Moore-Penroseinverse of A, we obtain the non-negativitycondition
0 - ly'A+#/4 ?0. But from the practicalpoint of view we are interestedin parameterizing
(3) in such a way that the estimatedconditionalvarianceis alwaysnon-negative.This can
be achievedby writinga as follows (cf. Schwallie(1982)):

't Mt' 1q, 1)[-b' I]0' c][' I ][1 (4)

whereP is a q x q unit lower-triangularmatrix,D a q x q non-negativediagonal matrix,

b a q x 1 vectorand c a non-negativescalar,so that PDP' is the Choleskydecomposition
of A. We can also write (4) equivalentlyas = c+(X -I,q- b)'A(Xt 1,q- b), which
is particularlyconvenient if A is (band-)diagonal(cf. Engle (1990) or Campbelland
Hentschel(1992)).
The above reparameterization is certainlyunique when A has full rank. But even
when A is positivesemidefinite(as in the linear"standarddeviation"model), uniqueness
is retainedif we set to zero the elements of b' and the sub-diagonalelements of P in
columnscorrespondingto the zero diagonalelementsof D.4 It is then straightforwardto
recoverthe parametersof interest, 0, q/ and A, or some combinationsof them such as
tr (A) from the underlyingparametersP, D, b and c, by using the relationshipsA = PDP',
w = -2PDP'b and 0 = c+b'PDP'b. The delta method (see e.g. Goldberger (1991)) can be
used to derivestandarderrorsand carryout Wald tests on the parametersof interest.A
word of caution is in order, though, since Wald tests are not usually invariantto the
parameterizationused. For instance,in the QARCH(1) case, the null hypothesisof no
"leverageeffects" can be equivalentlyexpressedas vyr=0 or b1=0 when all?0. But
whetherwe actually test one or the other results in a differentnumericalvalue for the
Wald test. In a MLE framework,the likelihoodratio test is invariant,but is hampered
by its non-robustnessin quasi-MLEprocedures.Robust LM tests a la Wooldridge(1990)
offer an attractivealternative.

4. In orderto optimizean objectivefunctionwith respectto D, P, b, and c, when A is suspectedto be

singular,it is advisableto use a constrainedalgorithmwhichensuresthat d and the elementsof D are non-
negative,and drops dj, bj and 1i,(i>j) from the currentactiveset if the constraintdj>O is binding,insteadof
carryingout unconstrainedoptimizationin termsof new "free"parametersv), with d>=vj (see Gill, Murray
and Wright(1981)).
 SENTANA QUADRATIC ARCH MODELS 643

2.2. Potential reinterpretationsof QARCH models

Just as AR(r) models can be understood as first-orderTaylor approximationsto the
unknown conditionalmean function p( ), the QARCH formulationcan be intuitively
interpretedas the two leadingtermsin a quadraticTaylorapproximationto a2( ). There
is a closely relatedinterpretationin termsof projections.Providedthat E(x, IX,_ l- ) = 0,
a, is also the conditional mean of x24 E(X21X,_ ,"). Under the assumption that
E(x4) < oo, it is well known that o2 is the (least squares)projectionof x2 on B,1, where
B,tI is the collection of all measurablerandom variableswith finite second moments
generated from X,_ O.
In this framework,the QARCH formulationmay be understoodas the least-squares
projectionof x2 on a "quadraticinformationset" Q, I B,t spannedby linearcombina-
tions of the vector t,,=-=vech [(, Xt,_o)', X,_ l o )] = {1, xt_ I, Xt-2*
2
x_,
2 .. We sal shall refer tto such a projection,
rjcin
X,XlXtX2,X,iX,-3,...; x,_2,xt-2x,_3,..-3
U,Q, as the quadraticprojectionof x4 on X, Importantly,by the law of iterated
projections(see e.g. Hansenand Sargent(1991)), 7UQ is also the projectionof cr2 on Q,-t .
Hence, E(U2) E(a2Q) a2, projectionand projectionerror (i.e. CrQ and a2 _- c2) are
=
uncorrelated,and V(a2Q) < V(a2 ). Therefore,we can perhapsbetterunderstandca2 as a
smooth, filtered"estimate"of at.
As usual, the predictionerrora2 _ a2Q will be orthogonalto any linearcombination
of lag values of x,, its squares or cross-products,but not independentbecause of the
heteroskedasticity.Furthermore,all QARCHparameterscan be interpretedas the (theor-
etical) regression coefficients of x2 (or a2) on X,o I For example, consider the
QARCH(1) model where a2 =0+ Vx_1 +a, I2_. Since this can be re-writtenas
X,2= 0+ flx,t I +a, x,2_ I + i,, with 1,-x,2 _- T2, under the additional assumptions that
E(4) = 0 and all necessary moments exist and are bounded, a, I = cor ( x,4x) and 1=
cov (x2, x,t )/V(x,). Hence,a,, is a measureof the correlationin the squareseries,while
VI measuresdynamicasymmetry.An importantresultis that if the joint distributionof
X,, is symmetric and the conditional variance function is also symmetric (i.e.
a 2(Xt_I 0) = a2(-Xt_ 10,)), then the quadratic projection will be symmetric too. The rea-
son for this is that cov (X2, x,j.) = E(cr2(X,_X )x,-j) + E([X2-_a2(X,t
O )]x,j) = 0, so
yrjmust be 0 for all j. This actuallysuggeststhat cov (2, x,j) could be a useful tool to
identifypotentialdynamicasymmetrieseven if QARCH is not the truemodel, sincewhen
&2(X,_l) # a2(-X,1 this thirdmomentmay differfrom 0.
Given that the QARCH formulationis linear in ,t-, its interpretationin terms
of projectionsis rather obvious.6 What be
might perhaps less obvious is that QARCH
formulationsunderstood in this weak sense are closed under both temporaland contem-
poraneous aggregation. In this respect, the result discussed in the previousparagraph
confirmsthose of Drost and Nijman(1993) and Nijmanand Sentana(1994) in the context
of temporaland contemporaneousaggregationrespectivelyof GARCH processes.These
papersshow that if suchtermsdo not appearin the quadraticprojectionsof the underlying
series, then they do not appear in the projectionof the aggregatedone either. But it is
easy to prove that if one series shows dynamicasymmetry,then so will the aggregated
one (see Drost (1993)).

5. DrostandNijman(1993)WeakARCHconceptis a particularcaseof whatwecallquadraticprojections,

in whichlinearand cross producttermsare not included.
6. The same is true of AR(r) models,whichmay be understoodas approximatingthe conditionalmean
p (X, l.r) by the Minimum Mean Square Error Linear Predictor of x,, say p,CL= a + y'X,_ I,. Again, this yields
the familiarinterpretationof yj as a (theoretical)regressioncoefficient.
644 REVIEW OF ECONOMIC STUDIES

2.3. Estimation and testing for QARCH effects

Given the interpretationof QARCH models as quadraticprojectionsof x2 on X,_loc, a
consistentmethod of estimatingthe parametersis providedby the OLS regressionof 4
on the relevantelementsof X, - (cf. Weiss (1986)). The preferredmethodof estimation
for ARCH models, though, has been maximumlikelihood,but since this involvesa non-
linearprocedure,it is of some interestto have a simplepreliminarytest for the presence
of QARCH effects.For the ARCH(q), Engle (1982) proposedan LM test which can be
computedas TR2 of the OLS regressionof x4on a constantand its first q lags. This test
is distributedas a x2 underthe null of no ARCH even if x, is not conditionallyGaussian
providedthat the fourth conditionalmoment of x, is constant and finite (see Koenker
(1981)). Bera, Higginsand Lee (1992) have extendedthis test to the case of AARCH(q)
by includingcross-productterms of the form x,-ix,-j in the above regression,resulting
in a Xq(q+ 1)/2 null distribution.It is straightforwardto check that if we also add the first
q lags of x, as regressors,an LM test for QARCH(q)is obtained,whichwill be distributed
as q(qq+3)/2 under homoskedasticity.It is worth noticing that this test is the dynamic
analogueof White's(1980) generaltest for static heteroskedasticity,which suggeststhat
it may also have good poweragainstmost other dynamicheteroskedasticalternatives.As
White's test, it can also be derived as a test for random coefficientsor as a selective
dynamicinformationmatrixtest (cf. Beraand Lee (1993)).
But as in the ARCH(q) model (see Demos and Sentana (1994)), some of the true
parameterslie at the boundaryof the parameterspace under the null of no QARCH.
Although this does not affect the distributionof the LM test, intuition suggests that a
more powerfultest could be achievedby taking the (partially)one-sidedalternativeinto
account.For the sake of claritylet's considertestingagainsta QARCH(1)alternative.As
we have seen, the two-sidedLM test is based on the regressionof x4on a constant,42_I
and x,- 1. Notice that if x, is unconditionallysymmetricallydistributed,the regressorsare
orthogonalunder the null, and we would expect each OLS coefficientto take any sign
independentlyof the other. But under the alternativewe would expect the coefficientof
X4 to be positive. Hence, a partiallyone-sided test of Ho: ==a,, =0 vs. H,: V/,I0,
a,, > 0 seemsmore appropriate.The test statisticwill be the resultof addingto the square
of the t-ratio associated with x,_1, the square t-ratio for x2_1 when the coefficientis
positive (cf. Yancey et al. (1980)). Under the null this statisticis a 50:50 mixtureof x2
and X2
All the above tests may have the wrong asymptoticsize if the assumptionof condi-
tionalhomokurtosisdoes not hold. Nevertheless,Wooldridge(1991)has recentlysuggested
a simple modificationthat can be used in the two-sidedLM tests for QARCH to allow
for heterokurtosisunderthe null. His robustprocedurewouldinvolveregressinga constant
on the product of de-meanedx2's times de-meanedlagged values of x,, its squaresand
cross-products,and using T- SSR as asymptoticallyX2(q?3)/2, where SSR is the sum of
squaredresidualsfrom the regression.
In some cases, more robust versionsof one-sided LM tests are also possible. Let's
consideragainthe QARCH(1)exampleabovewith an unconditionallysymmetricdistribu-
tion, and let's supposefor simplicitythat the conditionalkurtosis,K,, is a functionof x, 1
only, say KC(Xt, 1). A partiallyone-sidedLM test could be computedfrom the sameregres-
sion as Wooldridge (1991) by adding to the square of the t-ratio associated with de-
meanedx, 1 timesde-meanedx2, the squaret-ratiofor de-meanedx2_1 timesde-meaned
7. It can be provedthat the distributionof the correspondingLR and Wald tests is the same. To do so
it is moreconvenientto workwith the one-to-onereparameterization a, = o + 20o0x,- I + (O,, + 4021
/4o)x,.1
If x, is not unconditionallysymmetric,the mixingweightsin all threetests will change.
 SENTANA QUADRATIC ARCH MODELS 645

4x when the coefficientis positive. Under the null of conditionalhomoskedasticity,this

statisticis a 50:50 mixtureof x2 and x2 even if ictis time-varying,providedthat K( ) is
an even functionof x, 1. Anotherimportantcase ariseswhen we have an ARCH(l) null
and want to test if V/1is significantlynegative.The robustifiedtwo-sidedLM test would
consist in the squared t-ratio from the regressionof 1 on the regressionresidualsof
1 on a constantand
(x0_l/u)
(x,/&l)- times the regressionresidualsof (x, '/c) on a
constantand (x '/at), where a iS the estimatedARCH(1) variance.If the conditional
distributionof x, is symmetric,we would expect the regressioncoefficientto be negative,
and hence a more powerfulLM test could be computedas the squaredt ratio when the
coefficientis negativeand zero otherwise.Such a test would have a 50:50 mixtureof %
and x2 distributionunderthe null, and the relevantcriticalvalue would be 2 7 insteadof
3-8. In both cases, though, thereis a trade-offbetweenthe increasein power from using
the one-sidedtest and the reductionin robustnessderivedfrom the (weaker)maintained
assumptionsabout higher-ordermoments.

2.4. GQARCH(p, q) models

So far we have maintainedthe assumptionthat the relevantinformationset containsonly
a finite numberof lags, q. But this may be too restrictive(for instance,Nelson's (1992)
consistencyresultsdependon q beingunbounded,and the projectioninterpretationwould
also generallylead to infiniteq). Besides,even if q were finite, estimatingthe model for
largeq will be difficultas the numberof parametersin A is 0(q2), and we would need to
impose some structureon this matrix.For example,one could introducea rankk (k < q)
structurein A (e.g. if k = 1, we would thus obtain the linear"standarddeviation"model)
but even in that case, the numberof parameterscould still be excessivefor q large.Perhaps
the most naturalway to solve this problemis by introducinga decliningstructureon the
coefficients.Just as in conditionalmeanmodels,this structurecan be eitherrapidlydecay-
ing or moreslowly so (as in Robinson(1991) or Baillie,Bollerslevand Mikkelsen(1993)).
In particular,an exponentiallydeclininglag structurecan be obtainedby includinglagged
values of o2 on the right-handside of (3), as in Bollerslev(1986). That is:
( = 0+V11X,-1,,q+Xt1-1,qAX,_1,q+ffij 15.72 (5)

By analogy with Bollerslev's(1986) GARCH and Bera and Lee's (1990) GAARCH
models, we shall term these models GeneralizedQARCH models of ordersp and q, or
GQARCH(p, q) for short. As in the case of ARMA models, theseGQARCHmodelswill
generallyresultin longermemorymodels with a flexiblelag structure,which,at the same
time,could offera moreparsimoniousapproximationto the conditionalvariancefunction.
For instance,the generalGQARCH(1,2) model

(1 -SL)r2=O + yvx,t + yf2x,t2+al1x,, +a22x,_2+2a12x,tIx,t2 (6a)

is equivalentfor S1< 1 to the following tri-diagonalQARCH(oo)model:
o =0/(1-31)+ ViXt_I+a1, 0/2(v 2+S1 ,)37l X,-j

+Z02(a22+SaII)2 8aa2112x,_jx,1xj2.+2+2 (6b)

For higher-order(invertible)models the pattern of coefficientsof A,, and W will

becomemore complex,but notice that A,, will be band-diagonalwith 2q (off-)diagonals,
whereq is the order of the QARCH part. As Drost and Nijman (1993) and Nelson and
646 REVIEWOF ECONOMICSTUDIES

Cao (1992) point out for the standardGARCH(p, q) model, requiringthe positivityof
the finite QARCH part plus Sj>O for all j is unduly restrictive except in the
GQARCH(1, 1) case. Conditions for the positivity of the conditional variancein (5),
could be obtained as an extension to O., ip.Oand A,,, of the methods for finite q in
Section2.1.8

3. PROPERTIESOF THE QARCH MODEL

3.1. Stationarity conditions and moments of the unconditionaldistribution
It is possibleto prove (see Appendix)that a GQARCH(p,q) model is covariancestation-
ary if E'7 aii+ 8j is less than 1. The unconditionalvarianceof x, is then given by
-
V(x,) = 0/(I -LE,= 1 ai i= 8j). 7

Notice that the covariancestationarityof x, does not dependat all on the linearterm
in the conditional variance, #P'X,- I,q, only on the quadratic term associated with the matrix
A. Looselyspeaking,it is as if the quadratictermasymptoticallydominatesthe linearone.
Besides,the stationarityconditionfor GQARCH (and hence GAARCH) is the same no
matterwhat the off-diagonalelementsof A are, and so it coincideswith that of the nested
GARCH model. Notice also that the actual value of the unconditionalvariancedoes
not depend on yi. Besides,as (7) does not depend on the off-diagonalelementsaij, the
unconditionalvarianceof a GAARCH(p, q) process equals the unconditionalvariance
of the GARCH(p, q) process obtained from it by setting the off-diagonalelements of
A toO.
The formalproof of the above resultis obtainedby re-writingthe GQARCHprocess
as a randomcoefficientsmodelas in Tsay (1987) and Beraand Lee (1990). But an heuristic
proof of (7) may help us understandthe reason for the similarityof the unconditional
variancesfor GQARCH, GAARCH and GARCH models. Let's suppose that E(x,)
is bounded and thereforeequal to a2 = E(a2 ). Then, if we use the parameterizationof
the conditional variance given in (8) (i.e. _t=O+Z y,1,x, = aiix2_i+
2 E<. a,-x,_ x, +>j a2_j), take expectationsat both sides and solve for a2, we get
exactfyexpression(7$ as both the linear termsand the cross-productsvanish becausext
is a zero-meanuncorrelatedprocess.
As in standardGARCH models, the sum Eq aii+ Ej=I 3Aj providesa measureof the
persistenceof shocks to the varianceprocess (but see Nelson (1990)). Hence, we can
analogouslydefine IntegratedGQARCH process as those for which this sum is 1 (cf.
Engle and Bollerslev (1986)). For instance, the IGQARCH(1,1) will be defined as
2 -
+6U
+_yt,x I + (1 -
a, = 0 + )X, + 3_ , which shares with the IGARCH( 1, 1) the property
that E(2+j IX,, )=jo + a2 9
In orderto discusshighermoments,let's definex* as the standardizedvariableassoci-
ated with x, (i.e. x* = x,/at). So far we have mostly assumedthat E(x* IX,_l =0, and
E(x7*2IX,1,o ) = 1, but if we assumethat x, X,_ is symmetric,so is the unconditional
distributionof x,. To obtain unconditionalfourth moments,assume for simplicitythat
x* is i.i.d. with finite fourth moment K. Provided that the appropriatemoments are
8. See Nelson and Cao (1992) for some GARCH cases and Demos and Sentana (1991) for several
GQARCHones.
9. IGARCHprocessesareclearlynot covariancestationary,but theyarestrictlystationaryandergodic(cf.
Nelson (1990)andBougerolandPicard(1992).Giventhatthebehaviourof GQARCHprocessesis dominatedby
the quadraticterms,one wouldexpecta similarresultto be true of IGQARCH.
 SENTANA QUADRATIC ARCH MODELS 647

bounded,by Jenseninequalitythe degreeof leptokurtosisof the unconditionaldistribution

of any zero mean conditionallyheteroskedasticmodel is higherthan the degreeof lepto-
kurtosisof the assumeddistributionfor x*. In general,though,it is very tediousto derive
closed-formexpressionsfor the leptokurtosisof GQARCHmodels.
The GQARCH(1, 1) processis an importantexception.Assumingsymmetryfor x,*,
it is possible to show that, provided Ka21+ 2a11, +432< 1, the unconditionalfourth
momentof xt is given by:

- 1cl, - 2a 1s,
(I -O- [0(1+ a,I+1 ) + 21 (8)
which reducesto the expressionin Bollerslev(1986) for v' =0 and ic=3 (i.e. normality).
However,if V,I#0, the GQARCH(1, 1) processis moreleptokurticthanthe GARCH(1, 1)
model which it -nests,although the condition for boundednessof fourth momentsis the
same (see Bollerslev(1986)).

3.2. Dynamic correlation structure

QARCHprocessesareuncorrelatedbut not seriallyindependent.As an example,Bollerslev
(1986) proves that in the GARCH(p, q) case the autocorrelationfunctionsfor x2 corre-
sponds to that of an ARMA [max (p, q), p], and suggestsusing the sampleautocorrela-
tions as a tool for tentativelyselectingthe ordersp and q. In the generalGQARCH(p,q)
model,though,the processfor x2 is morecomplicated.Nevertheless,underthe assumptions
that x* is symmetricallydistributedand the unconditionalfourth moment of x, is finite,
it is easy to see that for k > max (p, q - 1) the autocorrelations,Pk, follow the same Yule-
Walkerrecursionas the GARCH(p, q) model, whereasfor k<max (p, q- 1), they will
depend on all the parameters (0, Vi, A, 8).
In particular,the autocorrelationsof the squaresof the GQARCH(1, 1) processare
exactly the same as those of the GARCH(1, 1) model. As a matterof fact, the similarity
between GQARCH(1, 1) and GARCH(1, 1) is remarkable:they both have the same
unconditionalmean, varianceand autocorrelationfunctions for both the series and its
squares. Nevertheless,the GQARCH(1, 1) has the advantagethat with a single extra
parameter(Vyr)it allows for both an asymmetriceffect on the conditionalvarianceand
higher kurtosis than the GARCH(1, 1) process which it nests, and therefore,it goes in
the rightdirectiontowardscapturingsome of the stylizedfacts characterizingmanyfinan-
cial time series.
As we saw earlier on, the asymmetryin the conditionalvariancecan be captured
more formally by dynamic third moments of the form cov (X2, x,_j) = pfoj , where
ij=cov
= (xt, x,t , x,_j) denotes the bicovariancefunctionof xt (see Hinich (1990)).?0In
the GQARCH(p, q) case, 3o,j also follow an analogueof the Yule-Walkerrecursionsfor
j> q. This apparently surprising result is again due to the quadraticstructureof the
conditionalvariancefunction, in which the squaretermseventuallydominateall others.
These moments, though, have to be understoodas unconditional,and thereforeunable
to capture conditional asymmetries.The AARCH(2) model provides an interesting
examplein which flojj=0 for allj, but E(x x,1 IX,t-2,)=2aI2x,t-2t-i*1.

10. Notice that our assumptionsimply that the remainingelementsof the bi-covariancefunctionare all
zero.
648 REVIEW OF ECONOMIC STUDIES

4. UNIVARIATE APPLICATIONS
4.1. Daily U.S. stock market returns
The analysisof financialtime serieshas turnedout to be the most fruitfulapplicationof
conditionallyheteroskedasticmodels (see Bollerslev,Chou and Kroner(1992) for a recent
survey). ARCH-type models have been used mainly as a reduced-formdescriptionof
conditionalvariances,but more recently,they have also been used as buildingblocks for
structuraleconomicmodelswith time-varyingvolatility.The asymmetricmodel of chang-
ing volatility in stock returnsof Campbelland Hentschel(1992) providesan illustrative
examplein which the QARCH model discussedhere is used to develop a formal model
of volatilityfeedback.
As part of their empirical exercise, Campbell and Hentschel (1992) estimate a
restrictedversionof the generalGQARCH(1, 2) model in (6a) for daily U.S. stock market
returnsoverthe period1926-1988(a total of 16,981observations).For tractabilityreasons,
they impose the restrictionsa,2=0 and YV2/a22= =V, /a,l, so that their model can then be
written as 52 = c+a11 (xt-, -b, )2 +a22(X,- 2-b, )2 + S1o-,2_ l. This restricted model gener-
ates non-negative variances whenever c, a,,, 3,> 0 and a22?-3,a,i (see Demos and
Sentana(1991)). Campbelland Hentschel(1992)findthat sucha modelappearsto capture
most of the "leverageeffect",and that unlike a symmetricGARCH model, it does not
significantlyoverestimateaveragerisk premia.12
In theirapplication,the parameterb,, which measuresdynamicasymmetries,is esti-
mated to be positiveand very significantlydifferentfrom zero in a GaussianMLE frame-
work. But their results also suggest that the maintainedassumptionabout conditional
Gaussianityis rejectedby the data. Althoughthis does not make the normalpseudo-MLE
inconsistent,the sizes of standardWald, LM and likelihoodratio tests could be affected.
For that reasonwe have computeda robustversionof the Wald statistic(as in Bollerslev
and Wooldridge(1992)), which shows that the significanceof b, is not an artefactof a
misspecifieddistribution.'3
We next test their restricted parameterization against our most general
GQARCH(1,2) model. A robustLM test of the two restrictions(see Wooldridge(1991))
producesa valueof 7 32, whichis significantat the 5%level (X2o o5= 5 99). The parameter
estimates of the general model together with standarderrors are reportedin Table 1.
Importantly,we have estimatedthe model both in termsof the free elementsin P, D, b
and c, as well as in terms of the parametersof interest 0, A and yr directly.The first
procedureguaranteesnon-negativevariancesin and out of sample(see Demos and Sentana
(1991) for details) while the second does not. In our case, both set of estimatescoincide,
which is reassuring.

11. Conditionalasymmetriesare importantto distinguishhow differentARCH modelscaptureleverage

effects or other dynamic asymmetries.For instance, in the conditionallynormal GQARCH(l, 1) model
cor (a2, , X,-2., ) approaches ?1 as a2- 1-*O, and 0 as -2 00oo. Hence, "leverage effects" are relatively
more important in times of low volatility. In contrast, cor (in a2, x,I X,-2,, ) is constant for the EGARCH
model, so these effectshave the same importanceregardlessof the level of oJ
12. To understandwhy, considera GARCH model with conditionalmean p in which the conditional
varianceis modelledin terms of squaredinnovations(x, -_P)2. Here, p is also forced to capturedynamic
asymmetries,and as a result,it is usuallyoverestimated.UnlikeGARCH and GAARCHmodels(cf. Beraand
Lee (1990) and Tsay (1987)), a quadraticARCH is obtainedfor linearconditionalmean modelswhetherwe
use innovationsor the actualseriesto model volatility.
13. Ourdata set differsslightlyfromthe one in Campbelland Hentschel(1992), in that we use arithmetic
daily returnsinsteadof geometricreturnsover the monthlyaverageof the safe rate, but this shouldnot affect
our qualitativeconclusions.Our resultsare also insensitiveto relaxingthe assumptionof constantconditional
meanby includingthe conditionalvarianceor an MA(I) component.
 SENTANA QUADRATIC ARCH MODELS 649
TABLE I
parameterestimatesof a GQARCH(1,2) model
likelihlood
Gaussianpseudo-maximum
U.S. daily stock returnsJanuary1926-December1988(16,981obs.)
2 = y+
2_ -
at 6+ V16,-1 + V28,-2+a,,e,-1 +a22,--2+2a12e-19,-2+3,t-,

=0/(I -81i ) +a,l2_ tE,- a2 lll)12_j

+ yi,x,_ | +,2
i' )3
VIr
(V2 + +52 2x,1 a124 X,jXt,-j1.-

Parameter Estimates
r 005352 (0 00543) [0 005731 {0-00541}
6 0 00859 (0 00075) [0-000611 {0.00095}
VIl -0 09857 (0.01013) [0 006381 {0-02010}
V2 0-06119 (0-00996) [0-006881 {0.01538}
all 0-13472 (0-01026) [0-005201 {0.02088}
a,2 -0-00029 (0.00263) [0-001821 (0 00387}
a22 -0 05926 (0-01056) [0.017971 (0 02073}
6, 0 91790 (0 00362) [0-002541 {0.00532}
Log-likelihood-21,100 07
Note: Hessian-basedstandarderrorsin parenthesis,outerproduct-basedstandarderrorsin square
brackets,robuststandarderrorsin curlybrackets.

Then we check whetheran even more generalGQARCH model is required.On the

basis of robust LM tests, it seems that extra terms in ,-t3, Ei_26,.3 and s,_ I?t3 are
required. Parameter estimates together with robust standard errors for such a
GQARCH(1,3) model are reported in Table 2. The "persistence"measure given by
all +a22 +35l(=0 99389) is very close to 1, but a robust Wald test indicates that it is
significantlysmaller.'4Note that the estimated model is effectivelya 5-band-diagonal
QARCH(oo) model, in which the coefficients for yfj, at,, aj, I and ajj12 decay exponentially
at the rate 31 afterjust a few free lags. Interestingly,the effectsof the first lags of levels,
squaresand cross-productsare the largest,since the estimatesof 3, a22 and a23have the
opposite signs to those of I/fI, a Il and a,2 . In the case of cross-producttermsof the form
E,_jc,_j_ the effect is dramatic:they practicallydisappearafter the first lag. In fact, a
robust Wald test confirmsthat a23+ 3Ia,2 is not significantlydifferentfrom zero, which
explainswhy the coefficientson c,_jct_j_I wereinsignificantin the GQARCH(1,2) model
in Table 1. The cross-productterms 61_jE6_-_2 have also very smallcoefficients(000832,
000767, . . . ). The leverageeffectis also shortlived, sinceaftertakingthe values -0' 09483
and -0-08911 for the first two lags, it drops to -002070, -0-01907, etc. In contrast,the
ARCH effectdecaysless dramaticallyfrom 0 14286to 0'06084,0 05608, etc. Thereforeit
seems that there is an implicitrankingof effects:the standardARCH effect is the most
important,followed by the "leverageeffect"and finallythe cross-productterms.
To see if the estimatedmodel adequatelycapturesthe autocorrelationin the square
returnswe have computed Ljung-Box statisticsup to 20-th order for the standardized
residuals.We also use robustLM tests to see whethernon-quadratictermssuch as IE?-I i,
Is,-21 and I?- l are requiredin the variance.Since we alreadyhave ?,_- , ?,-2 and 5t-3
in, this equivalentto testingan alternativemodel that imposesno restrictionsin the way
positiveand negativelaggedinnovationsaffectthe conditionalvariance(cf. Engleand Ng
14. Recentworkby Lumsdaine(1993) and Lee and Hansen(1994) showsthatthe Waldtest for IGARCH
has an asymptoticnormaldistributionin the GARCH(1, 1) model. We assumehere that the same is true for
the QARCHmodel, althoughit would be non-trivialto prove.
650 REVIEWOF ECONOMICSTUDIES
TABLE2
of a GQARCH(1,3) model
likelihoodparameterestinmates
Gaussianpseudo-maximum

U.S. daily stock returnsJanuary1926-December1988(16,981 obs.)

r,= r + E,.
ta2= + a,,l4, +a226,2-2+ VI S,-, + /26-2 + /3gt- 3

+ 2al26,_l,-,2 +2a238,_26t_3 +2a,3E,- ICt_3+ ,a_1

- 9/(1 - 3, ) + a,l c2,-1 + (a22 +3 IaII ) 7j=0 I t-2-j

+2a2e, -,e,-2+2(a23+3ala2) E.j 8 16,-2-j-,-3-j +2a13 jOO 5jl,'e-l-jt-3-j-

Parameter Estimates
r 0-05208 (000540) [0-005701 {0.00537}
0 0-00805 (0-00072) [0 00060] {0.00090}
VI, -0-09483 (0-01028) [0 006391 (0.02036}
V2 -0-00169 (0-01450) [0-009401 {0.02786}
Y/3 0-06145 (0-00920) [0-006181 {0-01768}
al, 0-14286 (0-01055) [0-006381 {0-01998}
a,2 -0-02709 (0.00751) [0-005541 {001 104}
a22 -0-07086 (0-01077) [0-006811 {0-01961)
a,3 0-00832 (0-00252) [0?001951 (0 00355}
a23 0-02842 (0 00731) [0-00546] (0-01058}
3, 0 92189 (0 00345) [0.002431 (0005101
Log-likelihood-21,061 18
Note: Hessian-basedstandarderrorsin parenthesis,outerproduct-basedstandarderrorsin square
brackets,robuststandarderrorsin curlybrackets.

(1993) or Hentschel(1994)). Both sets of tests suggestthat the purelyquadraticformula-

tion adequatelydescribesthe conditionalvariance.

4.2. Monthly U.K. stock market excess returns

We also look at the monthly FTA500 excess returnseries, rt, for the period 1971:2 to
1990:10. Given the small numberof observations(237), we initiallyfit a GARCH(1, 1)-
M model on the basis of a Gaussianpseudo-log-likelihoodfunction (see Table 3). Note
that the priceof risk coefficientis positiveand significantlydifferentfrom zero. However,
as in the U.S., the averageestimatedrisk premiumis substantiallyhigherthan the mean
excess return.
It is worth mentioningthat any GARCH-M model alreadyallows for correlation
betweenr2 and r,_,, since a2 is symmetriconly in the innovations.To check if there is
evidencefor an extradynamicasymmetriceffect,we have computeda robustLM test for
the inclusionof Et- I in the conditionalvariancealong the lines of Bollerslevand Woold-
ridge (1992).'5The x2 test statistictakes the value of 3-9, which suggeststhat it might be
worth estimatinga GQARCH(l, 1) model.
The parameterestimatestogetherwith robust standarderrorsare reportedin Table
3.16 Note that the Waldtest (t-ratio= -2-73) confirmsthat yris significantlynegative.Also

15. As the varianceparametersalso affect the mean, the computationsare more complicatedthan in
Section4.1 becausea matrixregressionis necessary.
16. In the GQARCH(I, 1) case, positivityof the varianceis achievedif a,,, 3, _0 and V1'<4a,1O. To
impose these restrictions,we have estimatedthe likelihoodfunctionwith o = c2+1,1(e, ,-b, )2+dlda,2,, so
that O=c2+l12l,b2,a, =1,2;,3a=d2 and =-2l1,2b,.
 SENTANA QUADRATIC ARCH MODELS 651
TABLE3
likelihoodparameterestimatesof
Gaussianpseudo-maximum
GARCH(I, l)-M and GQARCH(l, l)-M models
U.K. monthlyexcessstock returns1971:2-1990:10(237 obs.)
r,=pa2+ _e.
CF2-a,~+ VI E_ I+a?, 12 + 41,_ .2

GARCH GQARCH
Parameter estimates estimates
p 024348 0-16669
(0.07728) (0 08091)
[0 080671 [0-088151
{0*10514} {0.09905}
0 0-08926 0-14682
(003797) (0 05159)
[0-052051 [0-040561
{0.059281 (0-11980}

Y/t -0-23969
(0.06786)
[0-081441
{0-087781

all 020752 0-09783

(0-07198) (0-05324)
[0.099411 [0-057341
{0-09690} {0-06994}

31 0 67928 0-66351
(0 08757) (0-11312)
[0146861 [0-094711
{0.07172} {0-22878}
Log-likelihood -275 147 -268-081
Note: Hessian-basedstandarderrorsin parenthesis,outer product-
based standarderrorsin squarebrackets,robust standarderrorsin
curlybrackets.

note that the ARCH parametera,, is smallerin this case, whilethe GARCH parameterS,
is roughly the same, which taken togetherimplies a smallerpersistenceof the volatility
shocks. The differencebetween both models is even more obvious graphically.Figures
1(a) and I(b) plot the conditionalstandarddeviationof returnsimpliedby the GARCH
and GQARCH models around the two most significantepisodes in our sample: the
October1987crash (a 26 3% drop in stock prices)and the January1975 bounceback (a
51-6% surge). Both models capturethe effectof the 1987crash in a very similarmanner,
and the same happens for the bearish 1974. In contrast, the period immediatelyafter
January1975 is noticeablydifferent.The problemwith the GARCH(1, 1) model is that
it treatsboth episodesin the samemanner,despitethat the restof 1975was not particularly
volatile.
The implicationsfor risk premia are also different.The price of risk parameteris
smallerin magnitudeand less preciselyestimatedin the GQARCH(1, 1) model, but still
significantlypositiveif we considerthe relevantone-sidedtest.Again,the averageestimated
risk premia is much closer to the mean excess return.Figures 2(a) and 2(b) show the
652 REVIEWOF ECONOMICSTUDIES
3
(a)

2*5
GARCH(l,l)

1X5

05

0
1 73 1 74 1 75 1 76

EstimatedConditionalStandardDeviation

3
(b)

25

1-5

I - GARCH(l1,1) | \

0.5
GQARCH(1,1)

86 i 87 1 88 1 89
EstimatedConditionalStandardDeviation

FIGURE I
 SENTANA QUADRATIC ARCH MODELS 653

differencevery clearly.Like in the rest of the sample,the GARCH modelimpliesa higher

requiredreturnthan the GQARCH one around October 1987, but with a similartime
profile.However,in 1975the GARCH modelimpliesa maximumrequiredreturnof 200%
on an annualizedbasis, and a much slower reversionto the mean, which seems rather
implausible.17

5. MULTIVARIATEEXTENSIONS
Most of the existing literatureon ARCH models (includingthe precedingsections) is
concernedwith univariatemodels. Given that many issues in finance,and in particular,
asset-pricingtheories,are relatedto the variancesand covariancesof many assets, it is of
the utmost practicalimportanceto be able to extend univariatemodels so as to capture
time-variationin the (conditional) mean vector and covariancematrix.
Multivariategeneralizationsof the QARCHmodelare straightforward in theory.Let
y, be a multivariatestochasticprocess of dimensionm whose conditionalmean is 4,=
E(y,t Yt- 1,)-p( Y,t ,,) and whose conditional covariance matrix is t-= V(y, Yt- l ,) =
I( Y,- loo) where Y,- l= vec (y,- I, y,-2,. . . ). Again, for the sake of clarity,we shall
deal initiallywith the case in which the dependenceof the conditionalmomentson the
past is limited to a finite number of lags of y,. Let Y,_,q=vec (Y,- l, Yt -2, .. ., y,-q) be
the mq x 1 vector containing the values of the m series for those q lags, and let s, =
vech (I, ) = vech [(( Y,t- )]) S(ft,- l,q), be the vector-valuedfunctionwhich containsall
the distinctelementsof the conditionalcovariancematrix.SinceI, containsm conditional
variancesand m(m- 1)/2 differentconditional covariances,the dimensionof s, is n =
m(m+ 1)/2. For simplicitylet's assumethat p(Y,- ,,,)=0 so that s,=E,-, [vech(y,ty').
It is again clear that only an even-order polynomial can guarantee the positive
(semi-)definiteness of the conditionalcovariancematrixfor all possiblevaluesin the condi-
tioning set. On this basis, we can definea multivariateQARCH model by using second-
orderpolynomialsas follows:
S(Yt-1,q)= 0+ TYt - 1q +(In 0Yt'_1,q)A Yt -1,q (9)

whereE)is a n x 1 vector, T a n x mq matrixand A a nmqx mq matrixof coefficients.By

analogy,we shallcall this formulationthe MultivariateQARCH(q)model.'8If we partition
the matrixA in n squareblocks of size mq, and if, in additionto T being 0, each one of
these n blocks is in turna block-diagonalmatrixwith q squareblocksof size m, the right-
hand side of equation (9) reducesto the multivariateARCH(q) model introducedby
Kraft and Engle (1983):
s,t= Bo+ ,q_ Bicot-i (10)
where Bo is a n x I vector and the Bi's are n x n matricesand c,= vech (y,y,). On the
otherhand,whenT = 0 but the blocksof A arenot block-diagonal,equation(9) constitutes
a multivariategeneralizationof the AARCH model.
of S, can be obtained
Sufficientconditionsto guaranteethe positive(semi-)definiteness
by rewriting the multivariate QARCH(q) model above as a randomcoefficientsVAR(q)
model for y,. Just as in the univariatecase, one can then reparameterizein terms of the
17. Again we have computedLjung-Boxstatisticson the standardizedresidualsand tested for lagged
absoluteinnovationsin the conditionalvarianceby means of robustLM tests withoutbeingable to rejectthe
GQARCH(1,1) specification.
18. Infinitedependenceon the past can again be achievedby includinglaggedvaluesof s, on the right-
hand side of (9) as in Bollerslev,Engleand Wooldridge(1988).
654 REVIEWOF ECONOMICSTUDIES

2
(a)
1-75

1-5

1\25

1 GARCH(l,1)

075

0-5

0225
1 73 1 74 1 75 1 76
EstimatedRisk Premia

2
(b)
1-75

1*5

1*25

0*75 -

- GARCH(l,l)
05

-
025

0
GQARCH( l, l)
I 1 1A I I I I I I I I I I II I I ,II I I
-0-25 I I

I 86 1 87 1 88 1 89
EstimatedRisk Premia

FIGURE2
 SENTANA QUADRATIC ARCH MODELS 655

Choleskydecompositionof the covariancematrix of the randomcoefficients.The main

practicalproblem,though,is the sheernumberof parametersinvolved.Eventhe multivari-
ate ARCH(q) model containsn + qn2parameters,and in practicefurtherrestrictionshave
been imposed (e.g. the diagonalARCH model used in Attanasioand Edey (1987), Bol-
lerslev,Engle and Wooldridge(1988), or the BEKK model in Engleand Kroner(1995)).
But as in standardmultivariateARCH models, thereare severalalternativesto sim-
plify the multivariateQARCHmodel above. One would be a k-factorQARCHmodel (as
in Engle (1987)) in which k linear combinationsof the y,'s follow univariateQARCH
processes.Another would be a model in which the variancesfollow QARCH processes
but the conditionalcorrelationstructureis heldconstant(as in Bollerslev(1990)). Finally,a
conditionallyheteroskedasticfactormodelof the type introducedby Diebold and Nerlove
(1989), and extendedby King, Sentana and Wadhwani(1994) but with QARCH-type
effects on the underlyingfactors can also be entertained.Such a model is particularly
appealingin financialapplications,since the concept of factorsplays a fundamentalrole
in major asset pricing theories such as the ArbitragePricing Theory of Ross (1976).
Besides, it automaticallyguaranteesa positive (semi-)definiteconditional covariance
matrixfor y, once we ensurethat the conditionalvariancesof the factorsare non-negative.
However,muchcare has to be exercisedwhendealingwith conditionalvariancesthat
dependon past values of the underlyingfactors,ft, as the truevalue of the factorsdo not
necessarilybelong to the econometricianinformationset, Y, 1,". An argumentsimilarto
that in Harvey,Ruiz and Sentana(1992) shows that if the varianceof each of the factors
conditionalon an informationset that also containspast valuesof the factorsis given by
a GQARCH(p, q) model like (5), the variancesof each of the factors conditional on
Y,- ,oot,Rt-I= V(ft1,IY,1, ), will be given by:
At,t_-1= OI+y = l Ii i t + Ei#oja1yfi,-iit-ift-ft-j
-i
+ E. aziiffitr-il t- i + clt-it - + 4"lt
li tflj (1

wherefit - ilt-j = E(fi,t- i Yt-j,,, ) and co_

0l h-ii-j= V(fi,,- iIYt-oo ).
Notice that the only differencebetweenexpression(11) and a pure GQARCH(p,q)
variancelike (5) is the inclusionof a correctionin the standardARCH termswhichreflects
the uncertaintyin the factor estimates,but not in the AARCH or linear terms. Under
some additionalassumptions,this correctioncan be easilyevaluatedvia the Kalmanfilter.
In this respect,it is worth mentioningthat the Kalman Filter prediction,updatingand
smoothing equations derivedin Harvey, Ruiz and Sentana (1992) for the latent factor
modelunderthe assumptionthat the conditionaldistributionof the factorsis proportional
to a (standardized)multivariatet remainvalid here since they do not actuallydependon
the particularfunctionalform adopted for the conditionalvariances.But a non-trivial
advantageof the QARCH formulationin the context of conditionallyheteroskedastic
factor models is that it would be very difficultto handleanalyticallyany functionalform
other than the quadraticfor the conditionalvariancesof the underlyingfactors.

6. A MULTIVARIATEAPPLICATIONTO U.K. STOCK RETURNS

As we have seen, one of the strengthsof the QARCH formulationis that it is also able
to detect dynamicasymmetriesat a multipleasset level which could not be capturedby
multivariateARCH models.To gauge theirempiricalrelevance,we havejointly estimated
by Gaussian pseudo-maximumlikelihood a conditionallyheteroskedasticlatent factor
656 REVIEWOF ECONOMICSTUDIES

model of the type discussedin Section 5 for excess stock returnson 26 U.K. sectorsfor
the period 1971:2 to 1990:10 (see the Data Appendixfor details).
Given the relativelysmall number of observations(just over 200), we have only
consideredGARCH(1, 1) and GQARCH(1, 1) parameterizations for the commonfactor,
and constantconditionalvariancesfor the idiosyncraticterms.'9To model the conditional
mean, we use the dynamicversionof the APT in King, Sentanaand Wadhwani(1994).2?
Specifically,if r, is the vector of 26 sectorialexcess returns,the model that we estimate
can be writtenin vector notation as:
,p+ cft+wJtC(pt1t-,1
rt= CA1,-% +f,) + wt (12)
where f, is the common unobservable factor, w, represents idiosyncratic risks, c is the
vector of factor loadings, p is the price of risk parameter, , , = V(f,lr,t,...)=
O+iv,E(f,lir,t,,.. .)+a,iE(fr llIrt-,,.. .)+361A,)-1t-2, V(w,Ir,_,,. . .)=F diagonal
and E(ffw,j r, - I,. . 0.2

The novel resultthat we obtain at the multivariatelevel is that there appearsto be a

significantleverageeffect in the sector returnsthroughthe common unobservablefactor.
The robustWaldstatisticfor VY/ = 0 is 6 02. Comparingthe two estimatesof the conditional
variancesof the unobservablefactor,the "leverageeffect"is againparticularlynoticeable
around January1975. The estimateof p is positive for both conditionalvariancepara-
meterizations,but once more it is too largein the GARCH case, whichleadsto overestim-
ated risk premia.
The estimatesof the commoncomponentin the factormodel (i.e. pAt,- +f,t)for the
GQARCH(l, 1) and GARCH(1, 1) formulations,though, are remarkablyclose (r=
0.999). Importantly,the correlationof the common componentwith the FTA500 excess
returnseriesused in Section4.2 is also veryhigh for both GQARCH(0 984) and GARCH
parameterizations(0 984), and the same is true of the conditionalvariances(0.992 and
0 996 respectively).This may not be that surprising,though, if we note from (12) that if
the one factor model is realistic,a diversifiedlinearcombinationof r, such as the FTA500
index will have almost no idiosyncraticcomponent(see also Sentanaand Shah (1994)).
In fact, if the marketreturnis well diversified,the asset pricingrelationabove coincides
with that of a conditionalCAPM with constantmarketbetas. In this respect,we obtain
very similarparameterestimatesfor c and F when we regressr, on the FTA500 return.
Our results seem to provide empiricalsupportfor the QARCH formulationof the
conditional varianceas a potential candidate to generalizeGARCH models for time-
varyingvariancesat the multivariatelevel. In addition, they confirmthat QARCH also
providesa more adequaterepresentationof risk premia.

7. CONCLUSIONS
In this papera generalquadraticmodel for the conditionalvarianceof a time serieswhich
ensurespositivityis introduced.It turnsout that this model is the most generalquadratic
version possible of the class of AutoregressiveConditionallyHeteroskedastic(ARCH)
modelsintroducedby Engle (1982)and for that reasonwe havecalledit QuadraticARCH,
or QARCH. Its main distinctivefeatureis that it allows an asymmetriceffectof positive
19. Hence,we allow for possibleco-persistencein varianceas definedin Bollerslevand Engle (1993).
20. As far as the conditionalvarianceis concerned,similarresultsare obtainedif we assumea constant
conditionalmean.
21. For scaling purposes,estimationof the factor model was carriedout imposingthe restrictionO=
I - a,, - 3, so that the unconditionalvarianceof the unobservablefactoris 1.
 SENTANA QUADRATIC ARCH MODELS 657

and negative lagged values of the series. To allow for infinitedependenceon the past,
laggedvalues of the conditionalvariancehave been includedas in Bollerslev(1986).
In turn, the QARCH model may be nested into two non-parametricapproachesto
dynamicconditionalheteroskedasticity.First,a quadraticpolynomialconstitutesthe lead-
ing term in Gallant's(1981) FlexibleFourierForm approach,whereextra trigonometric
terms are added to the conditionalvariancefunction (see Pagan and Hong (1991) and
Pagan and Schwert(1990)). Second, a piecewise-quadratic spline approximationto the
unknownconditionalvariancefunctionwouldalso encompassQARCHas a trivialsmooth
example,as well as the modelsof Glosten,Jaganathanand Runkle(1993), Schwert(1989),
Taylor(1986), and Zakoian(1990). Hence,QARCHmay also providea usefulbenchmark
to comparethe relativeperformanceof thesemodels.Ourapproachis also similarin spirit
to the semi-non-parametric estimationtechniquebased on a polynomialexpansionput
forwardin Gallant and Tauchen(1989) (see also Gallant,Rossi and Tauchen(1992)).
The time-seriespropertiesof (G)QARCH processes are very similar to those of
Bollerslev's(1986) GARCH. In particular,GQARCH(1, 1) and GARCH(1, 1) models
are remarkablyclose: they both have the same mean, variance and autocorrelation
functionsfor both the seriesand its squares,as well as the sameforecastingrecursionrule.
Nevertheless,the GQARCH(1, 1) has the advantagethat by adding a single parameter,
it can allow for both an asymmetriceffect on the conditional variance and higher
(unconditional)kurtosis,which goes in the rightdirectiontowardscapturingsome of the
stylizedfacts characterizingmany financialtime series.
An additionaladvantageof quadraticARCH modelsis that they areeasy to integrate
in the analysis of structuraleconomic models, just as linearmodels for the conditional
mean are. The asymmetricmodel of changing volatility in stock returns of Campbell
and Hentschel (1992) provides an illustrativeexample. In their model, a multiplicative
parametrizationwould have been analyticallyintractable,while the QARCHmodel made
their derivationsstraightforward.
Multivariateextensionsof the QARCHmodel are straightforward in theory,but as in
multivariateARCH models,difficultto estimategiventhe numberof parametersinvolved.
However, QARCH formulationsare the only ones that can be easily adapted to the
context of conditionallyheteroskedasticlatent factor models, and do not entail a large
computationalburden.
Two univariateempiricalapplicationsto daily U.S. and monthlyU.K. stock market
returnsprovide support for the fact that the GQARCH conditional variancefunction
representsthe data substantiallybetterthan a standardGARCH model. The main reason
is that it is able to capturethe so-called leverageeffect, which the other is ruling out a
priori. An applicationof a factormodel to 26 U.K. industrialsectorsalso shows empirical
supportfor the GQARCHformulationvs. the GARCH one. In addition,all threeapplica-
tions show that it also providesa betterrepresentationof risk premia.
There is obviously no compellingtheoreticalreason why the conditional variance
function should be literallyquadratic,just as there is no reason why conditionalmeans
shouldbe linear.If anything,one could expectnon-linearitiesas well as "non-quadratities"
to be the rule, ratherthan the exception,and not surprisingly,the literatureon non-linear
conditionalmean and variancemodels is growingfast. At the same time, semi-parametric
and non-parametricmethodsare becomingever more popular.Henceno generalityclaim
should be made about our approach.But equally,it seems sensibleto use the quadratic
conditionalvariancefunctiondiscussedhere as a benchmark,just as one would initially
use linearmodelsfor conditionalmeansfor variablesthatcan takeon negativeand positive
values.
658 REVIEWOF ECONOMICSTUDIES

APPENDIX
COVARIANCE STATIONARITYOF GQARCH MODELS
The GQARCH(p,q) processin (5) is equivalentup to conditionalsecondmomentsto the followingrandom
coefficientsmodel:
X, = + Ej= I I?tx
x ?,qs+j, ,,+j I jtat _
tX_i + vEi'= Al
(Al)

where(?,, IIIt,. tqt, .It, . . p=

, (e, , ), are i.i.d. randomcoefficientswith zeromeanand covariance
matrix
L p//2 01
v/2 A 01,
0 0 Al

providedthat A and A are positivesemi-definite,with A diagonaland 0- y,'A'ip1/4>0. This equivalenceis not

unexpected,since the GQARCHmodel encompassesthe GAARCH model of Beraand Lee (1990), which is
introducedas a randomcoefficientsmodel,and the originalARCH modelof Engle(1982),whichwas expressed
as a randomcoefficientsmodel by Tsay (1987). The randomcoefficientsinterpretationalso suggestspotential
generalizationsof the GQARCHmodel. For example,one could allow for non-zerocovariancesbetween ,
and/or ?, with 4,, as well as non-zerocovariancesbetweendifferent{,,'s. For instance,the asymmetricnonlinear
GARCH model in Engle and Ng (1993) can be written as ? + tj I X, - I + ,Itat - I, with cov (ij ,, ,,) # 0. Similarly,
a model with termsof the form oa,-ja,, could be achievedwith a non-diagonalA. Neitherexample,though,
resultsin a QARCH(oo)model.
To provecovariancestationarity,it is actuallymore convenientto workwith the followingalternativere-
parameterization:
x, =(, + iq_, ti,(x, -i-bi) + E,j=, {j,tf,-j (A2)
wherethe bi'sare constantparametersand (4,, ,, ,) are i.i.d. randomcoefficientswith zero meanand covari-
ance matrix
cO 0 1
0 LL' Oj,
[ 0 A

with (,, iq, and 4, mutually independent.

Let s=max (p,q), and let z'=(x,, x, I,.., Xt-s+,), v,=(a,, a,,. a,-,+), and u'=(,,0,. ,0) be
s x I vectors.Also, definethe following[I + (s - 1)]x s matrices:

[I ?] [? ? ] [? O ]; 0 -
s-I I q s-q p s-P I s-I

so that the processcan be representedin state-spaceformas:

Zt= ((4 + Tt)zt-- I + ltvt- I +ut- T,b. (A3)
Let V,= E(z,z,). Then, recallingthe mutualand serialindependenceof i7,, {, and 4,, it follows that:
V,= ,4'D)+ E('P,z, , z, ,P) + E(Ht,v, v,' n,H)+ (c + b'LL'b)G. (A4)
E(z-z, -z,
Apart fromthe inclusionof b'LL'bin the constantterm,this expressionis the sameas the one derivedby
Bera and Lee (1990) for the GAARCH model (i.e. when b=0), and hence the stationaritycondition for
GQARCHmodels,statedin the followingproposition,is the sameas the stationarityconditionfor GAARCH
models (see Beraand Lee (1990) for the proof):
Let

Ls0=[ o ]O] and As = [0 ]S"P

q s-q p s-p

be lowertriangularand positivesemi-definitediagonals x s matricesrespectively.

 SENTANA QUADRATIC ARCH MODELS 659

Then x, as generatedby equation(A2) is covariancestationaryif and only if all the eigenvaluesof the
s x s matrixR are less than 1 in absolutevalue,where:
R=(4d?4 )+vec (G) vec'(LsL'+ A,). (A5)
Providedthat this conditionis satisfied,the unconditionalvarianceof z,, V, will be:
vec (V) = (c +b'LL'b)[(Is(?s)- RI-' vec (G). (A6)
Given the shape of G, the unconditionalvarianceof x, will be given by (c +b'LL'b)multipliedby the
element1, 1 of [(Is?s) - R5-', which is easily seen to be the reciprocalof the determinantof [(Il?ls) -RI-'.
Tediousalgebrashows that this determinantis equal to the familiarexpressionI - Eqa,aii- =, A>,wherea11
is the i-th diagonalelementof the matrixA = LL'.Thereforethe unconditionalvarianceof x, is indeedgivenby
equation(7).

DATA APPENDIX
The followinglist refersto the definitionof the 26 FinancialTimesActuariesSectorIndicesas of December31,
1990,and includesthe DATASTREAMfour-lettersectormnemonics(all startingwith FTA).
BANK Banks
BDIS Brewersand Distillers
BMAT BuildingMaterials
CHEM Chemicals
CONC Contracting,construction
ELEC Electricals
ENGG EngineeringGeneral
FDMG Food Manufacturing
FDRT Food Retailing
INBR Insurance(Brokers)
INCM Insurance(Composite)
INLF Insurance(life)
INVT InvestmentTrusts
LEIS Leisure
MERB MerchantBanks
METL Metalsand Metal Forming
MISC Miscellaneous
MISF OtherFinancial
MTRS Motors
NWSP Publishingand Printing
OILS Oil and Gas
PAPA Packagingand Paper
PROP Property
SHPT Shippingand Transport
STOR Stores
TEXT Textiles
Of these sectors,seven (Banks,Life Insurance,InsuranceGeneral,InsuranceBrokers,MerchantBanks,
Propertyand InvestmentTrusts)are not in the FTA 500 shareindex (FTA500I).The beginningof the month
3-monthTbill rate was used as the safe interestrate (UKTRSBL%).

Acknowledgements.This is an extensivelyrevisedversionof Sentana(1991). Threeanonymousreferees

have helpedme greatlyimprovethe paper.The author is also gratefulto ManuelArellano,Anil Bera,John
Campbell,RussellDavidson,AntonisDemos, FrankDiebold,SteveDurlauf,Rob Engle,ChristianGourieroux,
LudgerHentschel,Theo Nijman,Peter Robinson,MushtaqShah, SushilWadhwaniand seminarparticipants
at CEMFI, LSE, UniversidadComplutense,Universityof Southampton,the ESRC EconometricStudyGroup
Conference(Bristol,July 1991)and the EuropeanMeetingof the EconometricSociety(Cambridge,September
1991)for veryusefulcomments.Of course,the usualcaveatapplies.ErnstMaugcollectedthe U.K. sectordata.
Financialsupportfromthe LSE FinancialMarketsGroup,the SpanishMinistryof Educationand Scienceand
the ESRC is also gratefullyacknowledged.
660 REVIEW OF ECONOMIC STUDIES

REFERENCES
ATTANASIO,0. and EDEY, M. (1987), "Non-constantVarianceand ForeignExchangeRisks:An Empirical
Study"(LSE CLE DiscussionPaper285).
BAILLIE,R. T., BOLLERSLEV, T. and MIKKELSEN,H. 0. (1993), "FractionallyIntegratedAutoregressive
(KelloggGraduateSchoolof ManagementFinanceWorkingPaper168).
ConditionalHeteroskedasticity"
BERA, A. K., HIGGINS, M. L. and LEE, S. (1992), "InteractionBetweenAutocorrelationand Conditional
Heteroskedasticity: A Random Coefficients Approach", Journal of Business and Economic Statistics, 10,
133-142.
BERA, A. K. and LEE, S. (1990), "On the Formulationof a GeneralStructurefor ConditionalHeteroskedas-
ticity"(mimeo,Universityof Illinois).
BERA,A. K. and LEE, S. (1993), "InformationMatrixTest, ParameterHeterogeneityand ARCH: A Synthe-
sis", Review of Economic Studies, 60, 229-240.
from the AmericanStatistical
BLACK,F. (1976), "Studiesof Stock MarketVolatilityChanges",Proceedings
Association, Business and Economic Statistics Section, 177-181.
Journalof Econo-
BOLLERSLEV,T. (1986), "GeneralisedAutoregressiveConditionalHeteroskedasticity",
metrics, 31, 307-327.
BOLLERSLEV,T. (1990), "Modellingthe Coherencein Short-RunNominalExchangeRates:A Multivariate
Generalized ARCH Approach", Review of Economic and Statistics, 72, 498-505.
BOLLERSLEV,T., CHOU, R. Y. and KRONER, K. F. (1992), "ARCHModelingin Finance:A Reviewof
the Theory and Empirical Evidence", Journal of Econometrics, 52, 5-59.
BOLLERSLEV,T. and ENGLE,R. F. (1993), "CommonPersistencein ConditionalVariances",Econometrica,
61, 166-187.
BOLLERSLEV,T., ENGLE,R. F. and NELSON, D. B. (1994), "ARCHModels",in R. F. Engleand D. L.
McFadden (eds.), Handbook of Econometrics, Vol. 4 (Amsterdam: Elsevier).
BOLLERSLEV,T., ENGLE, R. F. and WOOLDRIDGE,J. M. (1988), "A CapitalAsset PricingModel with
Time Varying Covariances", Journal of Political Economy, 95, 116-131.
BOLLERSLEV,T. and WOOLDRIDGE,J. M. (1992), "Quasi-Maximum LikelihoodEstimationand Inference
in DynamicModelswith Time VaryingVariances",EconometricReviews,11, 143-172.
BOUGEROL,P. and PICARD,N. (1992), "Stationarityof GARCH Processesand Some Non-NegativeTime
Series", Journal of Econometrics, 52, 115-128.
CAMPBELL,J. and HENTSCHEL,L. (1992), "No News is Good News: An AsymmetricModel of Changing
Volatility in Stock Returns", Journal of Financial Economics, 31, 281-318.
DEMOS, A. and SENTANA, E. (1991), "InequalityConstraintson GQARCH Models" (mimeo, London
School of Economics).
DEMOS, A. and SENTANA, E. (1994), "Testingfor GARCH Effects:A One-sidedApproach"(mimeo,
CEMFI).
DIEBOLD, F. X. and NERLOVE,M. (1989), "The Dynamicsof ExchangeRate Volatility:A Multivariate
Latent Factor ARCH Model", Journal of Applied Econometrics, 4, 1-21.
DROST,F. C. (1993), "TemporalAggregationof TimeSeries",in J. Kaehlerand P. Kugler(eds.), Econometric
Analysis of Financial Markets (Heidelberg: Physica Verlag).
DROST, F. C. and NIJMAN, T. E. (1993), "TemporalAggregationof GARCHProcesses",Econometrica, 61,
909-927.
ENGLE,R. F. (1982),"Autoregressive ConditionalHeteroskedasticitywith Estimatesof the Varianceof United
KingdomInflation",Economnetrica, 50, 987-1007.
ENGLE, R. F. (1987), "MultivariateARCH with Factor Structures-Cointegrationin Variance"(mimeo,
Universityof California,San Diego).
ENGLE, R. F. (1990), "Discussion:Stock MarketVolatilityand the Crashof 87", Reviev of FinancialStudies,
3, 103-106.
ENGLE, R. F. and BOLLERSLEV,T. (1986), "Modellingthe Persistenceof ConditionalVariances",Econo-
metric Reviews, 5, 1-50, 81-87.
ENGLE, R. F. and KRONER, K. F. (1995), "MultivariateSimultaneousGeneralizedARCH", Econometric
Thleory,11, 122-150.
ENGLE, R. F. and NG, V. K. (1993), "Measuringand Testingthe Impactof News on Volatility",Journal of
Finance,48, 1749-1778.
GALLANT,A. R. (1981), "Onthe Bias in FlexibleFunctionalFormsand an EssentiallyUnbiasedForm:The
Fourier Flexible Form", Journal of Econometrics, 15, 211-244.
GALLANT,A. R., ROSSI,P. E. and TAUCHEN,G. (1992), "StockPricesand Volume",Reviewof Financial
Studies, 5, 199-242.
GALLANT,A. R. and TAUCHEN,G. (1989), "Seminonparametric Estimationof ConditionallyConstrained
57, 1091-1120.
HeterogeneousProcesses:Asset PricingApplications",Econometrica,
GILL, P. E., MURRAY, W. and WRIGHT,M. H. (1981) PracticalOptimization (London:AcademicPress).
GLOSTEN,L. R., JAGANNATHAN,R. and RUNKLE, D. (1993), "Onthe Relationbetweenthe Expected
Valueand the Volatilityof Nominal ExcessReturnson Stocks",Journalof Finance,48, 1779-1801.
GOLDBERGER,A. S. (1991) A Coursein Econometrics(Cambridge,MA.: HarvardUniversityPress).
HANSEN, L. P. and SARGENT,T. J. (1991) RationalExpectationsEconometrics(Boulder,CL.: Westview
Press).
 SENTANA QUADRATIC ARCH MODELS 661
HARVEY,A. C., RUIZ, E. and SENTANA, E. (1991), "UnobservableComponentTime SeriesModelswith
ARCH Disturbances", Journal of Econometrics, 52, 129-158.
HENTSCHEL,L. (1994), "All in the Family:Nesting Symmetricand AsymmetricGARCH Models",Journal
of Financial Economics (forthcoming).
on Acoustics,
HINICH, M. (1990), "Detectinga TransientSignalby BispectralAnalysis",IEEE Transactions
Speech and Signal Processing, 38, 1277-1283.
KING, M. A., SENTANA,E. and WADHWANI,S. B. (1994), "Volatilityand LinksbetweenNationalStock
Markets",Econometrica, 62, 901-933.
Journalof Econometrics,17,
KOENKER,R. (1981), "A Note on Studentizinga Test for Heteroskedasticity",
107-112.
KRAFT, D. and ENGLE, R. F. (1983), "ARCH in MultipleTime-seriesModels" (mimeo, Universityof
Californiaat San Diego).
LEE,S. W. and HANSEN,B. E. (1994),"AsymptoticTheoryforthe GARCH(l, 1) Quasi-Maximum Likelihood
Estimator", Econometric Theory, 10, 29-52.
LUMSDAINE, R. L. (1993), "Consistencyand AsymptoticNormality of the Quasi-MaximumLikelihood
Estimatorin GARCH(l, 1) and IGARCH(l, 1) Models"(mimeo,PrincetonUniversity).
NELSON, D. B. (1990), "Stationarityand Persistencein the GARCH(l, 1) Model",EconometricTheory,6,
318-334.
NELSON, D. B. (1991), "ConditionalHeteroskedasticityin Asset Returns:A New Approach",Econometrica,
59, 347-370.
NELSON, D. B. (1992), "Filteringand Forecastingwith MisspecifiedARCH Models: I. Getting the Right
Variancewith the WrongModel",Journalof Econometrics, 52, 61-90.
NELSON, D. B. and CAO, C. Q. (1992), "InequalityConstraintsin the UnivariateGARCH Model",Journal
of Business and Economic Statistics, 10, 229-235.
NELSON, D. B. and FOSTER,D. P. (1994), "AsymptoticFilteringTheoryfor UnivariateARCH Models",
Econometrica, 62, 1-41.
NIJMAN, T. and SENTANA,E. (1994), "Marginalization
and Contemporaneous Aggregationin Multivariate
GARCH Processes"(CEMFIWorkingPaper9419, forthcomingin Journalof Econometrics).
PAGAN, A. R. and HONG, Y. (1991), "Non-ParametricEstimationand the Risk Premium",in W. Barnett,
J. Powell and G. Tauchen (eds.), Semiparametricand Nonparamnetric
Methods in Econometricsand Statistics
(Cambridge:CambridgeUniversityPress).
PAGAN, A. R. and SCHWERT,G. W. (1990), "AlternativeModelsfor ConditionalStockVolatility",Journal
of Econometrics, 45, 267-290.
ROBINSON,P. M. (1991), "Testingfor StrongSerialCorrelationand DynamicConditionalHeteroskedasticity
in Multiple Regression", Journal of Econometrics, 47, 67-84.
ROSS, S. A. (1976), "The ArbitrageTheoryof CapitalAsset Pricing",Journalof EconomicTheory,13, 341-
360.
SCHWALLIE,D. P. (1982), "UnconstrainedMaximumLikelihoodEstimationof ContemporaneousCovari-
ances", Economic Letters, 9, 359-364.
SCHWERT,G. W. (1989), "Whydoes Stock MarketVolatilityChangeover Time?",Journal of Finance,44,
1115-1153.
SENTANA, E. (1991), "QuadraticARCH Models: A PotentialRe-interpretation of ARCH Models" (LSE
FinancialMarketsGroup DiscussionPaper 122).
SENTANA,E. and SHAH,M. (1994),"An Indexof Comovementsin FinancialTimeSeries"(CEMFIWorking
Paper9415).
TAYLOR,S. (1986) ModellingFinancialTimeSeries(New York: John Wileyand Sons).
TSAY, R. S. (1987), "ConditionalHeteroskedasticTime SeriesModels",Journal of the AmericanStatistical
Association, 82, 590-604.
WEISS,A. A. (1986), "AsymptoticTheoryfor ARCH Models: Estimationand Testing",EconometricTheory,
2, 107-131.
CovarianceMatrixEstimatorand a DirectTest for Hetero-
WHITE,H. (1980), "A Heteroskedastic-Consistent
skedasticity",Econometrica, 48, 817-838.
WOOLDRIDGE,J. M. (1990), "A UnifiedApproachto Robust,Regression-based SpecificationTests",Econo-
metric Theory, 6, 17-43.
WOOLDRIDGE,J. M. (1991), "On the Applicationof Robust, Regression-basedDiagnosticsto Models of
47, 5-46.
ConditionalMeansand ConditionalVariances",Journal of Econometrics,
WOOLDRIDGE,J. M. (1992), "SomeAlternativesto the Box-Cox RegressionModel",InternationalEconomic
Review, 33, 935-955.
YANCEY,T. A., JUDGE, G. G. and BOCK,M. E. (1981), "TestingMultipleEqualityand InequalityHypoth-
esis in Economics",EconomicLetters,7, 249-255.
ZAKOIAN,J. M. (1990), "ThresholdHeteroskedasticModels"(mimeoCREST,INSEE).