RESEARCH OUTPUTS / RÉSULTATS DE RECHERCHE
Author(s) - Auteur(s) :
Publication date - Date de publication :
Permanent link - Permalien :
Rights / License - Licence de droit d’auteur :
Bibliothèque Universitaire Moretus Plantin
Institutional Repository - Research Portal
Dépôt Institutionnel - Portail de la Recherche
researchportal.unamur.be
University of Namur
Hypothesis testing for tail dependence parameters on the boundary of the parameter
space
Kiriliouk, Anna
Published in:
Econometrics and Statistics
DOI:
10.1016/j.ecosta.2019.06.001
Publication date:
2020
Document Version
Publisher's PDF, also known as Version of record
Link to publication
Citation for pulished version (HARVARD):
Kiriliouk, A 2020, 'Hypothesis testing for tail dependence parameters on the boundary of the parameter space',
Econometrics and Statistics, vol. 16, pp. 121-135. https://doi.org/10.1016/j.ecosta.2019.06.001
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal ? Take down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
ContentslistsavailableatScienceDirect
Econometrics
and
Statistics
journalhomepage:www.elsevier.com/locate/ecosta
Hypothesis
testing
for
tail
dependence
parameters
on
the
boundary
of
the
parameter
space
Anna
Kiriliouk
Université de Namur, Faculté des sciences économiques, sociales et de gestion, Rempart de la vierge 8, Namur B-50 0 0, Belgium
a
r
t
i
c
l
e
i
n
f
o
Article history:
Received 25 January 2018 Revised 19 June 2019 Accepted 24 June 2019 Available online 29 July 2019
Keywords:
Brown–Resnick model Hypothesis testing Max-linear model Multivariate extremes Stable tail dependence function Tail dependence
a
b
s
t
r
a
c
t
Modellingmultivariatetaildependenceisoneofthekeychallengesinextreme-value the-ory. Multivariate extremes are usually characterized using parametric models, some of whichhavesimplersubmodelsattheboundaryoftheirparameterspace.Hypothesistests areproposedfortaildependenceparametersthat,underthenullhypothesis,areonthe boundaryofthealternativehypothesis.Theasymptoticdistributionoftheweightedleast squaresestimatorisgivenwhenthetrueparametervectorisontheboundaryofthe pa-rameterspace,andtwoteststatisticsareproposed.Theperformanceoftheseteststatistics isevaluatedfortheBrown–Resnickmodelandthemax-linearmodel.Inparticular, simu-lationsshowthatitispossibletorecovertheoptimalnumberoffactorsforamax-linear model.Finally,themethodsareappliedtocharacterizethedependencestructureoftwo majorstockmarketindices,theDAXandtheCAC40.
© 2019EcoStaEconometricsandStatistics.PublishedbyElsevierB.V.Allrightsreserved.
1. Introduction
Extreme-valuetheory isthebranchofstatistics concerned withthecharacterization ofextremeevents.Theseoccur in alargevarietyoffields,suchashydrology,meteorology,financeandinsurance,butalsoinsettingslikehumanlife spanor athleticrecords.Forsomeexamples,see EinmahlandMagnus(2008),Towleretal.(2010),Chavez-Demoulinetal.(2016),
Thomas etal.(2016) andRootzén andZholud(2017).In theunivariate case, the limitingdistribution ofsuitably normal-izedblock maximaorthresholdexceedancescanbecharacterizedentirelyusingthegeneralizedextreme-value(Fisherand Tippett,1928;Gnedenko,1943)andthegeneralizedParetodistribution(BalkemaandDeHaan,1974;Pickandsetal.,1975) respectively. However, manyextremeevents are inherentlymultivariate, and an importantchallenge is to modelthe tail dependencebetweentwoormorerandomvariablesofinterest.Ifdependencedisappearsasthevariablestakeonmoreand moreextremevalues,wesaythattheyareasymptoticallyindependent.Testingforasymptoticindependenceandmodelling asymptotically independentdatahasbeen doneinLedford andTawn (1996),Draismaetal.(2004),HüslerandLi(2009),
Wadsworthetal.(2016) andGuillouetal.(2018),amongothers.Inthefollowing,weconsidertheframeworkofasymptotic dependence.
Thefamilyoflimitingdistributionsformultivariatemaximaorthresholdexceedancesisinfinite-dimensional,sothat re-alisticandcomputationallyfeasibleparametricmodelsneedtobeproposed.ExamplesincludeGumbel(1960),Smith(1990),
Tawn(1990) andKabluchkoetal.(2009).Manydifferentapproachesofestimatingtaildependenceparametersexist.Fitting amultivariate extreme-valuedistribution tocomponentwise block maximaisdone in Padoanetal.(2010),Davisonetal. (2012),Castruccio etal.(2016) andDombryetal.(2017), alongothers.Multivariate thresholdexceedances, whichare the
E-mail address: anna.kiriliouk@unamur.be https://doi.org/10.1016/j.ecosta.2019.06.001
focus of thiswork, can be tackled using the stabletail dependence function (Drees andHuang, 1998) or the multivari-ategeneralizedParetodistribution(Rootzén andTajvidi,2006).Commonestimationmethodsincludemaximumlikelihood (WadsworthandTawn, 2014;Kiriliouket al., 2018a;de Fondevilleand Davison,2018) andminimum distanceestimation (Einmahletal.,2012;2016;2018).
Somepopularparametricmodelshavesimplersubmodelsattheboundaryoftheparameterspace.Themaingoalofthis paperisto proposehypothesis testsfortail dependenceparametersthat, underthenullhypothesis,are ontheboundary ofthealternative hypothesis.A firstexampleisthe Brown–Resnickmodel(Kabluchkoetal., 2009), popularinspatial ex-tremes,whichreducestothesimplerSmithmodel(Smith,1990)whentheshapeparameteroftheformerattainsitsupper bound. A second example is the max-linear model,where each component ofa vector can be interpreted as the maxi-mumshockamongasetofindependentfactors.Recentworkonmax-linearmodelsisnumerous.InGissiblandKlüppelberg (2018) andGissibletal.(2018),a recursivemax-linearmodelona directedacyclic graphisconsidered.InCui andZhang (2018) andZhaoandZhang(2018),max-linearmodelsareusedtomodelalatentfactorstructureforfinancialreturns. Fi-nally,theMarshall–Olkinmodel(Embrechtsetal.,2003;Segers,2012)isasubmodelthatisfrequentlyusedinpractice,see forinstanceBurtschelletal.,(2009),SuandFurman,(2017) orBrigoetal.,(2018).
Becauseoftheirnon-differentiability,max-linearmodelscannotbeestimatedbystandardlikelihoodmethodsandhence we usetheweightedleastsquaresestimatorofEinmahletal.(2018),whichiscentredaround thestabletaildependence function.Thehypothesistestingframeworkdescribedinthatpaperisvalidforparametersintheinterioroftheparameter space, and henceit cannot be usedto test whether one ormore factorloadings are equalto zero. The hypothesis tests proposed inthispapercanbeappliedtodecidehowmanyfactorsarenecessarytomodelacertain datasetortotest ifa morespecificmodel(eg,Marshall–Olkin)issufficient,bypassingthelimitationthatthenumberoffactorshastobechosen upfront.
Thepaperisorganizedasfollows.Section2 presentsbackgroundonmultivariateextremesandintroducestheweighted leastsquaresestimator.Section3 establishesasymptoticnormalityfortheweightedleastsquaresestimatorwhenthetrue parameter vector lies onthe boundary of theparameter space.Using results fromAndrews (1999),Andrews(2001) and
Andrews(2002),a deviance-anda Wald-type test statisticare proposed,whose asymptoticdistributions are easily com-putablewhenthedimensionoftheparametervectorontheboundaryismoderate.InSection4.2,weperformsimulations toevaluatetheperformanceoftheteststatisticsfortheBrown–Resnickandthemax-linearmodels.Finally,Section5 illus-tratesourmethodsontwomajorstockmarketindices,theDAXandtheCAC40.Technicaldefinitionsandproofsaredeferred totheappendix.
2. Background
2.1. Multivariateextreme-valuetheory
LetXi=
(
Xi1,...,Xid)
,i∈{
1,...,n}
,be randomvectorsinRd withcumulative distributionfunctionFandmarginalcu-mulative distribution functionsF1,...,Fd. LetMn:=
(
Mn,1,...,Mnd)
withMn j:=max(
X1j,...,Xn j)
for j=1,...,d.We saythatFisinthemax-domainofattractionofanextreme-valuedistributionGifthereexistsequencesofnormalizingconstants
an=
(
an1,...,and)
>0andbn=(
bn1,...,bnd)
∈Rd suchthatP
Mn− bn an ≤ x =Fn(
a nx+bn)
d −→G(
x)
, asn→∞. (1)Themargins,G1,...,Gd,ofGareunivariateextreme-valuedistributions,
Gj
(
xj)
=exp − 1+γ
xj−μ
jσ
j −1/γj + ,σ
j>0,γ
j,μ
j∈R,wherex+:=max
(
x,0)
.ThefunctionGisdeterminedbyG
(
x)
=exp{−
(
− logG1(
x1)
,. . .,− logGd(
xd))
}
,where:[0,∞)d→[0,∞)iscalledthe(stable)taildependencefunction,
(
x)
:=limt↓0t
−1P[1− F
1
(
X11)
≤ tx1 or ... or1− Fd(
X1d)
≤ txd]. (2)ThecumulativedistributionfunctionFisinthemax-domainofattractionofad-variateextremevaluedistributionGifand onlyifthelimitin(2) existsandthemarginaldistributionsin(1) convergetounivariateextreme-valuedistributions.Inwhat follows,weonlyassumeexistenceof(2),whichconcernsthedependencestructureofF,butnotthemarginaldistributions F1,...,Fd.
Becausetheclassofstabletaildependencefunctionsisinfinite-dimensional,oneusuallyconsidersparametricmodelsfor .Henceforthweassumethatbelongstoaparametricfamily{(· ;
θ
):θ
∈}with
⊂ Rp.Some examplescanbefound
below,seealsodeHaanandFerreira(2006),Falketal.(2010),Segers(2012) andreferencestherein. Example 2.1. Thed-dimensionallogisticmodel(Gumbel,1960)hasstabletaildependencefunction
If
θ
=1,thevariablesare(asymptotically)independent,whileθ↓
0correspondstothesituationofcompletedependence. Example 2.2. Thed-dimensionalBrown–Resnickmodeldefinedonspatiallocationss1,...,sd∈R2hasstabletaildependencefunction
(
x;θ
)
= d j=1 xjd−1
(
η
(j)(
1/x)
;ϒ
(j))
, whereη
(j)(
x)
=(
η
(j) 1(
x1,xj)
,...,η
(j−1j)(
xj−1,xj)
,η
(j+1j)(
xj+1,xj)
,...,η
d(j)(
xd,xj))
∈Rd−1,η
(j) l(
xl,xj)
=γ
(
sj− sl)
2 + log(
xl/xj)
2γ
(
sj− sl)
∈R,γ
(
s)
=(
||
s||
/ρ
)
α andϒ
(l)∈R(d−1)×(d−1)is the correlation matrix with entriesϒ
(j) lk =γ
(
sj− sl)
+γ
(
sj− sk)
−γ
(
sl− sk)
2
γ
(
sj− sl)
γ
(
sj− sk)
, l,k=1,...,d; l,k= j,
(Kabluchko etal., 2009; HuserandDavison, 2013). The parametervector is
θ
=(
ρ
,α
)
∈(
0,∞)
×(
0,2].The Smith model (Smith,1989)isobtainedwhenα
=2.Example 2.3. Themax-linearmodelwithrfactorshasstabletaildependencefunction
(
x;θ
)
= r t=1 max j=1,...,dbjtxj, x∈[0,∞)
d, (3)wherethefactorloadingsbjtarenon-negativeconstantssuchthat rt=1bjt=1forevery j∈
{
1,...,d}
andallcolumnsums ofthed× rmatrixB:=(bjt)j,t arepositive(Einmahletal.,2012).SincetherowsofBsumuptoone,theparametermatrixhasonlyd×
(
r− 1)
free elements. RearrangingthecolumnsofB willnot changethe valueof thestabletaildependence function.Foridentificationpurposes,wedefinetheparametervectorθ
bystackingthecolumnsofBindecreasingorderof theirsums, leavingout thecolumnwiththe lowestsum. Anexampleof arandom vectorZ=(
Z1,...,Zd)
thathasstabletaildependencefunction(3) is
Zj= max
t=1,...,rbjtSt, for j∈
{
1,...,d}
,whereS1,...,SrareindependentunitFréchetvariables,P[Sj≤ x]=exp
(
−1/x)
forx>0and j∈{
1,...,d}
.Example 2.4. LetI1,...,Id denoterandom(possiblydependent)Bernoulli randomvariables withpj=P[Ij=1]∈
(
0,1]forj∈
{
1,...,d}
. Let, for J⊂{
1,...,d}
, p(
J)
=P[{
j=1,...,d:Ij=1}
=J], so that(
p(
J))
J⊂{1,...,d} is a probability distribution. ThemultivariateMarshall–Olkinmodel(Embrechtsetal.,2003;Segers,2012)hasstabletaildependencefunction(
x;θ
)
= ∅=J⊂{1,...,d} p(
J)
max j∈J xj pj .Inthismodel,anysubsetofcomponentsofthevectorisassignedashockthatinfluencesallcomponentsofthatsubset.An interpretationintermsofcreditriskmodellingcanbefoundinEmbrechtsetal.(2003).
TheMarshall–Olkinmodelisobtainedasaspecialcaseofthemax-linearmodel(3) bysettingbjt=
p
(
Jt)
/pj1
(
j∈Jt)
,whereJt∈P
(
{
1,...,d}
)
, Jt=∅,andP(
A)
denotesthepower setofA.The Marshall–Olkinmodeliswell-knownind=2,where B=
b 11 0 1− b11 0 b22 1− b22 =⎛
⎜
⎝
P[I1=1,I2=0] P[I1=1] 0 P[I1=1,I2=1] P[I1=1] 0 P[I1=0,I2=1] P[I2=1] P[I1=1,I2=1] P[I2=1]⎞
⎟
⎠
.2.2.Estimationofthestabletaildependencefunction
2.2.1. Nonparametricestimationofthestabletaildependencefunction
Letk=kn∈
(
0,n]besuchthatk→∞andk/n→0asn→∞.AstraightforwardnonparametricestimatorofisobtainedbyreplacingPandF1,. . .,Fdin(2) bythe(modified)empiricaldistributionfunctionsandreplacingtbyk/n,yielding
n,k
(
x)
:=1 k n i=1 1{
Ri1>n+1/2− kx1or ... orRid>n+1/2− kxd}
. Here,Ri j= tn=11Xt j≤ Xi jdenotestherankofXijamongX1j,...,Xn j.Thisestimator,theempiricaltaildependencefunction,
n+1/2 rather than nallows forbetter finite-sample properties, that is,for a lower mean squarederror (Einmahl etal., 2012).
Anothernonparametricestimatoristhebetataildependencefunction,whichwasveryrecentlyproposedinKiriliouketal. (2018b).Contrarytotheempiricaltaildependencefunction,thebetataildependencefunctionisasmoothestimatorwhich leadstosomeimprovementinitsfinite-samplebehavior.Itsnamestemsfromthefactthatitisbasedontheempiricalbeta copula(Segersetal.,2017).Wedefine
βn,k
(
x)
=n k 1− Cβn 1−kx n , where Cβn(
u)
= 1 n n i=1 d j=1 Fn,Ri j(
uj)
, andfor r∈{
1,...,n}
, Fn,r(
u)
= rs=1nsus
(
1− u)
n−s isthe cumulative distribution function ofa Beta(
r,n+1− r)
randomvariable.
Adrawbackofn,k orβn
,k mightbetheir possiblygrowingbias ask increases:Fougèresetal.(2015) showthat, under
suitableconditions,theestimatorn,ksatisfiestheasymptoticexpansion
n,k
(
x)
−(
x)
≈ k−1/2Z(
x)
+α
(
n/k)
M(
x)
,whereZisacontinuouscenteredGaussianprocess,
α
isafunctionsuchthatlimx→∞α
(
x)
=0andMisacontinuousfunc-tion.When√k
α
(
n/k)
tendstoanon-zeroconstant,anasymptoticbiasappears.InFougèresetal.(2015) andBeirlantetal. (2016),thisbiasisestimatedandsubtractedfromtheestimatorn,k(
x)
inordertoproposenewbias-correctedestimators.Finally, manyother nonparametric estimators existof thePickands dependence function, whichis therestriction of to the unit simplex. See,for instance,Capéraà et al.(1997),Zhang etal. (2008),Gudendorf andSegers (2012), Berghaus etal.(2013),Vettorietal.(2016) andMarconetal.(2017).Whilethesecan be transformedinto estimatorsofthe stable taildependencefunction,theyrelyonstrongerassumptions,i.e.,theyarebasedondatafromanextreme-valuedistribution, andwewillnotconsiderthemhere.
2.2.2. Semi-parametricestimationofthestabletaildependencefunction
Nonparametricestimationformsasteppingstoneforsemi-parametricestimation(Einmahletal.,2012;2016;2018).Here, we focusonthemethodproposed inEinmahletal.(2018).Letn,kdenoteanyinitialestimatorofbasedon X1,...,Xn.
Letc1,...,cq∈[0,∞
)
d,withcm=(
cm1,...,cmd)
form=1,...,q,beqpointsinwhichwewillevaluateandn,k.Considerfor
θ
∈theq× 1columnvectors Ln,k:=n,k
(
cm)
q m=1, L
(
θ
)
:=(
cm;θ
)
q m=1, Dn,k
(
θ
)
:=Ln,k− L(
θ
)
. (4)The pointsc1,...,cq need to be chosen insuch a waythat themap L:
→Rq isone-to-one, i.e.,
θ
is identifiablefrom(
c1;θ
)
,...,(
cq;θ
)
.Inparticular,wewillneedtoassumethatq≥ p.For
θ
∈,let
(
θ
)beasymmetric,positivedefiniteq× qmatrixanddefinefn,k
(
θ
)
:=DTn,k
(
θ
)
(
θ
)
Dn,k(
θ
)
. (5)Thecontinuousupdatingweightedleastsquaresestimatorfor
θ
0isdefinedasθ
n,k:=argmin
θ∈ fn,k
(
θ
)
=argθ∈minDn,k
(
θ
)
T(
θ
)
Dn,k(
θ
)
. (6)InSection 4,wewillstudytheperformance ofthisestimatorfor
(
θ
)
=Iq,theq× qidentity matrix.Expression(6) thensimplifiesto
θ
n,k=argmin θ∈ q m=1 n,k(
cm)
−(
cm;θ
)
2 .
Einmahletal.(2018) showtheconsistencyandasymptoticnormalityof
θ
n,kundertheassumptionthatthetrue param-etervectorθ
0isintheinteriorof.Inthenextsectionwegivetheasymptoticdistributionof
θ
n,kwithoutthisrestriction.3. Inference on tail dependence parameters on the boundary of the parameter space
TheasymptoticresultspresentedinSection3.1 buildonAndrews(1999),whoestablishedtheasymptoticdistributionof ageneralextremumestimatorwhenoneormoreparameterslieontheboundaryoftheparameterspace.Themethodology to obtaintheasymptoticdistribution of√k
(
θ
n,k−θ
0)
canbe summarizedasfollows:first, oneshowsthat asn→∞,thecriterion functionfn,k(
θ
) in(5) isequaltoaquadraticfunctionqn,k(
√k(
θ
−θ
0))
plus atermthat doesnot dependonθ
.Ifθ
n,k isconsistent,then its asymptoticdistributionisonly affectedby thepartof
closeto
θ
0;equivalently,we areonlyconcerned withthe shifted parameterspace
−
θ
0 nearthe origin.If−
θ
0 canbe approximated nearthe originby aconvex cone
, one can show that minimizing fn,k(
θ
) overθ
∈is asymptotically equivalentto minimizing qn,k(
λ
) overInSection3.2,weshowhowaclosed-formexpressioncanbeobtainedfortheasymptoticdistributionof√k
(
θ
n,k−θ
0)
.For
β
⊂θ
,weshowinSection3.3 howtotestH0:β
=β
∗againstH1:β
=β
∗when,underthenullhypothesis,β
∗isontheboundaryofthealternativehypothesis.
Wesetupsome notationfirst.LetBε(
θ
)denoteanopen ballcenteredatθ
withradiusε
andletCε(θ
)denotean open cubecenteredatθ
withsidesoflength2ε
.Letcl()denotetheclosureof
.Aset
⊂ Rpissaidtobelocallyequaltoa
set
⊂ Rpif
∩B
ε
(
0)
=∩Bε
(
0)
forsomeε
>0.Finally,aset⊂ Rpisaconeif
λ
∈impliesa
λ
∈foralla>0.
3.1. Estimation,consistencyandasymptoticnormality
When
θ
0isontheboundaryof,themapLin(4) isnotdefinedandthusnotdifferentiableonaneighbourhoodof
θ
0.Wewillneedthefollowingassumption.
(A1)
includesaset
+ suchthat
+−
θ
0 equalstheintersectionofaunionoforthantsandanopencube Cε(0)forsome
ε
>0.Moreover,∩ Bε1
(
θ
0)
⊂+forsome
ε
1>0.If
−
θ
0 happenstobelocallyequaltoaunionoforthants,wecansimplyset+=
∩Cε
(
θ
0)
.ThisisthecaseforthemodelsconsideredinSection4.Wewillassume existenceoftheso-calledleft/right(l/r)partialderivativeson
+;aformal definitionisgivenintheappendix.Theshapeof
+issuchthatthesecanalwaysbedefined.
WriteL˙:=
(
∂
/∂
θ
)
L(
θ
)
∈Rq×p forθ
∈+,whereL˙ denotesthematrixofl/rpartialderivatives.Let
λ
1(θ
)>0denotethesmallesteigenvalueof
(
θ
).Theorem 3.1 (Existence, uniqueness and consistency). Let c1,...,cq∈[0,∞
)
d be q≥ p points such that the map L:θ
→(
(
cm;θ
))
qm=1 is a homeomorphism. Letθ
0∈cl() and assume that(A1) holds, that each element of L(
θ
) has continuous l/rpartialderivativesofordertwoon
+,thatL˙
(
θ
0)
isoffullrank,that:
→Rq×qhascontinuousl/rpartialderivativeson
+
andthatinfθ∈
λ
1(θ
)>0.Finallyassume,form=1,...,q,n,k
(
cm)
p
−→
(
cm;θ
0)
, asn→∞. (7)Thenwithprobabilitytendingtoone,theminimizer
θ
n,kin(6)existsandisunique.Moreover,θ
n,k
p
−→
θ
0, asn→∞.WeomittheproofofthistheoremsinceitisdirectlyobtainedbyreplacingBε(
θ
0)by+ intheproofofEinmahletal. (2018,Theorem1).
Whenfn,kisnotdefinedonaneighbourhoodof
θ
0,butthereexistsaset+whichsatisfies(A1),aTaylorexpansionof
fn,k(
θ
)aroundfn,k(θ
0)holds(Andrews,1999,Theorem6).Foreachθ
∈+,wehave fn,k
(
θ
)
=fn,k(
θ
0)
+D fn,k(
θ
0)
T(
θ
−θ
0)
+(
θ
−θ
0)
TD2fn,k(
θ
0)(
θ
−θ
0)
/2+Rn,k(
θ
)
,whereDfn,kandD2f
n,karebasedonl/rpartialderivativesandRn,k(
θ
)istheremainderterm.Wesuppressdependenceof,
˙
LandDn,kon
θ
0foreaseofnotation.FromEinmahletal.(2018,ProofofTheorems3.2.1and3.2.2)weknowthat D fn,k(
θ
0)
=−2DnT,kL˙+op
(
1)
, D2fn,k(
θ
0)
=2L˙TL˙+op
(
1)
,sothequadraticexpansionaboveisequalto
fn,k
(
θ
)
=fn,k(
θ
0)
− 2DTn,kL˙
(
θ
−θ
0)
+(
θ
−θ
0)
TL˙TL˙
(
θ
−θ
0)
+Rn,k(
θ
)
. Define J:=L˙TL˙∈Rp×p, Y n,k:= kJ−1L˙T
D n,k∈Rp, and qn,k
(
λ
)
=(
λ
− Yn,k)
TJ(
λ
− Yn,k)
.Thenfn,k(
θ
)canbewrittenasfn,k
(
θ
)
=fn,k(
θ
0)
− k−1YTn,kJYn,k+k−1qn,kk
(
θ
−θ
0)
+Rn,k
(
θ
)
.Weseethatfn,k(
θ
)isequaltoaquadraticfunction plusatermthatdoesnot dependonθ
plusRn,k(θ
).InAndrews(2002,Lemma3),itisshownthatundertheassumptionsofTheorem3.1 and(A3)below,theremaindertermRn,k(
θ
)issufficientlysmallsothatminimizing thequadraticapproximationtofn,k(
θ
)isequivalenttominimizingqn,k√k(
θ
−θ
0)
.Now,assume that
Then inf θ∈qn,k
k(
θ
−θ
0)
= inf λ∈qn,k(
λ
)
+op(
1)
,forsomecone
.Finally,assumethat (A3)√kDn,k
(
θ
0)
d
−→D∼ Nq
(
0,(
θ
0))
asn→∞,forsomecovariancematrix(
θ
0).Assumptions(A3)and(7) holdforthenonparametricestimatorspresentedinSection2.2.1 undersomegeneralregularity conditions.Thematrix
canbeexpressedintermsofanditsfirst-orderpartialderivatives.Wehave
Yn,k d −→Y:=J−1L˙T
D, q n,k
(
λ
)
d −→q(
λ
)
:=(
λ
− Y)
TJ(
λ
− Y)
, forallλ
∈Rp.Theorem 3.2 (Asymptoticnormality). Ifalloftheaboveassumptionshold,
k
(
θ
n,k−θ
0)
d−→
λ
=argminλ∈ q
(
λ
)
, asn→∞.Here
λ
can beinterpretedastheprojectionofYonwithrespecttothe normyJ:=yTJy.Thistheoremisa special
caseofAndrews(1999,Theorem3).Intheappendix,weverifythatourassumptionsaresufficient.Notethatif
includes aneighbourhoodof
θ
0,then=RpandwefindthesamedistributionasinEinmahletal.(2018,Theorem2)since
λ
=Y.3.2. Simplifyingtheasymptoticdistribution
The goalofthissectionis tosimplifytheasymptoticdistribution of√k
(
θ
n,k−θ
0)
andtogive conditionsunderwhich(part of)
λ
has a closed-form expression. LetG:=L˙TD, i.e.,Y=J−1G. We start by partitioningthe vector
θ
0∈Rp intotwosubvectors,
β
0∈Rc andδ
0∈Rp−c forc∈
{
1,...,p}
,whereδ
0 consistsofall parametersthatare intheinterioroftheparameterspace.Wepartition
θ
n,k,Y,J,Gandλ
accordingly:θ
n,k=β
n,kδ
n,k , Y= Yβ Yδ , J= Jβ Jβδ Jδβ Jδ , G= Gβ Gδ ,λ
=λ
βλ
δ .LetIcdenotethec× cidentitymatrix.ForH:=
Ic:0∈Rc×p,define qβ
(
λ
β)
:=(
λ
β− Yβ)
THJ−1HT−1
(
λ
β− Yβ)
. Supposethat(A4)Thecone
ofassumption(A2)isequaltotheproductset
β× Rp−c,where
β⊂ Rcisacone. Corollary 3.3. Ifalloftheaboveassumptionshold,then
k(
β
n,k−β
0)
d −→λ
β:=argmin λβ∈β qβ(
λ
β)
, k(
δ
n,k−δ
0)
d −→Jδ−1Gδ− J−1 δ Jδβλ
β.When
β isdefined byequality and/or inequality constraints,a closed-formexpressionfor
λ
β canbe computed; we givesomeexamplesthatarerelevantforSection4.Foramoreformalsolution,seeAndrews(1999,Theorem5).Example 3.1 ( c=1). Suppose that
=[0,1]p. If
β
0=0, then
β=[0,∞
)
andλ
β=max(
Yβ,0)
. Ifβ
0=1, thenβ=
(
−∞,0]andλ
β=min(
Yβ,0)
.Example 3.2 ( c=2). Supposethat
=[0,1]p.If
β
0=
(
1,1)
,thenβ=
(
−∞,0]2.Letρ
:=HJ−1HT.Then
λ
β:=1Yβ1<0,Yβ2<0 Yβ+1Yβ1−ρ
21Yβ2<0,Yβ2≥ 0(
Yβ1−ρ
21Yβ2,0)
T +1Yβ1≥ 0,Yβ2−ρ
12Yβ1<0(
0,Yβ2−ρ
12Yβ1)
T.If
β
0=(
0,0)
,β
0=(
0,1)
orβ
0=(
1,0)
,similar expressionscan be obtainedbyreversing thefirst and/or thesecond3.3.Hypothesistestingfortaildependenceparametersontheboundaryoftheparameterspace
Weareinterestedinconstructinghypothesistestsforparametervaluesthat,underthenullhypothesis,areonthe bound-aryofthealternativehypothesis.Weproposetwotest statistics,whoseasymptoticdistributionfollowsfromtheresultsin
Andrews(2001).
Assumethattherearenonuisanceparametersontheboundary,i.e.,all componentsof
θ
0 thatlieontheboundaryarepartofthenullhypothesis.Weareinterestedintesting
H0:
β
=β
∗ vs H1:β
=β
∗,β
∗∈Rc.Let
0:=
{
θ
∈:
θ
=(
β
∗,δ
)
forsomeδ
∈Rp−c}
denotetherestrictedparameterspace.Assumethat(A5)Forall
θ
∈0,
isaproductsetwithrespectto(
β
,δ
)localtoθ
.Thatis,forallθ
∈0,
∩Bε
(
θ
)
=(
B×)
∩Bε(
θ
)
forsomeB⊂ Rc,⊂ Rp−cand
ε
>0.Define
θ
(n0,k):=argminθ∈0 fn,k
(
θ
)
.AdevianceteststatisticcanbedefinedasTn(,k1):=k
fn,k
(
θ
n(0,k))
− fn,k(
θ
n,k)
.
Corollary 3.4. Suppose
θ
0∈0 andallpreviousassumptionshold.Then Tn(,k1)−→d
λ
TβHJ−1HT−1
λ
β.Recall that G=L˙T
D andlet J denote its covariance matrix,i.e., G∼ N
p
(
0 ,J)
with J=L˙TL˙∈Rp×p. Note that
J=J if and only if
=
−1. If
β=Rc (no parameters on the boundary) andJ =J, then
Tn,k(1)−→d YT
β
HJ−1HT−1Yβ∼
χ
c2,where
χ
2c denotesachi-squaredrandomvariablewithcdegreesoffreedom.
AWald-typeteststatisticcanbebasedonthequadraticformin
β
n,k−β
∗.LetVn−1:=(
HJn−1HT)
−1∈Rc×cdenoteaweightmatrixwhereJn:=J
(
θ
n,k)
.DefineTn(,k2):=k
β
n,k−β
∗ TV−1 nβ
n,k−β
∗ .Corollary 3.5. Suppose
θ
0∈0 andalloftheaboveassumptionshold.Then Tn(,k2)−→d
λ
Tβ
HJ−1HT−1
λ
β.Corollary 3.4 is a special caseof Andrews(2001,Theorem 4c) andCorollary 3.5 is a special case ofAndrews (2001, Theorem6d);theirproofsaredirectandthusomitted.
4. Simulation studies
Weconduct simulation experimentswhere wesimulate 1000 samplesof sizen=5000froma parametric modeland weassessthequality ofthe weightedleastsquaresestimatorintermsofitsrootmeansquarederror(RMSE) forsettings whereoneormoreofthe parametersare ontheboundaryof theparameterspace.Wetake k∈
{
25,50,...,300}
andwe usetheempiricalandthebetataildependencefunction asinitial estimators.Next,we studythe performanceofthetwo teststatisticsintroduced inSection3.3 intermsofempiricallevelandpowerfork∈{25,50,75,100}, aswe observethat higherkleadstoasteadilygrowingbiasthatquicklydeterioratestheperformanceofthehypothesistest.Inallexperiments, wetake=Iq,theq× q identitymatrix,becauseusingan optimalweightmatrixhasa veryminoreffectonthe quality
ofestimationwhileseverelyslowingdowntheestimationprocedure.Moreover,inverting
maybehinderedbynumerical problemsinthecaseofamax-linearmodel(Einmahletal.,2018).
4.1. Brown–Resnickmodel
Wesimulatedatafroma Brown–Resnickmodelwith
α
=2andρ
=1onaregular 3× 2grid (d=6)andona regular 4× 4grid(d=16).AsinEinmahletal.(2018),weletcm∈{0,1}dsuchthatexactlytwocomponentsofcmareequaltoone(meaningthat weconsider apairwise estimator)andwefocusonpairs ofneighbouringlocationsonly,i.e., locationsthat areatmostadistanceof√2apart.Thisleadstoq=11andq=42respectively.
Fig.1 showstheRMSEoftheparameterestimatesbasedontheempiricaltaildependencefunction(solidlines)andthe
betataildependencefunction(dashedlines).Weseethat theempiricaltaildependencefunctionoutperformsthebetatail dependencefunctionfor
α
,whiletheoppositeistrueforρ
,althoughdifferencesareveryminor.Forsmallvaluesofk,the RMSEislowerford=16thanford=6.Table 1
Empirical level of T n,k(2) based on the empirical and the beta tail dependence function for a significance level of 0.05.
k = 25 k = 50 k = 75 k = 100
emp beta emp beta emp beta emp beta
d = 6 6.0 4.3 5.1 5.0 7.0 5.5 8.8 8.0 d = 16 6.7 10.3 8.3 10.2 9.8 12.4 13.2 16.0 50 100 150 200 250 300 0.00 0.10 0.20 d = 6 k RMSE of α 50 100 150 200 250 300 0.00 0.04 0.08 0.12 d = 6 k RMSE of ρ 50 100 150 200 250 300 0.00 0.10 0.20 d = 16 k RMSE of α 50 100 150 200 250 300 0.00 0.04 0.08 0.12 d = 16 k RMSE of ρ
Fig. 1. RMSE for parameter estimators of the Brown–Resnick model based on the empirical tail dependence function (solid lines) and the beta tail depen- dence function (dashed lines) for (α, ρ) = (2 , 1) ; d = 6 (left) and d = 16 (right).
1.5 1.6 1.7 1.8 1.9 2.0 02 0 4 0 60 80 100 d = 6, k = 25 α P o w e r in % 1.5 1.6 1.7 1.8 1.9 2.0 02 0 40 60 80 100 d = 6, k = 50 α P o w e r in % 1.5 1.6 1.7 1.8 1.9 2.0 02 0 4 0 60 80 100 d = 6, k = 75 α P o w e r in % 1.5 1.6 1.7 1.8 1.9 2.0 0 20 40 60 80 100 d = 6, k = 100 α P o w e r in % 1.5 1.6 1.7 1.8 1.9 2.0 02 0 4 0 60 80 100 d = 16, k = 25 α P o w e r in % 1.5 1.6 1.7 1.8 1.9 2.0 02 0 40 60 80 100 d = 16, k = 50 α P o w e r in % 1.5 1.6 1.7 1.8 1.9 2.0 02 0 4 0 60 80 100 d = 16, k = 75 α P o w e r in % 1.5 1.6 1.7 1.8 1.9 2.0 0 20 40 60 80 100 d = 16, k = 100 α P o w e r in %
Fig. 2. Empirical power of T n,k(2) based on the empirical tail dependence function (solid lines) and the beta tail dependence function (dashed lines) for
k ∈ {25, 50, 75, 100}; d = 6 (top) and d = 16 (bottom).
Next, we study the empirical level and power of the test H0:
α
=2. Assumtion (A4) holds with=
(
−∞,0]× R. Table 1 showsthe empirical level of the test statistic Tn,k(2) based on a significance level ofα
=0.05. We do not display Tn,k(1)asitsbehaviourisidenticaltothat ofTn,k(2).Surprisingly,theempiricaltaildependencefunctionispreferredford=16 whilethebetataildependencefunctionperformsbestford=6.Ingeneral,wecanconcludethatlowvaluesofkaretobe preferredandthatthetestperforms betterforlower dimension.Fig.2 showstheempiricalpowerofTn,k(2) asafunction ofα
∈[1.5,2).Thepowerincreasesbothwiththedimensionandwithk.Table 2
Empirical level of T n,k(1) and T n,k(2) for models 1–3 (M1–M3) based on the empirical and the beta tail dependence function for a significance level of 0.05.
Tn,k(1) k = 25 k = 50 k = 75 k = 100
emp beta emp beta emp beta emp beta
M1 2.9 2.1 6.1 4.2 5.9 5.3 8.0 6.4
M2 0.3 2.3 8.7 5.0 12.4 12.6 40.3 27.7
M3 3.3 7.9 7.9 9.7 6.5 9.8 9.8 10.0
Tn,k(2) k = 25 k = 50 k = 75 k = 100
emp beta emp beta emp beta emp beta
M1 3.0 2.9 6.2 5.0 6.1 6.1 8.5 7.1
M2 0.3 2.4 9.2 5.3 12.5 13.0 41.3 28.3
M3 5.4 9.9 10.3 10.2 8.4 10.7 10.7 11.1
4.2.Max-linearmodel
Theconstantsc1,. . .,cqneedtobechosensuchthatthemax-linearmodelisidentifiable.InEinmahletal.(2018) itwas
observedthattakingextremalcoefficientsforcm,i.e., cm∈{0,1}dwhichatleasttwonon-zeroelements,isnotenoughfor
identifiability.Instead,oneneedstochoosecm suchthatitsnon-zeroelementsareunequal.Forsome theoretical
consider-ations,see alsoEinmahletal.(2018,AppendixB). Simulationexperiments showedthatin practice,takingalarge gridof values(meaningthat q p) on [0,1]d willlead to goodestimators interms ofRMSE. Hence, inall following simulation
experiments,wechoosec1,...,cqsuchthat
cm∈
{
0,0.01,0.1,0.2,...,0.8,0.9,0.99,1}
2, m∈{
1,...,q}
.4.2.1. Testingthestructureofamax-linearmodel Weconsiderthreedifferentscenarios:
Model 1 Letr=2andd=3,sothat
=[0,1]3.Let
θ
0=
(
b11,b21,b31)
=(
1,0.7,0.2)
.WetestH0:b11=1.Assumption(A4)holdswith
=
(
−∞,0]× R2.Model 2 Letr=2andd=3,sothat
=[0,1]3.Let
θ
0=
(
b11,b21,b31)
=(
1,0.7,0)
.WetestH0:(
b11,b31)
=(
1,0)
.As-sumption(A4)holdswith
=
(
−∞,0]× [0,∞)
× R. Model 3 Let r=3and d=2,so that=[0,1]4. Let
θ
0=
(
b11,b21,b12,b22)
=(
0,0.8,0.6,0)
. We test H0:(
b11,b22)
=(
0,0)
.Assumption(A4)holdswith=[0,∞
)
× R2× [0,∞)
.Ifthenullhypothesiscannotberejected,itmeansthataMarshall–Olkinmodelsuffices.
Fig.3 showstheRMSEoftheparameterestimatesofthethreemax-linearmodelsexplainedabove,basedonthe
empiri-caltaildependencefunction(solidlines)andthebetataildependencefunction(dashedlines).Weseeagainthatthesetwo initialestimatorsleadtosimilarresults.Parameterswhosetruevaluesareontheboundaryarebetterestimatedforlowk, whilehighervaluesofkarepreferredforparameterswhosetruevaluesareintheinterioroftheparameterspace.
Table2 showstheempirical levelofthetest statisticsTn,k(1)andTn,k(2) fork∈{25,50,75,100},usinga significancelevel of0.05.The testsperformwell formodel 1ingeneral,while lowvaluesofk arenecessaryformodels 2and3.Thebeta taildependencefunctionoutperformstheempiricaltaildependencefunctionformodel2,whiletheoppositeisthecasefor model3.Overall,teststatisticTn(,k1) performssomewhatbetterthanTn(,k2).
Toassessthepowerofthetwoteststatistics,weconsiderthemodelsdescribedabove,butwelet
b11∈
{
0.8,0.85,...,1}
formodel1,b11∈
{
0.8,0.85,...1}
,b31∈{
0,0.05,...0.2}
formodel2, and b11,b22∈{
0,0.05,...0.2}
formodel3.Fig.4 showstheempiricalpowerofTn(,k1)formodels1–3forthebetataildependencefunction(resultsfortheempiricaltail
dependencefunction areverysimilar). Themoststriking resultofthelower-leftpanel isthatthegraphisnotsymmetric, i.e.,thepowerishigherwhenb22 isnearzero(andb11isnot)thanwhenb11 isnearzero(andb22 isnot).
4.2.2. Testingthenumberoffactorsinamax-linearmodel
Westudiedthe performanceof theteststatistics incasewewanted toidentifya specificsubmodel ofthemax-linear model.Wewouldalsoliketoinvestigatethecapabilityoftheteststatisticstocorrectlyretrievethetruenumberoffactors, i.e.,fors∈
{
1,...,r}
,wewishtotestthehypothesisH0:(
b1s,...,bds)
=0 .However,asmentionedinSection2.1,themax-linearmodelisonly definedforparameter valueswith dj=1bjs>0foralls∈
{
1,...,d}
;when thisconditionisnot met,50 100 150 200 250 300 0.000 0.010 0.020 Model 1, b11 = 1 k RMSE of b 11 50 100 150 200 250 300 0.00 0.04 0.08 Model 1, b21 = 0.7 k RMSE of b 21 50 100 150 200 250 300 0.00 0.04 0.08 Model 1, b31 = 0.2 k RMSE of b31 50 100 150 200 250 300 0.000 0.010 0.020 Model 2, b11 = 1 k RMSE of b 11 50 100 150 200 250 300 0.00 0.04 0.08 Model 2, b21 = 0.7 k RMSE of b 21 50 100 150 200 250 300 0.000 0.010 0.020 Model 2, b31 = 0 k RMSE of b31 50 100 150 200 250 300 0.000 0.010 0.020 0.030 Model 3, b11 = 0 k RMSE of b 11 50 100 150 200 250 300 0.00 0.04 0.08 Model 3, b21 = 0.8 k RMSE of b 21 50 100 150 200 250 300 0.00 0.05 0.10 0.15 Model 3, b12 = 0.6 k RMSE of b12 50 100 150 200 250 300 0.00 0.01 0.02 0.03 0.04 Model 3, b22 = 0 k RMSE of b 22
Fig. 3. RMSE for models 1–3 based on the empirical tail dependence function (solid lines), and the beta tail dependence function (dashed lines).
Table 3
Empirical level of T (2)
n,k based on the empirical and the beta tail dependence function for a significance level of 0.05.
k = 25 k = 50 k = 75 k = 100
emp beta emp beta emp beta emp beta
Tn,k(2) 4.6 23.9 2.7 24.9 2.4 23.8 2.5 20.0
theasymptoticdistributionoftheteststatisticsusing
θ
n,kratherthanθ
0.Weconsideramodelwithd=2andr=3,withparametervector
θ
0=(
b11,b21,b12,b22)
=(
0.8,0.6,0.2,0.4)
,i.e.,thethirdcolumnofBcontainsonlyzeroesandthemodelhaseffectivelytwofactors.Weestimateathree-factormodelandwetestH0:
(
b13,b23)
=(
0,0)
.Fig. 5 showsthe RMSEof the parameterestimates based on the empiricaltail dependence function (solidlines) and
thebetataildependencefunction(dashedlines).Weseethat, contrarytoprevious experiments,thebetataildependence hasmuchhigherRMSEthantheempiricaltaildependencefunction:oversmoothingleadstoalargebiasfor(b13,b23)(and
hencefortheotherparametersaswell).
Table 3 showsthe empirical level ofthe test statistic Tn(,k2) usinga significance level of 0.05. We included theresults basedonthebetataildependencefunctionforcompleteness:becauseittendstooverestimate(b13,b23),itrejectsthenull
hypothesisfartoooften.TeststatisticTn(1)
,k isnotpresentedbecauseforanyk andanyinitial estimator,ithasanempirical
levelof0.Evenwhen
|
θ
n,k−θ
(0)
n,k
|
islarge,|
fn,k(
θ
n,k)
− fn,k(
θ
(0)
n,k
)
|
issmallandhencethedeviance-typetestisnotadapted0.80
0.85
0.90
0.95
1.00
0
2
04
06
08
0
1
0
0
T
n,k(1)for k = 50 and k= 75
b
11P
o
w
e
r in %
Power of Tn,k (1) based on beta stdf 0.80 0.85 0.90 0.95 1.000.00 0.05 0.10 0.15 0.20 0 20 40 60 80 100 b11 b31 power Power of Tn,k (1) based on beta stdf 0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0 20 40 60 80 100 b11 b22 powerFig. 4. Top: empirical power in % for model 1 and k = 50 (dashed line), k = 75 (solid line), based on the empirical tail dependence function. Bottom: empirical power in % for model 2 (left) and model 3 (right) and k = 50 based on the beta tail dependence function.
50 100 150 200 250 300 0.00 0.10 0.20 0.30 b11 = 0.8 k rm se of b11 50 100 150 200 250 300 0.00 0.10 0.20 b21 = 0.6 k rm se of b21 50 100 150 200 250 300 0.00 0.05 0.10 0.15 0.20 b12 = 0.2 k rm se of b 12 50 100 150 200 250 300 0.00 0.05 0.10 0.15 b22 = 0.4 k rm se of b 22
Fig. 5. RMSE based on the empirical tail dependence function (solid lines) and the beta tail dependence function (dashed lines).
Weremarkthatthepowerissimilarwhenoneparameterisfarfromitsboundaryandwhenbothparametersarefarfrom theirboundary.
5. Application to stock market indices
ConsidertwomajorEuropeanstockmarketindices,theGermanDAXandtheFrenchCAC40.Wetakethedailynegative log-returnsofthepricesofthesetwoindicesfromhttps://finance.yahoo.com fortheperiodofJanuary1st,1997toDecember 31st,2017.Weremove all datesforwhich atleastone ofthetwo serieshasmissingvalues,endingup withasampleof
sizen=5313.Fig.7 showsthetimeseriesplotsoflog-returnsandthedependencestructure forthereturns standardized
Power of T
n,k(2)based on empirical stdf, k = 50
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0 20 40 60 80 100 b13 b23 powerPower of T
n,k(2)based on empirical stdf, k = 75
0.00 0.05 0.10 0.15 0.20 0.00 0.05 0.10 0.15 0.20 0 20 40 60 80 100 b13 b23 power
Fig. 6. Empirical power in % of T n,k(2) , based on the empirical tail dependence function for k = 50 (left) and k = 75 (right).
2000 2005 2010 2015 −10 −5 05 1 0
Daily negative log−returns of the DAX index
Year retur ns in % 2000 2005 2010 2015 −10 −5 0 5 10
Daily negative log−returns of the CAC40 index
Year
retur
ns in %
0 2 4 6 8
02468
DAX versus CAC40
Fig. 7. Left and middle: time series plots of daily negative log-returns of the DAX and the CAC40; right: scatterplot of daily negative log-returns of the DAX versus the CAC40 on the unit Pareto scale, plotted on the exponential scale.
Table 4
Parameter matrix for a bivariate max-linear model with r = 4 factors for k = 40 ; standard errors are in parentheses.
DAX 0.41 (0.11) 0.14 (0.05) 0.43 (0.11) 0.03 (0.05)
CAC40 0.43 (0.12) 0.44 (0.10) 0.13 (0.05) 0.00 (0.01)
Dickey–Fullertest,whichgivesp-valuesbelow0.01forbothseries,andaKPSStest,givingp-valuesabove0.1forbothseries. Hence,theassumptionofstationarityisreasonable.Letujdenotethe0.95%quantileofX1j,...,Xn j for j∈
{
1,...,d}
.WefitageneralizedParetodistribution(GPD)toXi j− uj
|
Xi j>uj andobtaintheparameterestimatesσ
1=1.12(0.11),γ
1=0.03(0.08),
σ
2=1.10(0.10)andγ
2=0.02(0.07).Table4 showstheestimatedparametermatrixB=
(
bjt)
j,t for j=1,2andt=1,2,3,4obtainedbyfittingafour-factormax-linearmodeltoourdatabasedontheempiricaltaildependencefunction.Valuesofk∈
{
25,30,...,70,75}
were con-sidered:wechose k=40becauseparameterestimatesarestablearoundthisvalue. Standarderrorswerecalculatedusing theasymptoticvariance matrixoftheestimator. TheresultsinTable4 suggestto testH0:(
b14,b24)
=(
0,0)
;thevalue oftheteststatisticTn,k(2) is0.04.Comparingwithacriticalvalueof17.4basedonasignificancelevelof0.05,wefindthatwe cannot rejectthe nullhypothesis.Critical valuesarecalculated bysimulation fromtheasymptoticdistributionof thetest statistics,whichisgiveninCorollaries3.4 and3.5.WedonotconsiderTn,k(1)orthebetataildependencefunctionbecauseof theirbadperformance(seeSection4.2.2).
Table 5
Parameter matrix for a bivariate max-linear model with r = 3 factors for k = 40 ; stan- dard errors are in parentheses.
DAX 0.41 (0.11) 0.14 (0.06) 0.46 (0.09)
CAC40 0.43 (0.12) 0.44 (0.10) 0.13 (0.05)
DAX versus CAC40
x y 0.2 0.4 0.6 0.8 1 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 200 150 100 50 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Tail dependence coefficient χ^n,k
k 4 6 8 10 12
1e−05
1e−03
1e−01
probability of a joint exceedance
return threshold
Fig. 8. Left: level sets {(x1 , x 2) : (x1 , x 2) = c} for c ∈ {0.2, 0.4, 0.8, 1} based on the empirical tail dependence function (solid lines) and on the fitted max-linear model (dashed lines); middle: nonparametric (black dots) and model-based (red dots) estimates of the tail dependence coefficient χn,k for
k ∈ { 200 , 195 , . . . , 30 } ; right: probability of a joint exceedance based on the fitted max-linear model (solid line) and on the empirical excesses (dashed line). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Table5 showstheestimatedparameter matrixB=
(
bjt)
j,t for j=1,2andt=1,2,3forathree-factormodel.We testifa Marshall–Olkinmodel suffices,i.e., H0:
(
b12,b23)
=(
0,0)
. Thevalue ofthe teststatistic Tn(,k2) is75.9. Comparingto acriticalvalueof8.03,werejectthenullhypothesisofaMarshall–Olkinmodel.
Fig.8 showstwogoodness-of-fitmeasuresandtheestimatedprobabilityofajointexceedancesofthetwostocks.Onthe
left,weplottedthelevelsets
{
(
x1,x2)
:(
x1,x2)
=c}
forc∈{0.2,0.4,0.8,1}basedontheempiricaltaildependencefunction(solidlines)andonthefittedmax-linearmodel(dashedlines).
Acommonsummarymeasureofdependenceisthetaildependencecoefficient,
χ
(
q)
:=limq↑1P[F1
(
X11)
<q|
F2(
X12)
<q]=2−(
1,1)
,see(2).Fig.8 (middle)showsnonparametric estimatesofthetaildependencecoefficient (black dots)
χ
n,k=2− n,k(
1,1)
anditsmodel-basedcounterpart(reddots)fordecreasingk.Thehorizontalgreylinecorrespondstothevaluek=40that wasused forthe parameterestimatorsandthetests, correspondingto
χ
n,k≈ 0.68,whichis equaltothemodel-basedχ
. Thedottedlinescorrespondto95%pointwisebootstrapconfidenceintervals.TheprobabilityofajointexcessF
(
x1,x2)
=P[X11>x1,X12>x2]canbeapproximatedforlarge(x1,x2)by F(
x1,x2)
=x∗1+x∗2− 1+exp{
−(
x∗1,x∗2)
}
forx∗j=1− Fj(
xj)
, j=1,2.LetH
(
· ;γj,σ
j)
denotethesurvivalfunctionofaGPDwithshapeγ
jandscaleσ
j.Forx1>u1 andx2>u2,wecanestimateF
(
x1,x2)
by F(
x 1,x2)
=x1∗+x∗2− 1+exp −(
x∗1,x∗2;θ
n,k)
,where
(
· ;θ
n,k)
denotes themax-linearstabletaildependencefunctionbasedonthefittedparameterestimatesandx∗j=0.05× H
(
xj− uj;γ
j,σ
j)
.Theright-handplotofFig.7 showstheprobabilitiesF(
x,x)
onthelog-scaleforx∈{
2.5,3,...,12}
(solidline).Thesecanbecomparedtotheempiricalexcessprobabilitiesn−1 ni=11
{
X1i>x,X2i>x}
(dashedline),whicharezeroforx>8.5. 6. Discussion
Weproposedtwoteststatisticsforparametersthat,underthenullhypothesis,areontheboundaryofthealternative hy-pothesis.Theseteststatisticsareconvenientbecausetheirasymptoticdistributionhasanexplicitexpression,andhavebeen showntoperformwellinsimulationstudies.Inpractice,their asymptoticdistributionbecomescumbersomewhentesting fora higherdimensional (c>4) parametervector. Apossible solution mightbe to consider a multipletesting procedure withaBonferroni-typeofcorrection.