应用概率统计第二十六卷 第四期2010年8月 Chinese Journal of Applied Probability and Statistics Vo1.26 No.4 Aug.2010 Moment Estimation of Parameters for Discretely Sampled OU—compound Poisson Processes ZHANG SHIBIN (Depnnment of Mathem。tics,Shnn 。i Maritime University,Sh0nghni,2013061 ZHANG XINSHENG (Depnnment of Stntistics,School of Manngement,Fud。n University,Shnnghni,200433) Abstract Processes of Ornstein—Uhlenbeck type,driven by positive compound Poisson processes,are considered in this paper.We are interested in parametric estimation of those processes based on discrete observations.The parameter of the stationary distribution is estimated by the method of moments,and a consistent and asymptotically normal estimator is provided.The theoretical study is also generalized to the superposition case. Keywords:Process of Ornstein-Uhlenbeck type,compound Poisson process,martingale estimating function,simple estimating function. AMS Subject Classiifcation: 62M05 60Gl0. §1. Introduction Recently,processes of Ornstein—Uhlenbeck type are widely used in describing connec- tions between branching processes and L6vy processes[ ,stochastic volatilities of ifnance assets[2]and default intensities[引.The c1ass of Gamma-OU processes i8 a well known subclass of processes of Ornstein-Uhlenbeck type.And OU—compound Poisson(OU—CP) processes axe generalization of Gamma-OU processes.The OU—CP process x(t)is given by the solution to the stochastic diferentia1 equation dX(t)=一 x(t)dt+dZ(At), where >0 is a positive parameter,and the background driving L4vy process[ 1(BDLP) z(t)is a positive compound Poisson process.In detail,z(t)has the representation N(t) z(t)=∑ , =1 The research is supported by the National Natural Science Foundation of China(10901100)and the Science& Technology Program of Shanghai Maritime University(20100135). Received September 3,2007. 第四期 张世斌张新生:离散抽样OU一复合Poisson过程的参数矩估计 385 where.(Ⅳ(t),t 0)is a Poisson process of intensity ,and Y1,Y2,…is a sequence of in— dependent identically-distributed positive random variables having a common distribution function G(x, ).Here is an unknown parameter.Without loss of generality and for simplicity, is assumed to be a scalar.If (i=1,2,…)is a sequence of i.i.d.and each follows the exponential distribution,an OU—cP process is just a Gamma-OU process. There are many literatures on parametric inferences of discretely sampled L6vy driving OU processes.Among main references,there are quite a few aiming at Gamma-OU pro- cesses.To estimate the parameter of Gamma-OU stochastic volatility processes,Markov chain Monte Carlo(MCMC)methods are exploited in[4,5】.The maximum likelihood method to estimate the parameter in the marginal distribution of Gamma-OU processes is used in[6].The maximum likelihood method is also employed in[7,8】.However,asymp- totic properties of the estimator in[7,8]are not achieved by theoretical study.MCMC methods are often time—consuming.And likelihood methods are,in general,not practical since the transition density is intractable and,as a consequence,the likelihood function is unknown.In order to substitute the true score function of parametric inferences of dis・ cretely sampled difusion models,several papers(e.g.,[9,10])study martingale estimating functions.Nevertheless,martingale estimating functions usually involve some integrals of the transition density,these integrals and their numerical techniques are usually hard to hand1e with. The method in this paper uses a kind of simple estimating functions to estimate the parameter in the marginal distribution of our models.Our method extends simple estimating functions[11】for difusions to those ofr OU—CP processes.For OU—cP processes. the density of the marginal distribution and numerical approximation of the transition function are usually diifcult to obtain.After finding moment connections between the marginal distribution of the OU—CP process x(t)and the distribution of each component jump (i=1,2,…)of the BDLP z(t),we construct a simple estimating function. Because of moments of the marginal distribution being employed,the estimating method is referred to sa moment estimation.For processes of which neither the marginal distribution nor the transition distribution has a closed—form,parametric estimation can also employ this method. §2. Moment Connections The L6vy measure of Z(1)is W(dx)= z>0)dG( ; ).To be sure that there exists an invariant measure of x(t)E 引,G( ;Q)should be sasumed to satisfy 应用概率统计 第二十六卷 Assumption 2.1 logXdG(x;n1<。。. The subclass of OU—CP processes can be constructed by diferent choices of the r.v. in BDLPs.Here are two examples. Example 1 The OU-CP process with Weibull jumps in the BDLP.The L6vy dens Y。f z(1)is叫( ):Zg(x;c,Q),where 9( ;c, )=ac(ax)c-1 exp(一( )。) 。>0)is the density function of ( =1,2,…),and ,Ol,c>0.When c=1,the OU—CP process with Weibull jumps in the BDLP is just the Gamma-OU process with the marginal r(z,Q) distribution. Example 2 The OUICP process with Pareto jumps in the BDLP.The L6vy den— sity of Z(1)is ( )=13g(x;a),where 9( ; )= (1+ )- ̄-1I{ >o)is the density function of (i=1,2,…),and ,Ol>0. 0 l _∞.0 0.0 寸0 o By the structure of the BDLP ( )and the temporal homogeneity of processes (t) and z( ),simulations can be implemented.For the ifxed parameter,simulated paths of (t)and z( )in Example 1 are given in Figure 1.Simulations can also give some visual impressions of the stationary distribution of the OU—CP process;however,the stationary distribution is actually diifcult to be represented by a closed—form. (a)Simulations of z( n△)against n(△:1) (b)Simulations of x(nA)against (△=1) 0.0 0.5 1.0 1.5 (C)The empirical density of x(t)marginal Figure 1 The OU—CP process with Weibull jumps in the BDLP( =0.1, =3, Ot=6,C=2) 第四期 张世斌张新生:离散抽样Ou一复合Poisson过程的参数矩估计 387 Under Assumption 2.1,given that x(0)=X,x(t)can be represented to be x(t)lx(0): =d e-At +E ,N =1 where c:d,denotes equality in distribution’r.v. has a Poisson distribution of intensity ,●●●●●●●,、●●●●●_-/ 0 , ,and , ,…is a sequence of independent identically—dibuted random variables 一 stri一 ,< e e having a common density function ∞∑ + , ; : 。) (2.2) The representation(2.1)is given in[12],and the similar results for some other subclssa of processes of OU type,with their proofs can be seen in[13,14].The temporally homoge- neous transition function of the OU—CP process x(t)is Y<e-Atx, P(t,X, ; , , )= (Aflt)ne-; ̄Zt 丁胁 , Y=e-Atx, Y>e-) ̄tx, where P(t,X, ; , , )is the transition function of (・)over a time interval of length t, i.e.,P( (・+t) ylX(・)= ),and.厂礼x)(佗 1)are defined by 11( )=.厂( ), 厶( )= ‘厂( )厶一 ( 一 )d That is,fn( )is the convolution of f(x)and厶一1( ),where f(x)is defined in(2.2). Firstly,we try to find connections between moments of (・)and .If EQ ]<oo, then E,xtQ,[ ]= = G(e'Xtw;a) --G(w;a)d —: At o。。 1一e- d 蜥 E0 ] dGc ; : d = AtSimilarly,we have 1——e-2At E ,Q【 ]= [ 。] 2At E ] 。] if EQ ]<∞ 1——e-3At E ,。3At if E口 3】<∞; and E她。[ 4]= 1——e-4M 4At E。 4] if E 4】<oo 应用概率统计 第二十六卷 Assumption 2.2 Ea 4]<∞ By【12],the characteristic function of the OU—CP process x(t)is xp(/0。。(eizx-1)坠 d ) Under Assumption 2.2,liar X4(1一G( ; ))=0. After calculating妒 (z), ( ), (3)( ) X--- ̄(X) and (4)( ),by integration by parts,we obtain ,a (・)]={妒,(0)= E [ 】, 口【( 】= ,(0)= 。(E IN) +鲁 ], E ,a【( (.))3j=一{ (3)(0)= 3(EQIN)3+ 3 EQ 】EQ 2】+鲁EQ 3】, E ,a[( (・))4]= (4)(0)= 4(EaIN) +3 3(E0[ 】) E口【 】+ 3 。(Ea【 2】)。 4 E+ 。[ 】Ea【 。】+ E 【 4】. Under Assumption 2.2,moment connections among r-v. , and (・)call be induced in,IIable 1. T ble 1 Moment cOnnectiOns r.V. (・) Expectation E 】 At —LEa[l L 一’】J E0 】 2nd moment Ea 】 —1-2eA2At —t EQu JIF. 】 鲁E [ 。]+ (E口[ 】) 3rd moment E 。] —l_e3At-  ̄At —u JIV.。] 。(E 【 ])3+ 3 Ea[ 】Ea[ ]+鲁Ea【 3】 4( [Y-I)4+3 。(已 ]) E 】 -4th moment E 4] —1-e4At 4At —E I—u F.4]J + 。(E [ 】) + 4 Ea[ 】EQ[ 3】 +釉 4】 Proposition 2.1 Under Assumption 2.2.we have connections between central moments of (・)and moments of as , 【( (・)一EB, Ix(・)】) ] 釉蛸, (2.3) ,a[( (・)一 , IX(・) 。】, (2.4) , [ (・)一 。IX(・)】)4】 )2+ (2.5) 第四期 张世斌张新生:离散抽样OU.复合Poisson过程的参数矩估计 Connections between conditional mome ̄s and stationary moments hold: , ,d[x(t)lx(o)= 】=e-- ̄tx+(1一e-Xt)Ep,aIx(・)】, ,(2.6) (2.7) Varx, 。 ( )l (0)= 】=(1一e-2At)Mar,a (・)】. Proof Equalities(2.3)一(2.5)can be obtained by Table 1 and some mathematical calculations.From equalities(2.1)and(2.3),Table 1 and the Wald’S identity,we have equalities(2.6)and(2.7). 口 §3. Moment Estimation Assume we observe the process at dates{tk= △,k:0,1,…,礼},where A>0 is ifxed.For clarity,1et =X(ka)(k=0,1,…,佗),then the sample is(x0,X1,…, ). Let hi(x; ,oL)j x-EZ, ̄[ .],h2(x; ,oL)=( —E ,Q[ .】)。-Varfl,c ̄IX.】,and h( ; ,O/) :(hi(x; , ),h2(x; , )) .Let rot(dx)denote the invariant measure of the Ou—CP process (t).For any,∈L2( ,n)and any入>0,the transition operator on L ( 口,a)is defined by ,oo Ⅱ 'Q,( ; , ):/ ,( ; ,a)P(A,z,d ;入, , ). 一∞ Lemma 3.1 Under Assumption 2.1 and 2.2.we have Ⅱ ’ hffL。( 芦, )≤e—A△IlhllL ( , ). (3.1) Proof Under Assumption 2.1 and 2.2,it follows from(2.6)and(2.7)that hi(y; ,a)p(A, ,dy; , ,O/) :e-AA( — Ea[Y-1)=e- ̄ah1( ; , ), ,o。 /h2(y; ,Q)P(△, ,曲;入, , ) —O0 =Mar,x, , [xx IXo= 】+e-2XA ;( ; , )一Mar,aIx.] =e-2AA(( —E卢,aIX.】) 一Marfl,。[ .】)=e-2 ̄Ah2(x;fl, ). Thus,J JⅡ ' h1}f ( 卢,。) e— △ff I ff£z( 卢, )and ffⅡ ,Qh2ff , ) e— △ff 2ff ( 口, ). This proves the lemma. 口 We define the probability measure Q ,a on as the joint distribution of(x0,x1) in the stationary state,i.e.,Q p,a(d ,d )= , (dx)P(A, ,dy;A, , ).We denote by c_。 the potential operator defined by , , , ah(x)=∑ , ,口【h( ; , ̄)lX0= 】_In k=0 the following,we shall denote the true parameter value by ( o,Z0,s0). 应用概率统计 第二十六卷 Assumption 3.1 E。IV]and E IV.。]are all twice continuously diferentiable with respect to Q.And d(EQ[ ])E [ 】≠,E [ 】 (E [ 。】) d(EQ[ ])E [ ]≠Ea[ ] (EaIV.。】) Oh1.c Ac , ,=rdx ̄ 羹 : 三 ;):一 E4V . ]2 妻 ; ).c3.2 (( , ) 一(阮, 。) ) d N(0,V占), with v =A( , 0)_。VAh3o A r(Zo, 0)一1 (3.4) and VAh=/(h( )+uh(y)一Uh( ))(h( )+Uh(y)一Uh( )) Qa(d ,dy)・ (3・5) For clarity we have suppressed the argument(Ao, , 0). , Remark 2 An estimator obtained by solving equation(3.3)is equivalent to that obtained by solving Xk——1 ,。( .))=0, (3.6) 磺一 E口, ( .。))=0 身 期 张世斌张新生:离散抽样ou一 ̄#Poisson过程的参数矩估计 391 Under Assumption 2.1,2.2 and 3.1,the asymptotic variance vo V占= 1(fo,se A , 。 (Yn Ao, l\ (入。,,Q。o; where A( , )is defined by(3.2)and det(A(lf, ))= 鲁( [ ] (E。[ 】)一 d(E [ ])E [ 】), Ⅵl( , ,O1)= 1f3(1+e…)2EⅡ 。]( (E ))2 1f3 一 一t- 十e_-2AA  ̄ d/E ]) (E IF. ])Ea[ 3] +1f3 e_2Ax/)( ̄(E [ 】)) ( (EQIV.。】)。+三EnfY.4】), vI2(A, , )= 1f2 一。e-Ax/) (Ea[ 】) 未(E 。】) 1f 2 +e_2A/X)d ( IV. ]) 3] Q1f,, 2 /● 2 2 一一_0< 一< 0 e-2AA)E [ ] d(E [ 】)( (E [ 】)。+三E。[ 4]), 2(a, ,a), 、l,0 、l,O 、、l,鲁( 1+e-A/x) (E Iv.2】)。一鲁(1+e-Aa+e-2Aa)EQ 2 IV.3】 ++e-2Ax/)(EaIv.1) ( (E [ 】) + EQ[ 4】). Proof By the proof of Lemma 3.1.we have (h(x)+Uh(y)一 h( )) =( ( )_e-Aahl( ,南( 3,)_e-2AX/ ( ). /( ( )-e-AAhl( ))。Q , (d ,d ) =(1+e-2Ax/)var跏 】_2e-Ax// 跏( / ( , , ) =(1-t-e-2A/x)Va n 】-2e-2A/Xvar ,口 】=(1~e一 A)Varlf,aIx.]. Similarly,using(2.6)and(2.7),we also obtain h1( )_e-AA 1(z))( 2( )__e-2AA 2(z))Q a( 曲):(I-e-3A/x) n .)】 应用概率统计 第二十六卷 and /.( z( )_e-2 ̄,Ah2( )) Q (dx,dy)=(1-e-4XA)( 碍(x-)]一(Var#,a .】) ). Hence,VAh deifned by(3.5)can be represented by V△h= 1一e一2AA B( , , ) -'x e-2"XA)E[ h3( X.)] = (1+eA+】(1 △)[㈣】_(]). For clarity we have suppressed the argument( , ).Since we have(2.4),(2.5),(3.2)and (3.4),the result is established. 口 Remark 3 If there is no unknown parameter OL in the distribution of , is the only unknown parameter in the stationary distribution of the OU—CP process.To get the 一estimator of ,we can only exploit the simple estimating function hl(培 x)=X—E口 .】糕 . Thus.it is enough for asymptotic properties in Theorem 3.1 to repllJ ace Assumption 2.2 and 3.1 with ElY. ]<o。.Consequently, 卜 Vhhl-篝 Varzo IX.]and Example 3 (i)In Example 1,when follows a Weibull distribution and C is known,we have E0 】: 1 ,上 J,1、 and E 。】= 2 r/2、tc)’ Hence,we obtain = where = n k --1甄 : 叠(地一 一 ) (3.7) Especially,when follows an exponential distribution.we have m2=A /o 2= , / (ii)In Example 2,when follows a Pareto distribution,we have E0 ]:i/(a一1) if >1 and EQIF. 】=2/【( —1)(Q一2)】if Q>2.Hence,if OZ>2,we obtain 一=X竹(1+ / ), =2+ / , where—Xn and are defined as in(3.7). 一 第四期 张世斌张新生:离散抽样OU-复合Poisson过程的参数矩估计 393 Theorem 3.1 gives US a method to get the estimate of parameter( , ).Table 2 reports realizations of the estimator(ifn, )in Example 3(ii).For each case,we have all implemented 400 times.In the simulation,we change the sampling time interval A from 1 to 3.And for each value of A,we change the size of sample n from 500 to 1500. Table 2 Means and Standard deviations ofr 400 realizations of(ifn, ) for the OU—CP process with Pareto jumps in the BDLP(True parameter values are =0.5, =2,OL=8) 一 8n ^ (In △ n Mean S.D. Mcan S.D. 1 500 2.055 0.219 8.186 0.780 i000 2.005 0.162 8.017 0.617 1500 2.015 0.138 8.072 0.515 1.5 500 2.034 0.190 8.130 0.713 1000 2.016 0.140 8.062 0.546 1500 2.010 0.115 8.031 0.445 3 500 2.035 0.178 8.122 0.657 1000 2.029 0.118 8.110 0.453 1500 2.010 0.105 8.048 0.395 Remark 4 The parameter of the OU—CP process can be estimated by the es— timat。r n=(1/a) 1max< <{ln Xk.1一In Xk}as in[6】,or the estimat。r derived by the nempirical autocorrelation function sa in[7]. §4. Generalization to the Superposition of oU—CP Processes 上’he relative simplicity of the clsas of models outlined above implies some limitations. Combining independent Ornstein—Uhlenbeck processes with diferent rate parameters A is proposed to capture more realistic dependence structures in[2].The empirical findings in[15】suggest that the superposition of 2 independent processes with diferent rate is adequate for modeling daily financial data,SO we shall only consider this case in this paper.A superposition of processes Xl(t),X2(t)defines a new process x(t)through x(t)=Xl(t)+x2(t), 394 应用概率统计 第二十六卷 where each component process ( )(J=1,2)is defined by(1.1)with its own speciifc and driving process zj(t)while the latter are independent across components. Here that the process is observed at dates{tk=kA,k=0,1,…,礼)is still as— sumed,where△>0 is fixed.Let Xk= (尼△),( =0,1,…,礼),then the sample is ( , 1,…, ).Meanwhile,let Xlk=XI(kA),X2k:X2(ka),and realizations of samples( 1o,X11,…,X1n)and(X20,X21,…,X2n)are assumed to be seen.Factually, these realizations can not be seen in practice.And that we assume those is only the tech- nical means to get our results.In terms of the proof of Lemma 3.1 and(3.6),the estimator ,●●●●J ●●●一-, of the parameter in E[X1.]and E[X1. ]can be obtained by solving ∑ ∑ {l E[ Xlk一1一F[X1.]]=0, 2 X l E[(Xx ̄k一1一E[X1. ])+2E[X2.](xlk一1一E[X1.】)]=0. (4:2) Similarly】the estimator of the parameter in E[X2.】and E[X2.。]can be obtained by solving J =1 l E[X2k一1一E[X2.】]=0, =1 1L [南( 一l—E[X2.。])+2Xl 一1( 2 一1一E[X2.】)]=0. (4.3) By summing up corresponding parts of positions in(4.2)and(4.3),we reach the conclusion that the estimator of the parameter in E[X.】and E[X. 】can be obtained by solving 1一E[X.])=0, (4.4) 1一E[X.2])=0 Assumption 4.1 x(t)is defined by(4.1).The component processes are Xl(t) and x2(t).each of them satisifes Assumption 2.1,2.2 and 3.1.xj(£)has the BDLP (£),J:1,2. (£), :,. :( ):Ⅳ where 川,, , i.i.d. random variables( )= ∑ u , 川,…are ・・ V叭ab s, ,( 1,2.{ ( ))and{ ‘ ))are independent of each other and over i.They all have the cdf G( , ), >0.{Ⅳ(J)( ),t 0)is a Poisson process of intensity ,J=1,2・More0ver, {Ⅳ(1)(t),t 0)and{Ⅳ【 )(£),t 0)are independent. Since we haⅣe Theorem 3.1 and above discussions,the theorem for the superposition case is as follows if Assumption 4.1 is true. The0rem 4.1 Under Assumption 4.1,if we let = 1+ 2,then an estimator ( , ),which solves(4.4),exists with a probability tending to one as n 。O・Mor 第四期 张世斌张新生:离散抽样OU一复合Poisson过程的参数矩估计 395 oVer,as几__+∞,( ,an)-_÷(Z0, 0)in probability and(ifn,an)also has the asymptotic normality. Example 4 If the marginal distribution of (t)is r(岛, )distribution = then estimators of 8=fll+82 and仅are defined by -2:A /o 2, = / , where—Xn and are deifned sa in(3.7). Some simulation results for the process in Example 4 can be seen in Table 3,where lfl: =3:1 is assumed to be known.For each case,we have all implemented 400 times. In the simulation,for the fixed sampling time interval A=1,we change the size of sample n from 500 to 1500. Table 3 Means and Standard deviations ofr 400 realizations of(lfl礼, 2佗,an) ofr the superposition of Gamma-OU processes(True parameter values are AI=0.2, 2=1.2,fll=3, =1 and OL=6) 一 一 一 1 Oln n Mean S.D. Mean S.D. Mean S.D. 500 2.985 0.434 0.995 0.145 6.744 1.019 1000 2.913 0.313 0.971 0.104 6.561 0.730 1500 2.921 0.256 0.974 0.085 6.568 0.618 §5. When There Is No Unknown Parameter OL Under Assumption 2.1.if there is no unknown parameter OL in the distribution of and ElY. ]<∞,then( , )is the unknown parameter of the OU—CP process.By[10], the optimal linear estimating equation ofr estimating( , )is 一0. (5.1) where k=(kl,k2) ,and 1(A, , ; , )= F(A, ; , ) (△, ; , ) [Y—F(A, ; , )], k2(△, , ; , )= F(A, ; , ) 咖(△,z; , ) [Y—F(A, ; , )], 应用概率统计 第二十六卷 with F(A, ; , )=E , ̄[XllXo=X]and (△, ; , )=VarA, ̄[XllXo= 】.Let( n, ) be the estimator defined by n ,n 一 1 1 f I n∑Xk一1 一(∑ 1 、 =1 =甄1 C ,< n n∑X 2—1 一=1 ( ∑ 礼 =1 = /,,.,...。。(5.2) \ ∑( 一e =1 = h△ Jc一1) n(x—e-A, ̄A)E LY.] △ △ n_0(Then the estimator( ,3n)Solves(5.1)with a probability tending t3. 0 o one as8 Q Q Moreover,as n-_÷。o,(An, ) ( o, )in probability and d d (( , ) 一( 0, ) )一d N(0,w( 0, )), where w( 0, )=c( o,90)一 w△k 0, c (A0,Z0)一 ,and 一= 帆一厂/ 8Q Q △ △%一 ), w△ =/k(△,甄_lj凰; , )k(△, _1J ) Q (dx,dy) \、一、、一 /-一/一l aiag( △,(1-e-'xA) dx,dy dx,dy = diag(△2e一2AA(1--e-AA) ) (5.3) 一Actually, w( , ):diag( ( 一1),1一-I-e -'x ̄A△,32(E[Y . 2]. Proposition 5.1 Under Assumption 2.1.if there is no unknown parameter Q in the distributi。n。f and ElY. 】<。。,the estimat。r by s。lving∑n hi(xk1):0 and the estimator by solving(5.1)have the same symptaotic variance 1 q-e一 o△ Ely. ] 1一e-A0△2(E ]) Proof Since we have Remark 3 and equality(5.3),the result is established. Example 5 If has the pdf 口 g( )= 。 < )+ 1 1{ < )+ 。<4), 第四期 张世斌张新生:离散抽样OU一复合Poisson过程的参数矩估计 397 then ElY.】=17/24+1/(2 In 2)and ElY. 】=5/6+3/(2 In 2).The estimator :Xn/E[Y.], where Xn is defined as in(3.7).And the estimator is defined by(5.2). Table 4 is implementations of estimators 8n and 8n in Example 5.For each case,we have all implemented 400 times.In the simulation,we change the sampling time interval △from 1 to 2.And for each value of A.we change the size of sample n from 200 to 1000. From Table 4,we can see that the diference of standard deviations between 8n and 8n is negligible.which is consistent with Proposition 5.1. Table 4 Means and Standard deviations for 400 realizations of 8n and 8n for the OU—CP process in Example 5(True parameter values are =0.2 and =3) ^ ~ △ 礼 Mean S.D. Mean S.D. 1 200 2.998 0.329 2.999 0.329 500 2.990 0.206 2.989 0.208 1000 3.012 0.161 3.012 0.161 2 200 2.997 0.235 2.998 0.234 500 3.002 0.155 3.003 0.156 1000 3.001 0.105 3.001 0.105 Remark 5 Although asymptotic variances of estimators 8n and 8n are same{8n has the simpler form,and is easy to calculate. References 【1]Li,Z.H.,Ornstein-Uhlenbeck type processes and branching processes with immigration,Journal of Applied Probability,3r(2ooo),627-634. 【2l Barndorff-Nielsen,O.E.and Shephard,N.,Non-Gaussian 0rnstein U hlenbeck-based models and some of their uses in ifnancial economics(with discussion), R.Statist.Soc.B,63(2001),167-241. [3】Caxiboni,C.and Schoutens,W.,Jumps in intensity models:investigating the performance of Ornstein-Uhlenbeck processes in credit risk modeling,Metrika,69(2009),173—198. 【4】Grififn,J.E.and Steel,M.F.J.,Inference with Non.Gaussian Ornstein-Uhlenbeck processes for stochastic volatilityj lsnral of Econometrics,134(2006),605-644. 【5】Roberts,G.O.,Papaspiliopoulos,O.and Dellaportas,P.,Bayesian inference for non.Gaussian Ornstein-Uhtenbeck stochastic volatility processes,I,_R.Statist.Soc.B,66(2)(20oa),369—393. 【6】Zhang,s.B.,Zhang,X.S.and Sun s.G.,Parametric estimation of discretly sampled Gamma-OU processes,Science in China.Series A,49(9)(2006),1231~1257. 应用概率统计 第二十六卷 U M Valdivieso,L.,Parameter estimation for an IG・OU stochastic volatility model,Proceedings of 3rd Actuarial and Financial Mathematics Day,Vanmaele,M.et a1.,Koninkl ̄ke Vlaamse Acadamie van Belgie voor Wetenschappen en Kunsten,2005,95—104. Valdivieso,L.,Schoutens,W.and ̄erlinckx,F.,Maximum likelihood estimation in processes of Ornstein-Uhlenbeck type,Statistical Inference for Stochastic Processes,12(2009),1-19. Bibby,B.M.and Scrensen,M.,Martingale estimation functions for discretely observed difusion processes,Bernoulli,1(1995),17-39. Scrensen,M.,Estimating functions for discretely observed difusions:a review,Basawa,I.V.,Go- dambe,V.P.,Taylor,R.L.,Selected Proceedings of the Symposium on Estimating Functions,IMS Lecture Notes—Monograph Series,32,1997,305—325. Kessler,M.,Simple and explicit estimating functions for a discretely observed difusion process, Scand. Statist.,2 ̄'(2ooo),65—82. Zhang,S。B.and Zhang,X.S.,Transition laws of OU・compound Poisson processes,submitted. Zhang,S.B.and Zhang,X.S.,Exact simulation of IG—OU processes,Methodology and Computing in Applide Probability,10(3)(2008),337—355. Zhang,S.B.and Zhang,X.S.,On the transition law of tempered stable Ornstein—Uhlenbeck processes, Journal of Applied Probability,46(3)(2009),721—731. Barndorff-Nielsen,O.E.and Shephard,N.,Econometric analysis of realized volatility and its use in estimating stochastic volatility mode ̄,J.R.Statist.Soc.B(Statistical Methodology),64(2002), 253-280. 离散抽样OU.复合Poisson过程的参数矩估计 张世斌 张新生 (上海海事大学数学系,上海,201306) (复旦大学管理学院统计学系,上海,200433) 本文研究基于离散观测的正复合Poisson过程驱动OU型过程的参数估计.通过矩估计给出了过程平稳分 布参数的估计量,并得到了估计量的相合性和渐近正态性.进一步,将矩估计的方法和结论推广到叠加过程的 情况. 关键词:OU型过程,复合Poisson过程,鞅估计函数,简单估计函数. 学科分类号:O211.64.