1. Introduction In recent years, many scholars have found that some financial time series data tend to be shown as biased random walk, long memory, and self-similarity, etc., which made the stochastic differential equation model driven by Brownian motion no longer applicable to describe financial data. Perhaps the most popular approach for modeling long memory is the use of fractional Brownian motion (hereafter fBm) that has been verified as a good model to describe the long memory property of some time series.
Compared with the traditional efficient market hypothesis theory, fractional market theory can accurately depict the actual law of financial market, such as the Ornstein-Uhlenbeck process driven by fractional Brownian motion, which is more consistent with the characteristics of long-term memory, in place of Vasicek model that is suitable to simulate the short-term interest rate model.
Although the study of fractional Brownian motion has been going on for decades, statistical inference problems related are just in its infancy. Such questions have been recently treated in several papers [1–3]: in general, the techniques used to construct maximum likelihood estimators (MLE) for the drift parameter are based on Girsanov transforms for fBm and depend on the properties of the deterministic fractional operators related to the fBm. Generally speaking, these papers focused on the problems of estimating the unknown parameters in the continuous-time case. Prakasa Rao [4] gave an extensive review on most of the recent developments related to the parametric and other inference procedures for stochastic models driven by fBm. The latest study can be found in Xiao and Yu [5, 6], who developed the asymptotic theory for least square estimators for two parameters in the drift function in the fractional Vasicek model with a continuous record of observations. Another possibility is to use Euler-type approximations for the solution of the above equation and to construct an MLE estimator based on the density of the observations given "the past", for the case of stochastic equations driven by Brownian motion. “Real-world” data is, however, typically discretely sampled (e.g., stock prices collected once a day or, at best, at every tick). Therefore, statistical inference for discretely observed diffusions is of great interest for practical purposes and at the same time it poses a challenging problem. Some papers are devoted to the parameter estimation for the models with fBm and discrete observations; see, e.g., Hu and Nualart[1], Hu and Song [7], Mishura and Ralchenko [8], Zhang, Xiao, Zhang and Niu [9], and Sun and Shi [10].
In this paper, we shall consider the parameter estimation problem for fractional linear diffusion process (FLDP). Assume that we have the model(1)dXt=α-βXtdt+σdBtH,which can describe the intrinsic characteristics of interest rate more accurately in practical problem. The drift parameter α, β can characterize, respectively, the long-term equilibrium interest rate level and the rate of the short-term interest rates deviate from long-term interest rates. In general, the parameters of long-term equilibrium level of short-term interest rate are unknown. We assume β>0 throughout the paper so that the process is ergodic (when β<0 the solution to (1) will diverge), σ describes the volatility of interest rates, and (BtH)t≥0 is a fBm with Hurst parameter H∈(0,1). In this paper, we suppose the Hurst index H, the diffusion coefficients α,β, and the volatility σ are unknown parameters to be estimated. We will furthermore show the strong consistence of these estimators.
In the case of diffusion process driven by Brownian motion, the most important methods are either maximum likelihood estimation or least square estimation. Since fBm is not a Markov process, the Kalman filter method cannot be applied to estimate the parameters of stochastic process driven by fBm. Consequently, it is a convenient way to handle the estimation problem by replacing fBm with its associated disturbed random walk. In this paper, we follow Zhang et. al. [9] to use discrete expressions of fractional Bronwnian motion with Donsker type approximate formula, which can, to some extent, simplify calculation and simulation. Although we do not have martingales in the model, this construction involving random walks allows using martingales arguments to obtain the asymptotic behaviour of the estimators.
Our paper is organized as follows. In Section 2, we propose MLE estimators for FLDP from discrete observations. The almost sure convergence of the estimators is provided in the latter part of this section. In Section 4, an extension for generalized fractional diffusion process is briefly discussed. Finally, Section 5 includes conclusions and directions of further work.
2. Estimation Procedure It is worth emphasizing that the solution of (1) is given by (2)Xt=X0+∫0tα-βXsds+σBtH,where the unknown parameters included α,β,σ and H. We now proceed to estimate these parameters based on quadratic variation method and maximum likelihood approach.
Let {Xt,t∈R} be the FLDP with H>1/2 and suppose that πn={τkn,k=0,1,…,in},n≥1,in↑∞, be a sequence of partitions of the interval [0,T]. If partition πn is uniform, then τkn=kT/in for all k∈{0,1,…,in}. If in≡n, we write tkn instead of τkn. Assume that process Xt is observed at time points i/mnT,i=1,2,…,mn, where mn=nkn and kn grows faster than nlnn, but the growth does not exceed polynomial, e.g., kn=nlnθn,θ>1 or kn=n2.
In applications, the estimation of H∈(0,1) (called the Hurst index) is a fundamental problem. Its solution depends on the theoretical structure of a model under consideration. Therefore particular models usually deserve separate analysis.
According to the notation of Kubilius Skorniakov [11], suppose there are two hypotheses:(3)C1ΔXτkn=Xτkn-Xτk-1n=OωdnH-ε,C2Δ2Xτkn=ΔXτkn-ΔXτk-1n=σΔ2Bτk-1nH+Oωdn2H-ε,for all ε∈(0,H-1/2), where dn=max1≤k≤mn(τkn-τk-1n)Yn=Oω(an) means for a sequence of r.v. Yn, and an⊂(0,∞), and there exists a.s. non-negative r.v. ς, such that Yn≤ς·an. These two conditions are used to prove the strongly consistent and asymptotically normality of the estimator H from discrete observatios.
Denote (4)Wn,k=∑i=-kn+2knΔ2Xsin+tkn2=∑i=-kn+2knXsin+tkn-2Xsi-1n+tkn+Xsi-2n+tkn2,where 1≤k≤n and sin=i/mnT,
Then, the estimator of Hurst parameter H can be written as (5)H^=12+12lnknln2mn∑k=2mnΔ2Xtkn2Wn,k-1.
Next, we turn to the estimation problem of the diffusion parameter σ2. When H is known, Xiao et al. [12] obtained the estimators based on approximating integrals via Riemann sums with Hurst index H∈(1/2,3/4). In contrast, we suppose in this paper the Hurst index is unknown. Therefore in the next estimation, the estimator of H will be embedded in the equation. For simplicity, denote Xτin≜Xih,Bτk-1nH=BiH,mn,i=1,…,mn,h=T/mn. Thus, the full sequence of mn observations can be written as {Xh,X2h,…,Xmnh}.
For the diffusion parameter, we easily obtain an estimator for the diffusion parameter by using quadratic variations, such (6)σ^2=∑i=1mn-1Xi+1h-Xih2mn-1h2H^,which converges (in L2 and almost surely) to σ2.
Finally, we are in a position to estimate the drift parameter. Note that BtH-Bt-1H is not independent and the process BtH is not a semimartingale; therefore the martingale type techniques cannot be used to study this estimator. This problem will be avoided by the use of the random walks that approximate BtH. Based on the results on Sottinen [13], the fractional Brownian motion can be approximated by a "disturbed" random walk, which was called Donsker type approximation for fBm.
Lemma 1. The fBm with Hurst parameter H>1/2 can be represented by its associated disturbed random walk: (7)BtH,mn=∑i=1mntmn∫i-1/mni/mnKHmntmn,sdsεi,with KH(t,s)=cH(H-1/2)s1/2-H∫st(u-s)H-1/2uH-1/2du, which is the kernel function that transforms the standard Brownian motion into a fractional one, cH is the normalizing constant cH=2HΓ(2/3-H)/Γ(H+1/2)Γ(2-2H), and εi are i.i.d. random variables with Eεi=0 and varεi=1, and ⌊x⌋ denotes the greatest integer not exceeding x.
Sottinen (2011) proved that BtH,mn converges weakly in the skorohod topology to the fractional Brownian motion. With the estimators H^,σ^ plug-in, the replacing model still kept the main properties of the original process, such as long range dependence and asymptotic self-similar. Therefore, the martingales can be used to treat this replacing model.
In general, numerical approximation of model (1) can be presented by Euler scheme:(8)Xi+1h=Xih+α-βXihh+σ^Bi+1hH,mn-BihH,mn, i=1,…,mn-1.
Set(9)fiε1,ε2,…,εi=mn∑j=1i∫j-1hjhKHi+1h,s-KHih,sdsεjto denote the contribution of the n-1 first jumps of the random walk and(10)Fi=mn∫ihi+1hKHi+1h,sdsto denote the contribution of the last jump.
With the approximation of fBm (Lemma 1), we can write (11)Bi+1hH,mn-BihH,mn=fiε1,ε2,…,εi+Fiεi+1, i=1,…,mn-1.with which (8) can be written as (12)Xi+1h=Xih+α-βXihh+σ^fiε1,ε2,…,εi+Fiεi+1.Hence we have (13)EXi+1h-Xih∣Xih=α-βXihh+σ^fiε1,ε2,…,εi,varXi+1h-Xih∣Xih=σ^2Fi2.
We assume that random variables εi follow a standard normal law N(0,1). Then, the random variable X(i+1)h is conditionally Gaussian and the conditional density of X(i+1)h given Xh,X2h,…,Xih can be written as(14)fXi+1h∣Xh,X2h,…,Xihxi+1h∣xh,x2h,…,xih=12πσ^2Fi2exp-12xi+1h-xih-α-βxihh-σ^fiε1,ε2,…,εi2Fi2.The likelihood function can be expressed as(15)Lα,β=fXhxhfX2h∣Xhx2h∣xh⋯fXmnh∣Xh,X2h,…,Xmn-1hxmnh∣xh,x2h,…,xmn-1h=∏i=1mn12πσ^2Fi2exp-12xi+1h-xih-α-βxihh-σ^fiε1,ε2,…,εi2Fi2.
This leads to the MLE of α and β(16)α^=∑i=0mn-1yih-β^hxih/Fi2∑i=0N-1h/Fi2,(17)β^=1h∑i=0mn-1xih/Fi2∑i=0mn-1yih/Fi2-∑i=0mn-11/Fi2∑i=0mn-1yihxih/Fi2∑i=0mn-1xih2/Fi2∑i=0mn-11/Fi2-∑i=0mn-1xih/Fi22,
where yih=x(i+1)h-xih-σ^fi(ε1,ε2,…,εi)=(α-βXih)h+vi,vi=σ^Fiεi+1,i=1,2,…,mn-1.
Remark 2. Note that the parameter estimators of drift coefficients are related to the volatility σ, while, in fact, σ2 can be (at least theoretically) computed on any finite time interval. Furthermore, fBm is self-similar to stationary increments and it satisfies EBtH-BsH=t-s2H for every s,t∈[0,T]. For this reason, we may assume that the diffusion coefficient is equal to 1.
3. The Asymptotic Properties In this section, we turn to study the strong consistency of these estimators by (5), (6), (16), and (17).
Theorem 3. Assume that solution of (1) satisfies hypotheses (C1) and (C2), then estimator H^ converges to H almost surely as mn goes to infinity.
Detailed proof can be found in Kubilus and Skorniakov [11].
Theorem 4. The estimator σ^2 converges to σ2 almost surely as mn goes to infinity.
Proof. With the strong consistency of H^ to H and that σ~2≜∑i=1mn-1Xi+1h-Xih2/(mn-1)h2H→σ2 with probability 1 as mn goes to infinity, it can be easily shown that estimator σ^2 converges to σ2 almost surely as mn→∞.
Theorem 5. With probability one, α^→α,β^→β, as mn→∞.
Proof. Clearly, the consistency of α^ can be inferred combined with (16) and consistency of β^. We just prove that β^ is strong consistent.
A simple calculation shows that (18)β^-β=1h∑i=0mn-1vi/Fi2∑i=0mn-1xih/Fi2-∑i=0mn-11/Fi2∑i=0N-1vixih/Fi2∑i=0mn-1xih2/Fi2∑i=0mn-11/Fi2-∑i=0mn-1xih/Fi22=T/mn∑i=0mn-11/Fi2∑i=0mn-1vixih/Fi2-∑i=0mn-1vi/Fi2∑i=0mn-1xih/Fi2T/mn2∑i=0mn-1xih2/Fi2∑i=0mn-11/Fi2-∑i=0mn-1xih/Fi22=Mn<M>n,where Mn=T/mn∑i=0mn-11/Fi2∑i=0mn-1vihxih/Fi2-∑i=0mn-1yih/Fi2∑i=0mn-1xih/Fi2 is a square-integrable martingle and <M>n=(T/mn)2∑i=0mn-1xih2/Fi2∑i=0mn-11/Fi2-(∑i=0mn-1xih/Fi2)2 is quadratic characteristic of Mn.
Using the assumption of H∈(1/2,1) and fractional integral, we have the explicit solution of (1) that can be expressed as (19)Xt=1-e-βtαβ+σ∫0te-βt-sdBsH, t≥0,where the integral can be understood in the Skorohod sense.
As a consequence, for any i, we have
(20)
E
X
i
h
2
=
E
1
-
e
-
β
i
h
α
β
2
+
σ
∫
0
i
h
e
-
β
t
-
s
d
B
s
H
2
+
2
1
-
e
-
β
i
h
α
β
σ
∫
0
i
h
e
-
β
t
-
s
d
B
s
H
≤
2
1
-
e
-
β
i
h
α
β
2
+
σ
2
e
-
2
β
i
h
H
Γ
2
H
β
2
H
Hence, for any i, we obtain that E[Xih2] is bounded. Moreover, by using Cauchy-Schwartz inequality, we show that (see also in [14], with a slight modification below)(21)EBi+1hH,mn-BihH,mn2=E∑j=mnihmni+1hmn∫j-1/mnj/mnKHmni+1hmn,s-KHmnihmn,sdsεj2=∑j=mnihmni+1hmn∫j-1/mnj/mnKHmni+1hmn,s-KHmnihmn,sds2≤∑j=mnihmni+1hmn∫j-1/mnj/mnKHmni+1hmn,s-KHmnihmn,s2ds≤mn∫ihi+1hKHmni+1hmn,s-KHmnihmn,s2ds≤∫0TKHmni+1hmn,s-KHmnihmn,s2ds=mni+1hmn-mnihmn2H≤1mn2H,By standard calculations, we will have (22)Fi2+Efi2=EBi+1hH,mn-BihH,mn2≤1mn2H,and it holds that (23)Fj2≤1mn2H.Now, (20) combined with (23) shows that Mn/<M>n→0,a.s. as mn→∞.
Remark 6. The asymptotic normality of estimators is not involved in the results of this paper. In fact, Kubilius and Skorniakov [11] proposed the asymptotic normality of the estimators H^; in view of Remark 2, the asymptotic of σ^ is trivial. For the parameter estimation of fractional diffusion process (1), there are usually two key challenges: the likelihood is intractable and the data is not Markovian. With the Donsker type approximation formula, the statistical inference of fractional diffusion process (FDP) can be simplified to a certain extent. It has proved that the estimator of drift parameter is Lp(p≥1)- consistent and the asymptotic normality may be obtained with more complex operations by the future studies of this area.