HAL Id: hal-00517094
https://hal.archives-ouvertes.fr/hal-00517094
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.
THE ERRORS ARE UNCORRELATED BUT NONINDEPENDENT
Hamdi Raïssi
To cite this version:
Hamdi Raïssi. TESTING THE COINTEGRATING RANK WHEN THE ERRORS ARE UNCORRE-
LATED BUT NONINDEPENDENT. Stochastic Analysis and Applications, Taylor & Francis: STM,
Behavioural Science and Public Health Titles, 2009, 27, pp.24-50. �hal-00517094�
HAMDIRAÏSSI,
∗
Université Lille3
Abstrat
We study the asymptoti behaviour of the redued rank estimator of the
ointegratingspaeandadjustmentspaeforvetorerrororretiontimeseries
modelswithnonindependentinnovations. Itis shownthatthedistributionof
the adjustment spae an be quite dierent for models with iid innovations
and models with nonindependent innovations. It is also shown that the
likelihoodratiotestremainsvalidwhentheassumptionofiidGaussianerrors
is relaxed. MonteCarlo experimentsillustratethe nitesampleperformane
ofthelikelihoodratiotestusingvariouskindsofweakerrorproesses.
Keywords:Cointegration,reduedrankregression,likelihoodratiotest,strong
mixingondition,vetorerrororretionmodel.
1. Introdution
Multivariateproessesareoftenusedineonometriappliationsbeausetheyallow
tounderstandtheinterationsbetweendierentvariables. Inordertodesribelongrun
eonomirelationships,theointegrationtheoryhasbeendevelopedbyGranger(1981),
Engle and Granger(1987), Ahnand Reinsel (1990). This theory postulatesthat, in
someases,astationary proess oflowerdimensionis obtainedbyonsidering linear
ombinationsoftheomponentsofamultivariatenonstationaryproess. Thenumber
ofindependentlinearombinationsistheointegratingrankandisanimportantpiee
ofinformationfortheanalysisofeonomidata.
The dominant test for the ointegrating rank is the likelihood ratio (LR) test
developed by Johansen (1988, 1991), Perron and Campbell (1993), Lütkepohl and
Saikkonen (1999) in the framework of vetor error orretion models (VECM). For
theointegrationanalysis,theerrorstermsaregenerallysupposed tobeindependent
and identially distributed (iid). When applied to eonomi data (see for instane
Johansen andJuselius (1990),Clementsand Hendry (1996)orTrenkler(2003)),this
iidassumptionseemstoorestritivebeausemaroeonomitimeseries oftenexhibit
onditionalheterosedastiityand/orother formsofnonlinearity.
Rahbek,HansenandDennis(2002)studied theeetof ARCH innovations onthe
LRtest. AnimportantoutputoftheirworkisthattheLRtestremainsvalidwhenthe
errorproessisamartingaledierene. Howevertheassumptionthattheerrorproess
isamartingaledierenepreludesotherformsofdependene. Indeedthereexistmany
exampleswheretheassumptionofiidormartingalediereneontheinnovationsisnot
satised(see for instane Franq, Roy and Zakoïan (2005) in the univariateARMA
aseor Franq and Raïssi(2005)in theVARase). The rstaim ofthis paper isto
studythevalidityoftheLRtest inageneralontextofunorrelatederrors.
∗
Postaladdress: UniversitéLille 3,EQUIPPE UniversitésdeLille BP60149 VilleneuveD'Asq
Cedex,Frane. E-mail:hamdi.raissietu.univ-lille3.fr.
The seond aim is to study the asymptoti behaviour of the usual estimators of
the ointegration and adjustment spaes, in the general framework of VECM with
unorrelated, but possibly dependent errors. We will ompare our ndings to the
usualiidaseandresultsofSeo (2007)whihshowsin partiularthattheasymptoti
distribution of the redued rank estimator of the ointegrating spae is robust to
onditional heteroskedastiity. We will use the standardredued rank proedure to
estimatetheointegrationspae,relaxingtheassumptionofiidgaussianinnovations.
Thestrutureofthepaperisasfollows. InSetion2wepresentthemodelandwe
derivetheestimatorsoftheparameters.InSetion3wegivetheasymptotibehaviour
oftheLRtest. Insetion4westatetheonsistenyoftheointegrationspaeandthe
adjustment spae. In Setion 5Monte Carloexperimentsare performed. Theproofs
arerelegatedtotheappendix.
Inthesequelthefollowingnotationsareused. Weak onvergene isdenoted by⇒
andwedenote by
→P theonvergeneinprobability. For afullolumn rankmatrixA
of dimensiond×r withd > r, wedene theorthogonal omplement A⊥, whih is a
fullolumnrankmatrixofdimensiond×(d−r)andsuhthatA′A⊥= 0. Thesymbol
⊗ denotes the usual Kroneker produt and ve(A) denotes the vetor obtained by
stakingtheolumnofthematrixA. Wedenotebytr(B)thetraeofasquarematrix B. Wedenoteby[m]theintegerpartofagivenrealm.
2. Charaterization ofthe model
WeonsiderthefollowingVECMwithlineartrend
∆Xt= Π0Xt−1+
p−1
X
i=1
Γ0i∆Xt−i+µo0+µo1t+ǫt (2.1)
where µo0 and µo1 are d-dimensional parametervetors. The proess (ǫt) is usually
assumediidwithmeanzeroand positivedenite ovarianematrixΣǫ. Inthesequel
wewillonsideraweakerassumptionfortheerrorproess. TheΓ0i,i∈ {1, ..., p−1},
ared×dshortrunparametersmatries. Byonventionthesumvanishesin(2.1)when p= 1. Thefollowingassumption givesusthegeneralframeworkofourstudy.
AssumptionA1 (Cointegrationandrestritiononthetrendparameters)
(a) The matrixΠ0 is ofrankr0 (0≤r0< d). Ifr0 >0 thenΠ0 an bewritten as
Π0=α0β0′ whereα0 andβ0 arefullolumnrankmatriesofdimensiond×r0.
(b) The autoregressivepolynomialA(z) = (1−z)Id−Π0z−Pp−1
i=1 Γ0i(1−z)zi, is
suhthat|A(z)|= 0 impliesthat |z|>1orz= 1.
() Thematrixα′0⊥Γ0β0⊥ isoffullrankd−r0,whereΓ0=Id−Pp−1 i=1Γ0i.
(d) The vetor µo1 is suh that µo1 = −α0τ0, where τ0 6= 0 is an r0-dimensional vetor.
Note that if r0 = 0 the relation (2.1) is a vetor autoregressive model for the proess(∆Xt). Condition(d)isthelessrestritiveonditionontheparametersofthe
deterministipartof(2.1)whihallowsfortrendingbehaviourfor(Xt). Indeedunder
representation
Xt=C
t
X
i=1
ǫi+ρo1t+ρo0+Yt+A, (2.2)
where C =β0⊥(α′0⊥Γ0β0⊥)−1α′0⊥. ThetermA dependson initialvaluesand is suh
thatβ0′A= 0. Thestationaryproess(Yt)isoftheform Yt=
∞
X
i=0
ϕ0iǫt−i,
whereC(z) =P∞
i=0ϕ0iziisonvergentfor|z|≤1 +δ,forsomeδ >0. Notethat(2.2)
impliesthat(Xt)isanI(1)proess. From(a)and(d)weanwrite (2.1)as
∆Xt=ν0+α0β0′∗Z1t+
p−1
X
i=1
Γ0i∆Xt−i+ǫt (2.3)
whereZ1t= (Xt−1′ ,−t+ 1)′ andβ0∗= (β0′, τ0)′. Thed-dimensionalvetorofonstants
ν0 and the r0-dimensional vetor τ0 are funtions of the parameters in (2.1). Note
that in (2.2) the vetorρo1 is suh that β′0ρo1 =τ0. Then it an beseen from (2.2)
that (β0′Xt−E(β0′Xt))is trendstationary and the r0-dimensional proess (β0′∗Z1t− E(β0′∗Z1t))isstationary. Wesayin thisasethattheointegratingrankisr0. Inthis
studywetest,forsomer(0≤r < d),thenullhypothesis H0:r0=r vs. H1:r0> r.
Notethatin(2.3)theparametersα0,β0andτ0arenotidentied. Indeedforagiven
α01, β01, and sine we assumed that these matries havefull rank, wean takeany
nonsingularmatrixζ ofdimensionr0×r0suhthat β02=β01ζ andα02=α01(ζ′)−1
willgivethesamematrixΠ0. Togetridofthisproblemoneanonsiderthefollowing
normalization
β0c∗ = (β0c′ , τ0c)′ = ((β0(c′β0)−1)′,(β0′c)−1τ0)′ and α0c =α0β0′c,
wherethedimensionald×r0matrixcissuhthatc′β0hasfullrank. Thisnormalization ensuresidentiabilityin thesense thatwehaveβ01c=β02c. Toseethis, notethat
c′β01c =c′β02c =Ir0 ⇒ c′β01(c′β01)−1=c′β01ζ(c′β01ζ)−1
⇒c′β01
(c′β01)−1−ζ(c′β01ζ)−1
= 0. (2.4)
Thensinec′β01isafullrankmatrix,thisimpliesthat
(c′β01)−1−ζ(c′β01ζ)−1= 0. (2.5)
Multiplying(2.5)byβ01 ontheleft,weobtainβ01c=β02c. Onetheparameterβ0c is
identied,itiseasytoseethatα0c andτ0c arealsoidentied. It shouldbealsonoted
that theointegrationspaeandtheadjustmentspae,that isthespaesspanned by
respetivelyβ0c andα0c, donotdependonthehoieofthematrixc.
In general the assumption that (ǫt) is iid gaussian may appear to be toostrong.
IndeeditisquestionabletoassumethatalinearombinationofXt−1, . . . , Xt−p isthe
bestpreditorofXt. Inadditionnotethat, fromapratialpointofview,theorderp
isoften identiedusing teststhat are onlybasedontheautoorrelationsof(ǫt). For
instaneletusonsiderthedailyexhangeratesofU.S.DollarstooneBritishPound
andofU.S.DollarstooneEurofromJanuary2,2001toApril12,2007. Thelengthof
theseriesisT = 1578. TheanalyzeddataareplottedinFigure7.10. Weadjustedthe
model (2.1) to theseries with r0 = 1 and p= 2 using thesoftware JMulTi. Figures
7.11-7.12display the autoorrelations and rossorrelationsof the residuals. Figures
7.13-7.14displaytheautoorrelationsand rossorrelationsofthesquaredomponent
oftheresiduals. InviewofFigures7.11-7.12thehypothesisofunorrelatederrorsseems
plausible. Indeed most of the autoorrelations and rossorrelations are inside the
5%signianelimits. Howeversinemanyautoorrelationsandrossorrelationsare outsidethe5%signiane limitsin Figures 7.13-7.14,thehypothesisof independent errorsis learlyrejeted.
Rahbeket al (2002)onsidered VECMwith martingaledierene innovations. In
ourframeworkwewillonsideramoregeneralassumptionallowingforalargelassof
errorproesses.
Assumption A2 The error proess (ǫt) is stritly stationary and suh that Cov(ǫt, ǫt−h) = 0forallt∈Zandallh6= 0.
Suh errorproessesareommonly namedweakwhitenoise. Note that Granger's
representationtheorem stillholds whenthe assumptionofiid gaussianinnovationsis
replaed by A2. Thefollowingare examplesof error proesseswhih verifyA2 but
arenotiid.
Example2.1. Considertheproess(ǫt)dened bytherelation
ǫt=at+ Φ{ǫt−1⊙at}, (2.6)
where ⊙denotesthe Hadamardprodut, (at)is ad-dimensional iidentered proess suhthat |E(aitajt)|≤1, andthematrixΦis diagonalofdimensiond×dand suh
that|Φii |<1. TakingΦ0=Id,theequation(2.6)hasastationarysolutionoftheform
ǫt=P∞
i=0Φiat−i⊙ · · · ⊙at.Itis easyto seethattheǫt'sareunorrelated. However
Cov(ǫ2it, ǫ2it−1) =E(a2it)Cov((1 + Φiiǫit−1)2, ǫ2it−1)6= 0,
ingeneral,showingthattheproess(ǫt)isnotiid.
Example2.2. The univariate all-pass models (see for instane Breidt, Davis and
Trindade (2001)) onstitute an important lass whih an be extended to the mul-
tivariate ase. Assume that the proess (ǫt) is the unique solution to the following
equation
ǫt−φ01ǫt−1− · · · −φ0qǫt−q =wt+φ0q−1φ−10q wt−1+· · ·+φ01φ−10qwt−q+1−φ−10qwt−q,
where φ(z) = Id−φ01z· · · −φ0qzq is suh that φ(z)6= 0 for| z |≤ 1. The entered
proess (wt)is iidwith varianeΣw. Assume alsothat the matriesφ01, . . . , φ0q are
diagonal. Writingthespetraldensityforeahomponent(ǫit),it anbeshownthat
theproess(ǫt)isunorrelated(seeAndrews,DavisandBreidt(2006)). Howeverify0
isnotgaussiantheproess (ǫt)isnotindependent. Toseethis onsiderthefollowing bivariatesimpleexample
ǫt−φǫt−1=wt−φ−1wt−1
where φ =
φ1 0 0 φ2
and | φ1 |< 1, | φ2 |< 1. Let us introdue ϑt = ǫ1t− φ1ǫ1t−1. Sine (ǫt)is unorrelated, the proess (ϑt) followsan ausalM A(1). Then
we have ǫ1t = P
i≥0φiϑt−i. Straightforwardomputations show that E(ǫ1tϑ2t−1) = E[ǫ1t(ǫ1t−1−ǫ1t−2)2] =Ew3t(1−φ−21 )(1+φ1)andE(ǫ1tϑ3t−1) =E[ǫ1t(ǫ1t−1−ǫ1t−2)3] = (Ewt4−3)(1−φ−21 )2φ1. Usingthefat that ϑt−1 belongsto theσ-eld generated by {ǫ1u, u < t},wehaveE{ϑ2t−1E(ǫ1t|ǫ1t−1,· · ·)} 6= 0forEwt36= 0andE{ϑ3t−1E(ǫ1t| ǫ1t−1,· · ·)} 6= 0for Ew4t 6= 0. Thusthe (ǫt)proessis notamartingaledierene in
general.
2.1. Derivation of thequasi maximumlikelihood(QML) estimators
Now we turn to the derivation of the QML estimators of α0c and β0c∗. We use
here the QML method beause we assume that the errors terms are unorrelated
but not neessary gaussian independent. Note that the estimation proedure we
will desribe is performed under H0. In the framework of the VECM we shall see
that the methodology in Johansen (1988,1991) in the iid ase remains valid under
unorrelatederrorsassumption. Wewill usethe followingnotation. Let Z0t= ∆Xt,
Z2t = (∆Xt−1′ , . . . ,∆Xt−p+1′ ,1)′, Ψ0 = (Γ01, . . . ,Γ0p−1, ν0) where Xt = 0 for t ≤ 0.
Theexpression(2.3)beomeswiththesenotations
Z0t=α0cβ0c′∗Z1t+ Ψ0Z2t+ǫt. (2.7)
HereweanremarkthatsineXtisI(1)thentheproessesZ0tandZ2tarestationary.
Using(2.7)andgiventheobservationsX1, . . . , XT wewritethequasilog-likelihoodas follows
logL(Ψ, αc, βc,Σǫ) =−1
2Tlog|Σǫ|
−1 2tr
( T X
t=1
Σ−1ǫ (Z0t−αcβc′∗Z1t−ΨZ2t)(Z0t−αcβ′∗c Z1t−ΨZ2t)′ )
,
where
βc∗= (β′c, τc)′= ((β(c′β)−1)′,(β′c)−1τ)′ and αc=αβ′c.
Themaximum likelihood estimationmethod fortheVECM with unorrelatederrors
impliatesseveralsteps. WerstestimatetheparametersinthematrixΨ0andobtain
Ψ(αˆ c, β∗c) =M02M22−1−αcβ′∗c M12M22−1
where
Mij =T−1
T
X
t=1
ZitZjt′ .
Now dening by R0t and R1t the residualsof respetivelytheregressionsof Z0t and
Z1tonZ2t,wegettheonentratedlog-likelihood
logL(αc, β∗c,Σǫ) = −1
2Tlog|Σǫ|
−1 2tr
( T X
t=1
Σ−1ǫ (R0t−αcβc′∗R1t)(R0t−αcβc′∗R1t)′ )
(2.8)
where
R0t=Z0t−M02M22−1Z2t and R1t=Z1t−M12M22−1Z2t.
Sine the R1t's are the residuals of the regressionof the Z1t's on the Z2t's, and
noting that the proess (Z1t) is I(1) and the proess (Z2t) is I(0), then theproess (R1t) is I(1). The expression of the onentrated log-likelihood orresponds to the regressionequation
R0t=α0cβ0c′∗R1t+ ˜ǫt, (2.9)
sothatweobtainthefollowingunfeasibleestimatorsofα0candΣǫin(2.9)byordinary
leastsquares
ˆ
αc(β∗0c) =S01β∗0c(β0c′∗S11β∗0c)−1, (2.10) Σˆǫ(β∗0c) =S00−αˆc(β0c∗)(β0c′∗S11β0c∗)ˆα′c(β∗0c)
where
Sij =T−1
T
X
t=1
RitRjt′ .
Notethatreplaingαc andΣǫ bytheirestimatesin(2.8)wewrite
logL(ˆα(βc∗), β∗c,Σˆǫ(β∗c)) =−1
2Tlog|Σˆǫ(βc∗)| −1 2dT.
Finallytheparametersin β0c∗ an beestimatedusingtheresultsofthewellknown
reduedrankmethodofAnderson(1951). Inthisendweshallminimizethefollowing
expression
|Σˆǫ(βc∗)|=|S00−S01βc∗(βc′∗S11β∗c)−1β′∗c S10|.
Usingtherelation
A11 A12
A21 A22 =|A11||A22−A21A−111A12|=|A22||A11−A12A−122A21|,
wend
|S00−S01βc∗(βc′∗S11β∗c)−1β′∗c S10|=|S00| |β′∗c (S11−S10S00−1S01)β∗c |
|βc′∗S11βc∗| .
UnderthenullhypothesisandusingLemma7.1theexpression|βc′∗(S11−S10S00−1S01)βc∗| /|βc′∗S11β∗c |isminimizedforthefollowingnormalizedexpression
βˆc∗= ( ˆβc′,τˆc)′= (( ˆβ(c′β)ˆ −1)′,(( ˆβ′c)−1τ))ˆ ′,
where
βˆ∗= ( ˆβ′,τ)ˆ ′ =S−
1 2
11 (v1, . . . , vr)
andv1, . . . , vr areeigenvetorsorrespondingto ther largestsolutionsλˆ1≥ · · · ≥λˆr
oftheeigenvalueproblem
|λI−S−
1 2
11 S10S00−1S01S−
1 2
11 |= 0. (2.11)
Inadditionthematrixc′βˆisoffullrank. Weobtainαˆc=S01βˆ0c∗( ˆβ0c′∗S11βˆ∗0c)−1. Noting
thatwehave|Σˆǫ( ˆβc∗)|=Qr
i=1(1−λˆi),thelikelihood ratiotestforris givenby Q−
2
rT = Qr
i=1(1−λˆi) Qd
i=1(1−λˆi)=
d
Y
i=r+1
(1−λˆi)−1.
Thentotestthenullhypothesis,weonsider theLR teststatisti
−2 logQr=−T
d
X
i=r+1
log(1−ˆλi),
where ˆλ1 ≥ · · · ≥λˆd arethed greatersolutionsof theeigenvalueproblem (2.11). In
thenextsetionwewillstudy theasymptotibehaviouroftheLR teststatisti.
3. Asymptotipropertiesof the LR statisti
To state the main results of the paper, the assumption that the proess (ǫt) is
unorrelated is not enough. Indeed we haveto ontrol the serial dependene of the
proess (ǫt). To this end we introdue the mixing oeients αξ(h) for a given
stationaryproess(ξt)
αξ(h) = sup
A∈σ(ξu,u≤t),B∈σ(ξu,u≥t+h)
|P(A∩B)−P(A)P(B)|,
whihmeasuresthetemporaldependeneoftheproess(ξt). Denekξtkq = (Ekξtkq)1/q,
wherek.kdenotestheEulideannorm. Thenweneedtomakethefollowingassumption
ontheproess(ǫt).
Assumption A3 The proess (ǫt) satises kǫtk2+ν+η < ∞ and the mixing oef-
ientsof theproess (ǫt)are suh that P∞
h=0{αǫ(h)}ν/(2+ν)<∞ for some ν >
0 and η >0.
Note that the kind of dependene induedby A3 is mild for the error proess (ǫt).
ThefollowingpropositiongivesustheasymptotidistributionoftheLRteststatisti.
Proposition 3.1. UnderA1,A2andA3,theLRteststatistihasthesameasymp-
totidistributionasin the iidgaussianase, that is
−2 logQr0 ⇒tr (Z 1
0
F(dB)′ ′Z 1
0
F F′du
−1Z 1 0
F(dB)′ )
, (3.1)