Statistics& Statistics&OperationsResearchTransactions OperationsResearch SORT40(2)July-December2016,267-302 © Institutd’Estad´ısticadeCatalunya Transactions ISSN:1696-2281 [email protected] eISSN:2013-8830 www.idescat.cat/sort/ A construction of continuous-time ARMA models by iterations of Ornstein-Uhlenbeck processes ArgimiroArratia1, AlejandraCaban˜a2 and EnriqueM.Caban˜a3 Abstract Wepresentaconstructionofafamilyofcontinuous-timeARMAprocessesbasedon piterations ofthelinearoperatorthatmapsa Le´vyprocessontoanOrnstein-Uhlenbeckprocess. Thecon- structionresemblestheproceduretobuildanAR(p)fromanAR(1).Weshowthatthisfamilyisin facta subfamilyofthewell-knownCARMA(p,q) processes,withseveralinterestingadvantages, including a smaller number of parameters. The resulting processes are linear combinations of Ornstein-UhlenbeckprocessesalldrivenbythesameLe´vyprocess.Thisprovidesastraightfor- wardcomputationofcovariances,astate-spacemodelrepresentationandmethodsforestimating parameters.Furthermore,the discrete and equallyspaced samplingof the process turnsto be an ARMA(p,p−1) process. We propose methods for estimating the parameters of the iterated Ornstein-UhlenbeckprocesswhenthenoiseiseitherdrivenbyaWieneroramoregeneralLe´vy process,andshowsimulationsandapplicationstorealdata. MSC: 60G10,62M10,62M9960M99. Keywords: Ornstein-Uhlenbeckprocess,Le´vyprocess,ContinuousARMA,stationaryprocess. 1. Introduction Thelink betweendiscrete time autoregressivemovingaverage(ARMA) processesand stationaryprocesseswithcontinuous-timehasbeenofinterestformanyyears,seeforin- stance,Doob(1944),Durbin(1961),Bergstrom(1984,1996)andmorerecentlyBrock- well (2009), Thornton and Chambers (2013). Continuous time ARMA processes are better suited than their discrete counterpartsfor modelling irregularly spaceddata, and when the white noise is driven by a non-Gaussian process it becomes a more realistic modelinfinanceandotherfieldsofapplication. 1UniversitatPolite`cnicadeCatalunya,Barcelona,[email protected] SupportedbySpain’sMINECOprojectAPCOM(TIN2014-57226-P) andGeneralitatdeCatalunya2014SGR 890(MACDA). 2UniversitatAuto`nomadeBarcelona,[email protected] SupportedbySpain’sMINECOprojectMTM2015-69493-R. 3UniversidaddelaRepu´blica,Montevideo,[email protected] Received: November2015 Accepted: May2016 268 Aconstructionofcontinuous-timeARMAmodelsbyiterationsofOrnstein-Uhlenbeckprocesses A popular continuous-time representation of ARMA(p,q) process (known as CARMA(p,q))canbeobtainedviaastate-spacerepresentationoftheformalequation a(D)Y(t)=σb(D)DΛ(t), where σ > 0 is a scale parameter, D denotes differentiation with respect to t, Λ is a second-orderLe´vyprocess,a(z)=zp+a zp−1+...+a isapolynomialoforder pand 1 p b(z)=b +b z+...+b zqapolynomialoforderq≤ p−1withcoefficientb 6=0(see, 0 1 q q e.g., Brockwell, 2004, 2009, Thornton and Chambers, 2013). The parameters of this modelare estimated by adjusting first an ARMA(p,q), q< p to regularly spaceddata. Then obtain the parameters of the continuous version whose values at the observation timeshavethesamedistributionofthefittedARMA.Hence, p+q+1parametershave tobeestimated. We proposein this work a parsimoniousmodelfor continuousautoregression,with fewer parameters (as we shall see exactly p plus the variance). Our construction de- partsfromtheobservationthataOrnstein-Uhlenbeck(OU)processcanbethoughtofas continuous-timeinterpolationofanautoregressiveprocessoforderone(i.e.anAR(1)). Thisis shownin Section2,wherewealsoreviewsomewellknownfactsonLe´vypro- cesses,ARMAmodelsandtheirrepresentations.Themodelisobtainedbyaprocedure that resembles the one that allows to build an AR(p) from an AR(1). Departing from thisanalogy,wedefineandanalysetheresultofiteratingtheapplicationoftheoperator that maps a Wiener process onto an OU process. This operator is defined in Section 3 anddenotedOU,withsubscriptsdenotingtheparametersinvolved. The p iterations of OU, for each positive integer p, give rise to a new family of processes, the Ornstein-Uhlenbeck processes of order p, denoted OU(p). They can be used as models for either stationary continuous-time processes or the series obtained by observing these continuous processes at equally spaced instants. We show that an OU(p) process can be expressedas a linear combination of ordinary OU processes,or generalizedOUprocesses,alsodefinedinSection3.Thisresultresemblestheaggrega- tionsofGaussian(andnon-Gaussian)processesstudiedwiththeideaofdeconstructing a complicated economic model into simpler constituents. In the extensiveliterature on aggregations (or superpositions) of stochastic processes the aggregated processes are driven by independent Le´vy processes (see, e.g., Granger and Morris, 1976, Granger, 1980,Barndorff-Nielsen,2001,Eliazar andKlafter, 2009,amongmanyothers). A dis- tinctivepointofourconstructionisthatthestochasticprocessesobtainedbyconvolution oftheOU operatorresultin a linearcombinationcomprisedofprocessesdrivenbythe sameLe´vyprocess. Another consequence of writing the OU(p) process as the aggregation of simpler onesisthederivationofaclosedformulaforitscovariance.Thishasimportantpractical implications since it allows to easily estimate the parameters of a OU(p) process by matchingcorrelations(aprocedureresemblingthemethodofmoments,tobedescribed inSection6.2),andbymaximumlikelihood. ArgimiroArratia,AlejandraCaban˜aandEnriqueM.Caban˜a 269 In Section 4 we show how to write the discrete version of a OU(p) as a state-space model, and from this representation we show in Section 5 that for p>1, a OU(p) be- haves like an aggregation of AR processes (in the manner considered in Granger and Morris (1976)), that turns outto be anARMA(p,q),with q≤ p−1.Consequentlythe OU(p)processesareasubfamilyoftheCARMA(p,q)processes.Notwithstandingthis structural similarity, the family of discretized OU(p) processes is more parsimonious than the family of ARMA(p,p−1) processes, and we shall see empirically that it is abletofitwelltheautocovariancesforlarge lags.Hence,OUprocessesofhigherorder appear as a new continuous model, competitive in a discrete time setting with higher order autoregressive processes (AR or ARMA). The estimation of the parameters of OU(p)processesisattemptedinSection6.Simulationsandapplicationstorealdataare providedinSection6.5.OurconcludingremarksareinSection7. 2. Preliminaries LetusrecallthataLe´vyprocessΛ(t)isaca`dla`gfunction,withindependentandstation- ary increments, that vanishesint =0. As a consequence,Λ(t) is, for eacht, a random variable with an infinitely divisible law (Sato, 1999). A Wiener process W is a cen- tred Gaussian process, with independent increments and variance E(W(t)−W(s))2 = σ2|t−s|. Wiener processes are the only Le´vy processes with almost surely continu- ous paths. For parameter λ>0 the classical Ornstein-Uhlenbeckprocess is defined as t e−λ(t−s)dW(s)(UhlenbeckandOrnstein,1930). −∞ Wiener processcanbereplacedbya secondorderLe´vyprocessΛ to definea Le´vy R drivenOrnstein-Uhlenbeckprocessas t x(t)(=x (t)):= e−λ(t−s)dΛ(s) (1) λ,Λ Z−∞ Thepreviousequationcanbeformallywrittenindifferentialform dx(t)=−λx(t)dt+dΛ(t) (2) We may think of x as the result of accumulating a random noise dΛ, with reversion to themean(thatweassumetobe0)ofexponentialdecaywithrateλ. WhentheOrnstein-Uhlenbeckprocessxissampledatequallyspacedtimes{hτ :h= 0,1,2,...,n},τ >0,theseriesX =x(hτ)obeysanautoregressivemodeloforder1(i.e. h (h+1)τ an AR(1)), because X =e−λτX +Z , where Z = e−λ((h+1)τ−s)dΛ(s), h+1 h h+1 h+1 hτ isthestochasticinnovation. Z Hence,wecanconsidertheOUprocessascontinuous-timeinterpolationofanAR(1) process. Notice that both models are stationary. This link between AR(1) and OU(1) suggeststhedefinitionofiteratedOUprocessesintroducedinSection3. 270 Aconstructionofcontinuous-timeARMAmodelsbyiterationsofOrnstein-Uhlenbeckprocesses An ARMA(p,q) or autoregressive moving average process of order (p,q) has the followingform x =φ x +···+φ x +θ ǫ +θ ǫ +···+θ ǫ t 1 t−1 p t−p 0 t 1 t−1 q t−q where φ , ...,φ are the autoregressiveparameters,θ , ...,θ are the movingaverage 1 p 0 q parameters,andthewhite-noiseprocessǫ hasvarianceone.DenotebyB thebackshift t operatorthatcarries x into x .By consideringthepolynomialsin thebackshiftoper- t t−1 ator, φ(B)=1−φ B−···−φ Bp and θ(B)=θ +θ B+···+θ Bq 1 p 0 1 q theARMA(p,q)modelcanbewrittenas φ(B)x =θ(B)ǫ (3) t t This compact expression comes in handy for analysing structural properties of time series. It also links to the representation of ARMA processes as a state-space model, useful for simplifying maximum likelihood estimation and forecasting. A state-space modelhasthegeneralform Y =AY +ηηη (4) t t−1 t x =KTY +N (5) t t t where (4) is the state equation and (5) is the observation equation, with Y the m- t dimensionalstatevector,AandK arem×mandm×kcoefficientmatrices,KT denotes thetransposeofK,ηηη andN are mandk dimensionalwhite noises.N wouldbepresent onlyiftheprocessx isobservedsubjecttoadditionalnoise(seeBox,Jenkins,andRein- t sel,1994forfurtherdetails).WepresentinSection4astate-spacemodelrepresentation ofourgeneralizedOUprocess. 3. Ornstein-Uhlenbeck processes of order ppp The AR(1) process X = φX +ǫ , where ǫ , t ∈Z, is a white noise, can be written t t−1 t t as (1−φB)X = ǫ using the back-shift operator B. Equivalently, X can be written t t t as X = MA ǫ, where MA is the moving average that maps ǫ onto MA ǫ ,= t 1/ρ t 1/ρ t 1/ρ t ∞ 1 ǫ ,andρ(=1/φ)istherootofthecharacteristicpolynomial1−φz. j=0ρj t−j P ArgimiroArratia,AlejandraCaban˜aandEnriqueM.Caban˜a 271 p Moreover, the AR(p) process X = φ X +ǫ ( or φ(B)X =ǫ ), where φ(z)= t j t−j t t t j=1 X p p 1− φ zj=∏(1−z/ρ )hasrootsρ =eλj,j=1,...,p,canbeobtainedbyapplying j j j j=1 j=1 thecXompositionofthemovingaveragesMA tothenoise,thatis: 1/ρj p X =∏MA ǫ t 1/ρj t j=1 NowconsidertheoperatorMAe−λ thatmapsǫt onto MAe−λǫt = e−λ(t−l)ǫl l≤t,integer X AcontinuousversionofthisoperatorisOU thatmapsy(t),t ∈Ronto λ t OU y(t)= e−λ(t−s)dy(s), (6) λ Z−∞ whenever the integral can be defined. The definition of OU is extended to include λ complexprocesses,byreplacingλbyκ=λ+iµ,λ>0,µ∈Rin(6).Thesetofcomplex numberswith positiverealpartisdenotedbyC+, andthe conjugateofκis denotedby κ¯. For p≥1andparametersκκκ=(κ ,...,κ ),thepreviousargumentsuggeststodefine 1 p thefollowingprocessobtainedasrepeatedcompositionsofoperatorsOU , j=1,...,p: κj p OU y(t):=OU OU ···OU y(t)=∏OU y(t) (7) κκκ κ1 κ2 κp κj j=1 ThisiscalledOrnstein-Uhlenbeckprocessoforderpwithparametersκκκ=(κ ,...,κ )∈ 1 p (C+)p. The composition ∏p OU is unambiguouslydefined becausethe application j=1 κj ofOU operatorsiscommutativeasshowninTheorem1(i)below. κj The particular case of interest where the underlying noise is a second order Le´vy processΛ,namely, p OU Λ(t):=OU OU ···OU Λ(t)=∏OU Λ(t) (8) κκκ κ1 κ2 κp κj j=1 is called the Le´vy-driven Ornstein-Uhlenbeck process of order p with parametersκκκ= (κ ,...,κ )∈(C+)p. 1 p 272 Aconstructionofcontinuous-timeARMAmodelsbyiterationsofOrnstein-Uhlenbeckprocesses Fortechnicalreasons,itisconvenienttointroducetheOrnstein-Uhlenbeckoperator OU(h) ofdegreehwithparameterκthatmapsyonto κ t (−κ(t−s))h OU(h)y(t)= e−κ(t−s) dy(s) (9) κ Z−∞ h! andΛonto t (−κ(t−s))h ξ(h)(t)= e−κ(t−s) dΛ(s) (10) κ Z−∞ h! We call the process (10) generalized Ornstein-Uhlenbeck process of order 1 and degree h. For the remainder of the paper we restrict the underlying noise to a second orderLe´vyΛ,butnotethatthegeneralpropertiesoftheOU operatorthatwearegoing κ toshowholdforanyrandomfunctiony(t)forwhichtheintegral(6)isdefined. 3.1. Properties of the operatorOOOUUU κκκ Thefollowingstatementssummarizesomepropertiesofproducts(compositions)ofthe operatorsdefinedby(7)and(9),andcorrespondingly,ofthestationarycentredprocesses ξ(h), h ≥ 0. In particular, the Ornstein-Uhlenbeck processes of order 1 and degree 0, κ ξ(0)=ξ aretheordinaryOrnstein-Uhlenbeckprocesses(1). κ κ Theorem1 (i) Whenκ 6=κ ,theproductOU OU canbecomputedas 1 2 κ2 κ1 κ κ 1 OU + 2 OU κ −κ κ1 κ −κ κ2 1 2 2 1 andisthereforecommutative. (ii) Thecomposition∏p OU constructedwithpairwisedifferentκ ,...,κ isequal j=1 κj 1 p tothelinearcombination p p ∏OU = K (κ ,...,κ )OU , (11) κj j 1 p κj j=1 j=1 X withcoefficients 1 K (κ ,...,κ )= . (12) j 1 p ∏ (1−κ/κ ) κl6=κj l j ArgimiroArratia,AlejandraCaban˜aandEnriqueM.Caban˜a 273 (iii) Fori=1,2,...,OU OU(i)=OU(i)−κOU(i+1). κ κ κ κ (iv) Foranypositiveinteger pthe p-thpoweroftheOrnstein-Uhlenbeckoperatorhas theexpansion p−1 p−1 OUp = OU(j). (13) κ j κ j=0(cid:18) (cid:19) X (v) Letκ ,...,κ bepairwisedifferentcomplexnumberswithpositiverealparts,and 1 q p ,...,p positiveintegers,andletusdenotebyκκκacomplexvectorin(C+)pwith 1 q componentsκ repeated p times, p ≥1, h=1,...,q, q p = p. Then, with h h h h=1 h K (κκκ)definedby(12), h P q q 1 q ∏OUph = OUph = K (κκκ)OUph. h=1 κh h=1 ∏l6=h(1−κl/κh)pl κh h=1 h κh X X An immediate consequenceis that the operatorOU with p-vectorparameterκκκ canbe κκκ written as a linear combination of p operators OU or OU(h) for suitable scalar values κ κ κ and non-negative integer h. Therefore, the process OU Λ can be written as a linear κκκ combinationofOUprocessesdrivenbythesameLe´vyprocess,asstatedinthefollowing Corollary. Corollary1 q (i) TheprocessOU (Λ)=∏OUph(Λ)canbeexpressedasthelinearcombination κκκ κh h=1 q ph−1 OU (Λ)= K (κκκ) ph−1 ξ(j) (14) κκκ h j κh h=1 j=0 X X(cid:0) (cid:1) ofthe pprocesses{ξ(j):h=1,...,q,j=0...,p −1}(see(10)). κh h (ii) Consequently, q ph−1 OUκκκΛ(t)= Kh(κκκ) ph−j 1 −t∞e−κh(t−s)(−κh(jt!−s))jdΛ(s) h=1 j=0 X X(cid:0) (cid:1)R Corollary2 For real λ,µ, with λ>0, the product OU OU is real, that is, ap- λ+iµ λ−iµ pliedtoarealprocessproducesarealimage. TheproofsofTheorem1andcorollariesareinAppendixA. 274 Aconstructionofcontinuous-timeARMAmodelsbyiterationsofOrnstein-Uhlenbeckprocesses 3.2. Computingthe covariances Therepresentation q ph−1 p −1 x:=OU (Λ)= K (κ) h OU(j)(Λ) κκκ h j κh h=1 j=0 (cid:18) (cid:19) X X of x as a linear combination of the processes ξ(i) = OU(i)(Λ) allows a direct compu- κh κh tation of the covariances γ(t) = Ex(t)x¯(0) through a closed formula, in terms of the covariancesγ(i1,i2)(t)=Eξ(i1)(t)ξ¯(i2)(0): κ1,κ2 κ1 κ2 q ph′−1 q ph′′−1 p −1 p −1 γ(t)= K (κκκ)K¯ (κκκ) h′ h′′ γ(i′,i′′) (t) (15) h′ h′′ i′ i′′ κh′,κh′′ h′=1 i′=0 h′′=1 i′′=0 (cid:18) (cid:19)(cid:18) (cid:19) X X X X withv2=VarΛ(1), 0 (t−s)i1 (−s)i2 γ(i1,i2)(t)=v2(−κ )i1(−κ¯ )i2 e−κ1(t−s) e−κ¯2(−s) ds κ1,κ2 1 2 Z−∞ i1! i2! i1 i tj 0 =v2(−κ )i1(−κ¯ )i2e−κ1t 1 e(κ1+κ¯2)s(−s)i1+i2−jds 1 2 j=0(cid:18)j(cid:19)i1!i2!Z−∞ X v2(−κ )i1(−κ¯ )i2e−κ1t i1 tj(i +i − j)! 1 2 1 2 = (16) i2! j!(i1− j)!(κ1+κ¯2)(i1+i2−j+1) j=0 X Arealexpressionforthecovariancewhentheimaginaryparametersappearasconjugate pairscanbeobtainedbutitismuchmoreinvolvedthanthisone. 4. The OU(ppp) process as a state-space model Theorem 1 and its corollaries lead to express the OU(p) process by means of linear state-space models. The state-space modelling provides a unified methodology for the analysisoftimeseries(seeDurbinandKoopman,2001). In the simplest case, where the elements of κκκ are all different, the process x(t) = OU Λ(t)isalinearcombinationofthestatevectorξξξ (t)=(ξ (t),ξ (t),...,ξ (t))T, κκκ κκκ κ1 κ2 κp whereξ =OU (Λ). κj κj ArgimiroArratia,AlejandraCaban˜aandEnriqueM.Caban˜a 275 Moreprecisely,thevectorialprocess ξξξ (t)=(ξ (t),ξ (t),...,ξ (t))T, ξ =OU (Λ) κκκ κ1 κ2 κp κj κj andx(t)=OU Λ(t)satisfythelinearequations κκκ ξξξ (t)=diag(e−κ1τ,e−κ2τ,...,e−κpτ)ξξξ (t−τ)+ηηη (t) (17) κκκ κκκ κκκ,τττ and x(t)=KKKT(κκκ)ξξξ(t), (18) t ηηη (t)=(η (t),η (t),...,η (t))T, η (t)= e−κj(t−s)dΛ(s), κκκ,τττ κ1,τ κ2,τ κp,τ κj,τ Zt−τ t 1−e−(κj+κ¯l)τ Var(ηηη (t))=v2((v )), v = e−(κj+κ¯l)(t−s)ds= (19) κκκ,τττ j,l j,l κ +κ¯ Zt−τ j l andthecoefficientsfrom(12),KKKT(κκκ)=(K (κκκ),K (κκκ),...,K (κκκ)). 1 2 p The initial value ξξξ(0) is estimated by means of its conditional expectation ξξξˆ(0)= KKKT(κκκ)Vx(0) 1 E(ξ(0)|x(0))= ,withV =Var(ξξξ(0))= . KKKT(κκκ)VKKK (cid:18)(cid:18)κj+κ¯l(cid:19)(cid:19) AnapplicationofKalmanfiltertothisstate-spacemodelleadstocomputethelikeli- hoodofxxx=(x(0),x(τ),...,x(nτ)).SomeKalman filterprograms includedin software packages require the processes in the state-space to be real. That condition is not ful- filled by the modeldescribedby equations(17) and (18). An equivalentdescriptionby meansofrealprocessescanbeobtainedbyorderingtheparametersκκκwiththeimaginary componentspairedwith theirconjugatesinsuchawaythatκ =κ¯ ,h=1,2,...,c 2h 2h−1 andtheimaginarycomponentℑ(κ )=0ifandonlyif2c< j≤ p. j Then the matrix M=((M )) with all elements equal to zero except M = j,k 2h−1,2h−1 M =1,−M =M =i, h=1,2,...,candM =1, 2c< j≤ p,induces 2h−1,2h 2h,2h−1 2h,2h j,j thelineartransformationξξξ 7→Mξξξ thatleadstothenewstate-spacedescription Mξξξ(t)=Mdiag(e−κ1τ,e−κ2τ,...,e−κpτ)M−1Mξξξ(t−τ)+Mηηη(t), (20) x(t)=KKKTM−1Mξξξ(t), (21) wheretheprocessesMξξξ arereal. 276 Aconstructionofcontinuous-timeARMAmodelsbyiterationsofOrnstein-Uhlenbeckprocesses Observethatthereisnolossofgeneralityinchoosingthespacingτ betweenobser- vations as unity for the derivation of the state-spaceequations. Hence, we set τ =1 in thesequeland,inaddition,τ willbeomittedfromthenotation. When κ ,...,κ are all different, p ,...,p are positive integers, q p = p and 1 q 1 q h=1 h κκκ is a p-vector with p repeated components equal to κ , the OU(p) process x(t) = h h P OU Λ(t)isalinearfunctionofthestate-spacevector κκκ ξ(0),ξ(1),...,ξ(p1−1),...,ξ(0),ξ(1),...,ξ(pq−1) κ1 κ1 κ1 κq κq κq (cid:16) (cid:17) where the components are given by (10), and the transition equation is no longer ex- pressedbyadiagonalmatrix.Inthiscasethestate-spacemodelhasthefollowingform ξξξ(t)=Aξξξ(t−1)+ηηη(t) x(t)=KKKTξξξ(t) (22) Weleavethetechnicaldetailsofthis derivationto AppendixB. Thetermsξξξ(t),A,ηηη(t) and KKK are precisely defined in (36). The real version of (22), when the processξξξ has imaginary components is obtained by multiplying both equations by a block-diagonal matrix C (which is defined precisely in the Appendix), giving us the real state-space model Cξξξ(t)=(CAC−1)(Cξξξ(t−1))+Cηηη(t), (23) x(t)=(KKKTC−1)(Cξξξ(t)). (24) 5. The OU(ppp) as an ARMA(ppp, ppp−−−111) Thestudiesofpropertiesoflineartransformationsandaggregationsofsimilarprocesses haveproducedagreatamountofworkstemmingfromtheseminalpaperbyGrangerand Morris (1976) on the invariance of MA and ARMA processes under these operations. These results and extensions to vector autoregressive moving average (VARMA) pro- cessesarecompiledinthetextbookbyLu¨tkepohl(2005). The description of the OU(p) process x = OU (Λ) with parameters κκκ as a linear κκκ state-space model, given in the previous section, will allow us to show that the series x(0), x(1), ..., x(n) satisfies an ARMA(p,q) model with q smaller than p. We refer the reader to (Lu¨tkepohl, 2005, Ch. 11) for a presentation on VARMA processes and, in particular, to the following result on the invariance property of VARMA processes underlineartransformations,whichwequotewithaminorchangeofnotation:
Description: