You are on page 1of 6

Copulas:PartII...

acontinuationofPartI
We'retalkingaboutgeneratingjointdistributionsfortwovariables,sayxandy,withprescribed
properties.
>Justtwo?
Payattention!Beforewerun,wewalk.
Forexample,weintroducedthe1parameterfamilyof
Frank'sCopulas:
[F1]

where(forthemoment)eachofxandyliein[0,1]...andso
doesC(x,y).
Wealsowrotethisas:
[F2]

Notethatifuandvareuniformlydistributedrandom
variableson[0,1],
thenC(x,y)istheprobabilitythatuxandvy.
Thatis:C(x,y)=Prob[ux,vy].

Figure1

TogenerateFrank'sCopula,weintroduce(ta
DUM!):
g(z)=log[(1edz)/(1ed)]
ThenwegenerateFrank'sCopulavia:
[A]g(C)=g(x)+g(y)
Thatis:
log[(1edC)/(1ed)]=log[(1edx)/
(1ed)]log[(1edy)/(1ed)]
IfwesolvethisforC,we'llget[F1].
>Andwhatdoesthatgguylooklike?
LikeFigure2a.See?It'sjustasimple,
monotone(decreasing)functionon[0,1].
Indeed,forznear0,g(z)behaveslikelog[z].

Ofcourse,that'snottheonlysuchg(z)onecan
invent.Howabout:
g(z)=(1/d)(1zd)
SeeFigure2b?It'salsomonotone...but
decreasing(ford>1).
Ifwewriteg(C)=g(x)+g(y)andsolveforC,

Figure2a

wegetClayton'sCopula:
C(x,y)=[xd+yd1]1/d
Then,ofcourse,there'sGumbel'sCopulawhere
g(z)=[log(z)]dwherelog(z)<0when0<
z<1
sothat,puttingg(C)=g(x)+g(y)andsolving
forCgives...
>Ithinkthat'senough.IfIpickone,sayGumbo,
whatdoIdowithit?
Goodquestion...
Figure2b

theRightStuff
Supposewehavetwosetsofdata,sayxandy,andwewishtoidentifytheMeanandStandard
DeviationofeachandtheCorrelation(ormaybeweshouldcallittheDependence)betweenthem.
That'sfivevariables:TheMeansandSDs,sayM[x],M[y],S[x],S[y]andthe"Dependence",say
K.
>We'llcalculatetheDependenceusingCopulas?
Whenwe'refinishedwe'llbeabletogeneratexandydatawiththeprescribedMeansand
StandardDeviationsandtheprescribedcorrelationor"dependence".
Althoughwewillusuallybespeakingofstockreturns,thetwosetsofdatamaybetheaverageage
ofdriversin50states(that'sx1,x2,...x50)andtheannualnumberofautoaccidentsinthose50
states(that'sy1,y2,...y50).Ormaybethexandydatarepresenttheannualcostofinsurance
claimsandthecostoflawyer'sfees.Ormaybetheefficacyofadrugortherelationshipbetween
$USDand$CADormaybe...
>Canwecontinue?
Okay,butrememberthatwe'reinterestingintheDependenceofoneontheother.
Infact,wecouldeasilylookatthedistributionofthexandydataandinventsomedistribution
functionthatfitsprettywell.
(Normalorlognormal,forexample).
Aah,butifwethentrytosimulatethesevariablesbyrandomselectionfromourtwoinvented
distributions,itishighlyunlikelythatwe'llgetanymeaningfulcorrelation.
Forexample,supposewedothis,assumingthatbothsetslookliketheyarenormallydistributed:
1. ThexsethasMean=1.0andSD=3.0andtheysethasMean=2.0andSD=4.0.
2. Wegenerateahundredvaluesofxandyvariablesbyrandomselectionfromtwonormal
distributionswiththeseparameters.
3. Weplotthepoints(xk,yk)thegetascatterplotandwecalculatesomecorrelation.
4. Eachtimewedosteps2and3we'llgetdifferentresults

>SouseCopulas!
Goodidea!
We'lltalkaboutArchimedeanCopulaswhicharegeneratedvia[A],above,whereChasarguments
whichbothliein[0,1].
IfF1(x)andF2(y)arethecumulativedistributionfunctionsforthexandydata(xandy,ofcourse,are
NOTrestrictedto[0,1]),
thenF1(x)andF2(y)willliein[0,1].
Foragivenxvalue,sayU,what'stheprobabilitythatxU?It'sF1(U)...thecumulativeprobability
distributionforthexvariable..
Foragivenyvalue,sayV,what'stheprobabilitythatyV?It'sF2(V)...thecumulativeprobability
distributionfortheyvariable.

Andwhat'stheprobabilitythatxUandyV?
It'sajoint,cumulativeprobabilityFthatwe'llgenerateusingCopulas,namely:
[B1]F(U,V)=C(F1(U),F2(V))whereF1(U)andF2(V)aretheprobabilitiesthatxUandyV.
ThisdescribesajointdistributionfunctionevaluatedatU,Vwheretheindividualdistributions
(calledthemarginaldistributions)areF1andF2.
IfweselectsomeCopulaandweknow(orassume)distributionsforeachvariable,then[B1]will
provideajointdistribution.
>Wait!IfF1(U)istheprobabilitythatxUandF2(V)istheprobabilitythatyVthenthe
probabilitythatxUandyVisF1(U)F2(V).
Goodthinking!Considerthis:

Wemakearandomselectionof100jillionvaluesofxand60jilliontimeswefindthatxU.
WeconcludethatF1(U)=0.60fromwhichweconcludethat,60%ofthetime,xU.
Wealsomakearandomselectionof100jillionvaluesofyandfindthat30jilliontimeswegotyV.
WeconcludethatF2(V)=0.30fromwhichweconcludethat,30%ofthetime,yV.
Wenowhave100jillionselectionsofbothxandyand60jilliontimeswehadxUso,ofthese,we'd
expect30%willhaveyVand30%of60jillionis18jillionsotheprobabilityofhavingxUandy
Vis18outof100jillionor0.18whichisF1(U)F2(V)=0.60*0.30.

>Sowhyallthefussaboutcopulasandjointdistributionsand...?
We'duseF=F1F2ifthetwovariableswereindependent.
However,we'retalkingabout"dependence"so,havingselectedanx,theyvaluemayberestricted
becauseofthatdependence.
>Whydidn'tyousaythatbefore?
Hey!I'mjustlearning'boutthisstuff.AfewdaysagoIdidn'tknowacopulafromacupola.
>Okay,butin[B1],you'reassumingyouknowtheindividualdistributions,right?
Yes,inthiscase,but...
>Andyou'dgetadifferentjointdistributioneverytimeyoudecidetochangecopulas,right?
Well,yesbut...
>IfIjustgaveyouthejointdistribution,wouldyouacceptitorwouldyoupickacopula?
Aah,goodpoint!Infact,nomatterwhatjointdistributionFunctionyougaveme,there'llalwaysbe
aCopulasatifying[B1].
>You'rekidding,right?NomatterwhatjointdistributionIinvent?
Yes...that'sSklar'sTheorem.Infact:
ACopulaconnectsagivenjointdistributionFunctiontoitsmarginaldistributions...them'sthe
individualdistributionfunctionsF1,F2.
TheCopuladescribesthedependencebetweenvariables,regardlessoftheirindividual
distributions.
Copulaswereinvented(in1959?)inordertoisolatethedependencestructurebetween
variables.
NomatterwhatF1andF2are,thenumbersF1(U)andF2(V)willbeuniformlydistributedon
[0,1]!
SincetheargumentsofCarejustuniformlydistributedrandomvariabes(namelyF1(U)and
F2(U)),what'sleftistheirdependence.
Infact,wecanlinktogetheranytwomarginaldistributions(F1,F2)andanyCopulaandwe'll
getavalidjointdistributionFunction.
Infact...

>Okay!Canwecontinue?
Sure.Let's...
>Wait!Don'tyoufindthatconfusing?Firstyousayxandy,
thenuandv,thenUandV?

Here'sapicture...Figure3.
xisourrandomvariableandUisaparticularvalueofxand
ifweask:
"What'stheprobabilitythatxU?"
theanswerisu=F1(U).
NotethatxhenceUcanhaveanyvalue,but0u1.

Figure3

Morestuff
We'venotedthatu=F1(U)andv=F2(V)arejustuniformlydistributedvariables,regardlessofthe
functionsF1andF2.
Further,uandvareuniformlydistributedin[0,1]...sincethey'reprobabilities.
Further,ifwesolveforUandVwe'dget:U=F11(u)andV=F21(v)...whereF11andF21arethe
inversefunctions.
>Inversewhat?
Ifa=f(b)=b3wherefisthe"cubefunction",andwesolvefor
b=f1(a)=a1/3,wecallthat"cuberoot"functiontheinverseofthecubefunctionandusethe
notationf1.Gotit?
Wedothatwheneverwecansolve"backwards",givingthenamef1totheinverseoffandthe
nameg1totheinverseofgand
...uh,thatremindsme.Ifwelookbackat[A],itsays:g(C)=g(x)+g(y).
IfwesolveforCwe'dwritethisasC=g1(g(x)+g(y)).
>Yeah...Igotit.
Ihaveabetterexample.LookagainatFigure3.
IfweknowUwegettoubygoingNorththenWest,alongtheredlinetou=F1(U).
However,ifweknowu,wegoEastthenSouthtoU=F11(u).
>Igotit!
Anyway,wecannowwrite[B1]as:
[B2]C(u,v)=F(F11(u),F21(v))
>zzzZZZ
We'realmostthere!Ournewequationinvolvesjusttheuniformlydistributedvariablesuandv,
see?
Tomodelthemwejustselectnumbersatrandomin[0,1]using,forexample,ExcelsRAND()
function,see?
>zzzZZZ

Indeed,wecannowconclude(viaSklar'stheorem)that:
IfweknowF1andF2,thedistributionsofxandy(whereF1(U)istheprobabilitythatxUandF2(V)
istheprobabilitythatyV),
andwepickourfavouriteCopula,thenthejointdistributionFisgivenby[B1]:
F(U,V)=C(F1(U),F2(V))
Also:
IfweknowthejointdistributionFandmarginaldistributionsF1andF2,thenthereisa
uniqueCopulaaccordingto[B2],namely:
C(u,v)=F(F11(u),F21(v))
Theneatthingis,wecanpickourCopulatomodelthedependencebetweenvariableswithout
consideringwhichtypeofmarginals.
>Huh?
Thexvariablemaybenormallydistributedandtheyvariablemaybelognormalormaybeat
distributionormaybeuniformlyormaybe...
>Yeah,butwhatCopulaandhowdoyouchangethedependence?
HavingchosenyourfavouriteCopula,youchangetheparameter,suchasthedinClayton's
Copula:g(z)=[xd+yd1]1/d
Aah,butwhatCopula,eh?
>That'swhatIsaid!

forPartIII

You might also like