HAL Id: hal-00517078
https://hal.archives-ouvertes.fr/hal-00517078
Submitted on 13 Sep 2010
HAL
is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire
HAL, estdestinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
models with uncorrelated but nonindependent errors
Christian Francq, Hamdi Raïssi
To cite this version:
Christian Francq, Hamdi Raïssi. Multivariate Portmanteau test for Autoregressive models with un-
correlated but nonindependent errors. Journal of Time Series Analysis, Wiley-Blackwell, 2007, 28,
pp.454-470. �hal-00517078�
DENT ERRORS
CHRISTIANFRANCQ,
∗
GREMARSUniversitéLille3
HAMDIRAÏSSI,
∗∗
GREMARSUniversitéLille3
Abstrat
Inthispaperweonsiderestimationandtestoftformultipleautoregressive
timeseriesmodelswithnonindependentinnovations.Wederivetheasymptoti
distributionoftheresidualautoorrelations. It isshownthatthis asymptoti
distributionanbequitedierentformodelswithiidinnovationsandmodelsin
whihtheinnovationsexhibitonditionalheterosedastiity orother formsof
dependene. Consequently,theusual hi-squaredistributiondoesnotprovide
adequate approximation to the distributionof the Box-Piere goodness-of-t
portmanteau test in the presene of nonindependent innovations. We then
proposeamethodtoadjusttheritialvaluesoftheportmanteautests. Monte
Carlo experiments illustrate the nite sample performane of the modied
portmanteautest.
Keywords: VetorweakARmodel,Goodness-of-ttest,Residualautoorrela-
tion,Diagnosti Cheking,Box-PiereandLjung-Boxportmanteautests.
1. Introdution
In multivariate time series analysis,the Vetorial AutoRegressive (VAR) models are
muhemployed(seeLütkepohl(1993)). TheVARmodelspostulatethatthed-dimen-
sionalvetorXtoftheobservationsattimetanberepresentedasalinearombination of ppast valuesXt−1, . . . , Xt−p plus anerror ǫt. A theoretial argument in favor of
the VAR models is that any stationary proess an be approximated by a VAR(p)
model with suiently largepand unorrelatederrors. Thereason ofthe suessof these modelsishoweverofpratialnature,andisertainly dueto thefatthat itis
relativelyeasytodealwithalinearfuntionofanitenumberofpastvalues.
ItishoweverlearthattheVARmodelsarenotuniversalandthatthehoieofthe
order pis ruial. Thus it isimportantto hekthe validity of aVAR(p)model, for
agivenorderp. Inmultivariate,thehoieofpispartiularlyimportantbeausethe numberofparameters,pd2,quiklyinreaseswithp,whihentailsstatistialdiulties.
Thispaperisdevotedtotheso-alledportmanteautestsonsideredbyChitturi(1974)
andHosking(1980)forhekingtheoverallsignianeoftheresidualautoorrelations
of a VAR(p) model (see also Ahn (1988), Hosking (1981a, 1981b), Li and MLeod
(1981)).
∗
Postaladdress: GREMARS,UFRMSES,UniversitéLille3,Domaine duPontdebois,BP149,
59653Villeneuved'AsqCedex,Frane
∗∗
Postaladdress: GREMARS,UFRMSES,UniversitéLille3,DomaineduPontdebois,BP149,
59653Villeneuved'AsqCedex,Frane
ForthestatistialanalysisofVARmodels,theerrorsǫtaregenerallysupposedtobe
independent(as in Lütkepohl (1993),Denition 3.1). Thisindependene assumption
is restritive beause it preludes onditional heterosedastiity and/or other forms
ofnonlinearity(see Franq, Royand Zakoïan (2005)andFranqand Zakoïan (2005)
forthestatistialinfereneofunivariateARMAmodelswith unorrelatedbut nonin-
dependent errors,and see Dufour andPelletier (2005)forweak VARMA modelling).
Themaingoalofthepresentpaperisto studythebehaviouroftheabove-mentioned
goodness-of-tportmanteautestswhentheǫt'sarenotindependent. Wewillseethat theresultsobtainedbythestandardportmanteautestsanbequitemisleadinginthis
framework.A modiedversionofthese testsisproposed.
Brüggemann, Lütkepohl and Saikkonen (2004) is an important reent referene
dealing with portmanteau tests and other tests for residual autoorrelation in VAR
models with iid innovations when somevariables are ointegrated. Inthis paperwe
donotonsiderointegratedvariables,buttheindependeneassumptionoftheǫt'sis
relaxed.
Therest of thepaperis organizedas follows. Setion 2 provides examplesof AR
models withunorrelatedbut nonindependenterrors,presentsseveralexpressions for
theleastsquares (LS)estimatorofthe ARoeients,and givesonditionsensuring
theonsisteny andasymptoti normality oftheLS estimator. Setion 3studies the
asymptotibehaviourof the ARresiduals. Weobtainthe asymptotidistribution of
vetorsof residual autoorrelations, under the assumption of the tted AR(p) is an
adequatelinearmodel with awellhosenorder p. The resultsareapplied in Setion
4to obtainthe asymptoti distribution of the portmanteau tests and to modify the
ritialvaluesofthesetestswhentheyareappliedtoVARmodelswithnonindependent
errors. Setion5isdevotedtothepratialimplementationofthemodiedversionof
thetests. Setion6proposesnumerialillustrations. Thetehnialproofsarerelegated
totheappendix.
Thefollowing notationswill be used throughout. For amatrix A of generi term A(i, j)weuse thenormkAk=P
|A(i, j)|. Thespetralradiusof asquarematrixA
is denoted by ρ(A), its traeis denoted by Tr. We denote by A⊗B the Kroneker
produt oftwomatriesA andB, veA denotesthevetorobtainedbystakingthe
olumnsofA,andA⊗2standsforA⊗A(seee.g. Harville(1997)formoredetailsabout
thesematrixoperators). Thesymbol⇒denotestheonvergeneindistribution.
2. LSestimator of weak VARmodels
ConsiderthevetorialAR(p)model Xt=
Xp
i=1
A0iXt−i+ǫt forallt∈Z={0,±1,±2, . . .} (2.1)
where theǫt'sare d-dimensionalerror terms,theXt'sared-dimensional vetors,and theA0i'sared×dmatries. Itisustomarytosaythat (Xt)isastrongAR(p)model
if(ǫt)isastrongwhitenoise,thatis, ifitsatises
TheauthorsaregratefultotheProfessorLütkepohlwhosequestionsheaskedduringtheMadrid
A1: (ǫt) isa sequeneof independent and identially distributed (iid)random vetors,Eǫt= 0andVar(ǫt) = Σǫ.
Wesaythat(2.1)isaweak AR(p)model if(ǫt)isaweakwhitenoise,thatis, ifit
satises
A1': Eǫt= 0,Var(ǫt) = Σǫ,and Cov(ǫt, ǫt−h) = 0forallt∈Zandallh6= 0.
Assumption A1 is learly stronger than A1'. The lass of strong AR models is
oftenonsidered toorestritivebypratitioners. IndeedAssumption A1amountsto
assume that the best preditor of Xt is a linearombinationof its p pastvalues. If pis hosenlargeenough,it is reasonableto onsider that thebest preditor ofXt is
wellapproximatedbyafuntion ofXt−1, . . . , Xt−p,butitisquestionabletoassumea linearformforthisfuntion. Itiswellknownthatnumerousnonlinearproessesadmit
weaklinearrepresentations(seeExample2.1below). Weaklinearrepresentationsare
alsoobtainedfromtransformationsofstronglinearproesses(seeExample2.2below).
In Examples 2.1-2.3 below, Assumption A1' holds but A1 is not satised. Other
examplesofunivariateweak linearmodels an befound in Franq, Roy andZakoïan
(2005),andreferenestherein.
For the statistial analysis of multivariate AR time series models, it is therefore
of interest to replae the standard strong white noise assumption A1 by the more
exible weak white noiseassumption A1'. Inthis paperwe fous on theestimation
andvalidationstagesofthestatistialanalysisoftheseweakmultivariateARmodels.
2.1. Examples of weak VARmodels
Theexamplesgivenin thissetionaremainly hosenfortheirsimpliity. Therst
oneisaweakwhitenoiseinspiredbyexamplesgivenbyRomanoandThombs(1996)
intheunivariatease. Theseondissimplytheausalrepresentationofanonausal
AR(1) proess. This example shows that A1 must by replaed by A1' when one
wantsto make,withoutlossofgenerality, theusualassumption that therootsof the
ARpolynomialareoutsidetheunit irle. Thethird belongstothelass ofGARCH
models.
Example2.1. Intheunivariatease,Romanoand Thombs(1996)builtweak white
noises(ǫt) bysetting ǫt =ηtηt−1· · ·ηt−k where (ηt) is astrongwhite noiseand k ≥ 1. This approah an be extended to the d-multivariate framework by setting ǫt = B(ηt)· · ·B(ηt−k+1)ηt−k,where{ηt= (η1t, . . . , ηd t)′}tisad-dimensionalstrongwhite noise, and B(ηt) = {Bij(ηt)} is a d×d random matrix whose elements are linear
ombinationsof theomponentsof ηt. It isobvioustohekthat ǫtis awhitenoise,
butingeneralthisnoiseisnotastrongone. Indeed,assumingforsimpliitythatk= 1
andB11(ηt) =η1t andB1j(ηt) = 0forallj >1,wehave
Cov ǫ21t, ǫ21t−1
=
Eη21t 2Var
η21t 6= 0,
whihshowsthattheǫt'sarenotindependent.
Example2.2. (NonausalAR(1))LettheAR(1) model
Xt=AXt−1+ǫt, ǫt iidand E(ǫt) = 0, E(ǫtǫ′t) = Σǫ,
where A is an invertiblematrix whose all the eigenvalues λi,1 ≤i ≤d, are greater
thanone in modulus. Thisequation hasastationary andantiipativesolutionof the
formXt=−P∞
i=1A−iǫt+i. Theautoovarianefuntion of(Xt)isthengivenby ΓX(h) =A−h
X∞
i=1
A−iΣǫA−i, h≥0.
Let ǫ∗t = Xt−A−1Xt−1. We have E(ǫ∗t) = 0, Var(ǫ∗t) = ΓX(0) + ΓX(1)(A−1)′ + A−1ΓX(1) +A−1ΓX(1)(A−1)′, andCov(ǫ∗t, ǫ∗t−h) = 0 forh6= 0. Thus Xt admits the
ausal AR(1) representation Xt = A−1Xt−1+ǫ∗t. However, in general, ǫ∗t is not a
martingale dierene. To see this, assume for simpliity that the matrix A is suh
that |a11| >1 and a1j =aj1 = 0, ∀j ∈ {2,· · · , d}. Wethen have EX13t−1 = (1− a311)−1Eǫ31t, EX1tX12t−1=a−211(1−a311)−1Eǫ31tandE(ǫ∗1tX12t−1) =EX1tX12t−1− a−111EX13t−16= 0whenEǫ31t6= 0. Inthisaseǫ∗t isnotamartingaledierenebeause E(ǫ∗1tX12t−1) =E
X12t−1E ǫ∗1t|ǫ∗t−1, . . . 6= 0. Thusthewhitenoiseǫ∗t isnotstrong.
Example2.3. Intheunivariatease,theGARCHmodelsonstituteimportantexam-
plesofweakwhitenoises. Thesemodelshavenumerousextensionstothemultivariate
framework. The simplest of these extensions is ertainly the multivariate GARCH
model with onstant orrelation proposed by Jeantheau (1998). In this model the
proess(ǫt)veriesthefollowingrelationǫt= ∆tηtwhereηt= (η1t, . . . , ηd t)′ isaniid
entered proesswith Var(ηi t) = 1, and∆t isadiagonal matrixwhose elementsσii t
verify
σ211t
.
.
.
σdd t2
=
c1
.
.
.
cd
+ Xq
i=1
Ai
ǫ21t−i
.
.
.
ǫ2d t−i
+ Xp
j=1
Bj
σ211t−j
.
.
.
σdd t−j2
.
TheelementsofthematriesAiandBj,aswellastheci's,aresupposedtobepositive.
In addition suppose that the stationarity onditions hold (see Jeantheau (1998) for
more details). The stationary solution of this GARCH equation satises A1′, but doesnotsatisfyA1ingeneral. ConsiderforinstanetheARCH(1)asewithA1 suh
that a11 6= 0and a12 =· · · =a1n = 0. Thenit iseasy to seethat Cov(ǫ21t, ǫ21t−1) =
Cov
(c1+a11ǫ21t−1)η12t, ǫ21t−1 =a11Varǫ21t6= 0,whihshowsthattheiidassumption A1is notsatised.
2.2. Derivation of theLS estimator
Itiswellknownthat(Xt)an bewrittenas aMA(∞)oftheform Xt=
X∞
i=0
ψ0iǫt−i, ψ00=Id, X∞
i=0
kψ0ik<∞, (2.2)
undertheassumption
A2: detA0(z)6= 0forall|z| ≤1,where A0(z) =Id−Pp
i=1A0izi.
Denote by θ0 = ve (A01· · ·A0p) the vetor of the unknown AR parameters. For any θ ∈ Rd2p, let A1 = A1(θ), . . . , Ap = Ap(θ) be d×d matries suh that θ =
ve(A1, . . . , Ap). With thisnotation(2.1)an berewrittenas Xt=
Xt−1′ , . . . , Xt−p′
⊗Id θ0+ǫt (2.3)
usingtheelementaryrelationve(ABC) = (C′⊗A)veB formatriesofappropriate dimensions.
One of the most popular estimation proedure is that of the least squares (LS)
estimator. For linear proesses of the form (2.1), the LS estimator of θ oinides
withthegaussianquasi-maximumlikelihood(QML)estimator. Giventheobservations
X1, . . . , Xn,theLS/QMLestimatorsofθandΣǫ aredenedby
(ˆθn,Σˆǫ) = arg min
θ,Σǫ
(
nlog (det Σǫ) + Xn
t=1
ǫ′t(θ)Σ−1ǫ ǫt(θ) )
where
ǫt(θ) =Xt− Xp
i=1
AiXt−i.
To give an expliit expression for these estimators, the d-dimensional AR(p) model
(2.1)anberewrittenasthedp-dimensionalAR(1)model
X˜t= ˜A0X˜t−1+ ˜ǫt, (2.4)
where
A˜0=
A01 · · · A0p−1 A0p
Id 0
.
.
.
0 Id 0
, X˜t=
Xt
Xt−1
.
.
.
Xt−p+1
, ˜ǫt=
ǫt
0
.
.
.
0
.
NotethatA2isequivalenttoρ A˜0
<1. LetΣˆX˜t,X˜t−h= n1Pn
t=1X˜tX˜t−h′ andwrite ΣˆX˜t−1
instead of ΣˆX˜t−1,X˜t−1
. Note that ΣˆX˜t−1
is a onsistent estimator of ΣX˜t = EX˜tX˜t′,whihisgivenby
ve ΣX˜t
=
I(dp)2−A˜0⊗A˜0
−1
ve (Σ˜ǫt), Σǫ˜t=
Σǫ 0′d(p−1)
0d(p−1) 0d(p−1)×d(p−1)
.
It iseasy tosee that theLSestimatorsof theAR parametersofmodels(2.1)and
(2.4)aregivenby
b˜ A=
Ab1 · · · Abp−1 Abp
Id 0
.
.
.
0 Id 0
= ˆΣX˜t,X˜t−1
Σˆ−1˜
Xt−1, (2.5)
provided ΣˆX˜t−1
is non singular, and that the LS estimators of the varianes of the
noises(˜ǫt)and(ǫt)aregivenby Σbǫ˜= Σbǫ 0′d(p−1)
0d(p−1) 0d(p−1)×d(p−1)
!
= ˆΣX˜t−ΣˆX˜t,X˜t−1Σˆ−1˜
Xt−1
ΣˆX˜t−1,X˜t, (2.6)
withtheonventionthat Xt= 0whent≤0 ort > n.
2.3. Asymptoti behaviorofthe LS estimator
Toestablishthestrongonsistenyoftheestimatorsdenedin (2.5)and(2.6),we
needthefollowingassumptions.
A3: MatrixΣǫ ispositivedenite.
A4: Thesequene(ǫt)isstritlystationary andergodi.
Note that A4 isentailedby A1, but notby A1′. A straightforwardonsequeneof theergoditheoremisstatedinthefollowingproposition.
Proposition 2.1. UnderassumptionsA1-A2-A3orA1'-A2-A3-A4,thematrixΣˆX˜t
isalmost surelynonsingular, andalmost surely
b˜
A→A˜0, Σb˜ǫ→Σ˜ǫ and θˆn→θ0
asn→ ∞.
ToobtaintheasymptotinormalityoftheLSestimatorofθ,additionalassumptionsare neededwhen(ǫt)isnotiid. Intheunivariatease,FranqandZakoïan(1998)showed
the asymptoti normality under mixing assumptions. We will extend this result to
VARmodels. ThemixingoeientsofastationaryproessX = (Xt)aredenotedby αX(h) = sup
A∈σ(Xu,u≤t),B∈σ(Xu,u≥t+h)|P(A∩B)−P(A)P(B)|.
Thereader isreferredto Davidson(1994)for details aboutmixing assumptions. Let
kXkr= (EkXkr)1/r,where kXkdenotestheEulideannorm.
A5: TheproessX= (Xt)issuhthatP∞
h=0{αX(h)}ν/(2+ν)<∞andkXtk4+2ν
<∞forsomeν >0.
Theasymptotidistribution oftheLSestimatorisgiveninthefollowingproposition.
Proposition 2.2. Under assumptionsA1A3 orunderA1' andA2A5,
√nve b˜ A−A˜0
⇒ N(0,Ω), (2.7)
where
Ω = X∞
h=−∞
En Σ−1X˜
t
X˜t−1X˜t−h−1′ Σ−1X˜
t ⊗˜ǫt˜ǫ′t−ho
. (2.8)
Moreover
√n θˆn−θ0
⇒ N 0,Σθˆn
, (2.9)
where
Σθˆn = X∞
h=−∞
En Σ−1X˜
t
X˜t−1X˜t−h−1′ Σ−1X˜
t ⊗ǫtǫ′t−ho
. (2.10)
Notethatunder A1,wehave
Ω = Σ−1˜
Xt ⊗Σ˜ǫt.
and
Σθˆn = Σ−1X˜
t ⊗Σǫt. (2.11)
ThisstandardresultanbefoundinJohansen(1995,Theorem2.3p. 19).
Example2.4. The following example illustrates the dierene between the asymp-
totivariane(2.10)intheweakaseandtheasymptotivariane(2.11)ofthestrong
AR ase. Consider a bivariate AR(1) model Xt = AXt−1 +ǫt, with true AR(1)
parameterA0= 0andǫt=B(ηt). . . B(ηt−k+1)ηt−k oftheformgiveninExample2.1.
WehaveXt= ˜Xt=ǫt. For simpliity assumethat B(ηt) = Diag(η1t, η2t)and that η is gaussianwith Var(η1t) =Var(η2t) = 1 and Cov(η1t, η2t) =ρ ∈(−1,1). Thus,
using Eη13tη2t = 3ρ and Eη12tη22t = 1 + 2ρ2, we ndthat Var(ǫ1t) = Var(ǫ2t) = 1,
Cov(ǫ1t, ǫ2t) =ρk+1,and,using(2.10), Σθˆn = E
Σ−1ǫt ǫt−1ǫ′t−1Σ−1ǫt ⊗ǫtǫ′t
= Σ−1ǫt ⊗I2
E
(ǫt−1⊗ǫt) (ǫt−1⊗ǫt)′ Σ−1ǫt ⊗I2
,
withΣǫ=
1 ρk+1 ρk+1 1
andE
(ǫt−1⊗ǫt) (ǫt−1⊗ǫt)′ equalto
3k ρ(3ρ)k ρ(3ρ)k ρ2(1 + 2ρ2)k ρ(3ρ)k (1 + 2ρ2)k ρ2(1 + 2ρ2)k ρ(3ρ)k ρ(3ρ)k ρ2(1 + 2ρ2)k (1 + 2ρ2)k ρ(3ρ)k ρ2(1 + 2ρ2)k ρ(3ρ)k ρ(3ρ)k 3k
.
Whenk= 1thematrixΣθˆn
isequalto
ΣW := 1
(ρ2+ 1)(1−ρ4)×
−2ρ4+ 3ρ2+ 3 ρ2(ρ2+ 3) −3ρ4−ρ2 ρ2(−2ρ4−3ρ2+ 1) ρ2(ρ2+ 3) 3ρ2+ 1 ρ2(−2ρ4−3ρ2+ 1) −ρ2(3ρ2+ 1)
−ρ2(3ρ2+ 1) ρ2(−2ρ4−3ρ2+ 1) 3ρ2+ 1 ρ2(ρ2+ 3) ρ2(−2ρ4−3ρ2+ 1) −3ρ4−ρ2 ρ2(ρ2+ 3) −2ρ4+ 3ρ2+ 3
.
Byomparison, forthestrongAR(1)withaniidnoisewithvariane
Σǫ=
1 ρ2 ρ2 1
,
theasymptotivarianeΣθˆn
isgivenby
ΣS := 1 1−ρ4
1 ρ2 −ρ2 −ρ4 ρ2 1 −ρ4 −ρ2
−ρ2 −ρ4 1 ρ2
−ρ4 −ρ2 ρ2 1
.
Figure 2.1displays theratioΣW(1,1)/ΣS(1,1) asfuntion ofρ. Itis learfrom this
examplethattheasymptotivarianeoftheLSestimatormaybequitedierentwhen