Int. J. Contemp. Math. Sciences, Vol. 7, 2012, no. 38, 1861 - 1869

Estimating Fractional 1

Hoang Thi Phuong Thao

Hanoi University of Natural Sciences Vietnam National University 334 Nguyen Trai Str. Thanh Xuan, Hanoi, Vietnam [email protected]

Tran Hung Thao

Institute of Mathematics Vietnamese Academy of Science and Technology No 18 Hoang-Quoc-Viet Road, Hanoi, Vietnam. [email protected]

Abstract

The aim of this paper is to find the best state estimation for frac- tional from observations by a method of approximation.

Mathematics Subjects Classification: 60H30, 91B28

Keywords: State estimation, Fractional stochastic volatility, L2-approximation, Point process

1 Introduction

It is known that stochastic volatility models for options and derivatives were developed out of a need to modify the Black-Scholes model, which failed to effectively take the volatility in the price of underlying security into account. The Black-Scholes model assumed that the volatility of the underlying security was constant, while stochastic volatility models categorized the price of the un- derlying security as a random variable or more general, a . In its turn, the dynamics of this stochastic process can be driven by some other process (commonly by a Brownian motion). Various models have been created for stochastic volatility such as , Constant Elasticity of 1This research is funded by Vietnam National Foundation for Science and Technology Development (NAFOSTED) under grant number 101.02-2011.12. 1862 H. T. P. Thao and T. H. Thao

Variance model (CEV model), SABR volatility model, GARCH model, 3/2- model, ,. . . In many situations, the dynamics of the volatility has a long-range effect, for instance changes of very strong bluechips in a stock market. One thinks of a kind of stochastic volatility perturbed by a process of long memory such as a fractional Brownian motion. In this paper we consider the problem of state estimation of a stochastic volatil- ity Vt given by the fractional stochastic differential equation

H dVt = −bVtdt + σdBt ,

H where Bt is a fractional Brownian motion of Liouville form, from a point process observation Yt. H Recall that a fractional Brownian motion Wt with Hurst index H is a centered H H such that its covariance function R(s, t)=EWs Wt is given by 1 R(s, t)= (t2H + s2H −|t − s|2H ) where 0

 t H 1 α Wt = [Zt + (t − s) dWs], Γ(1 + α) 0 W α H − 1 where is a standard Brownian motion, = 2 and

 0 α α Zt = [(t − s) − (−s) ]dWs. −∞

Since Zt has absolutely continuous trajectories [1] one can see that the term BH t t − s αdW ρ t := 0 ( ) s has long memory in the following sense (see [4]). If n = H H H ∞ E(B1 (Bn+1 − Bn )) then the series n=0 ρn is either divergent or convergent H with very late rate. Bt is not a semimartingale as well and it is known as a fractional Brownian motion of Liouville form. Every stochastic processes here are supposed to be adapted to a filtration (Ft) in a probability space (Ω, F,P). Now we try to estimate the stochastic volatility Vt satisfying

H dVt = −bVtdt + σdBt , 0 ≤ t ≤ T, (1.1.1) Estimating fractional stochastic volatility 1863

2 V0 ∼ N(μ0,ν0 ),b and σ are positive constants from an observation given by a point process of intensity λt  t Yt = λsds + Mt, (1.1.2) 0  λ E t λ2ds < ∞ M F where t is a progressive process such that 0 s , t is a t- mar- H tingale and V0,Bt and Mt are independent. We will use an approximation approach given in [7] to solve this problem.

H 2 Approximation for the fractional process Bt

In this section we will make a brief recall on a principal result for L2-approximation of the fractional Brownian motion (refer to [7]). For every >0 we define  t H, H− 1 Bt = (t − s + ) 2 dWs, (2.2.1) 0 H 1 ,

H, H 2 Theorem 2.1. (See[7]) Bt converges to Bt in L (Ω) when  tends to 0. This convergence is uniform with respect to t ∈ [0,T]. Moreover we have H, H α+ 1 ||Bt − Bt ||L2(Ω) ≤ C(α) 2 where C(α) depends only on α. By this theorem we can see now any fractional stochastic dynamical sys- tem driven by a fractional Brownian notion can be approximated in L2 by  a system governed by a semimartingale Bt , where the traditional stochas- tic calculus can be applied. Then the remaining work is to prove the con- vergence of approximate solutions to the exact solution. Notice that the exact solution of (1.1) can be defined as an adapted process Vt satifying V −b t V ds BH ,V ∼ N μ ,ν2 . t = 0 s + t 0 ( 0 0 ) And our method of estimating stochastic volatility Vt is included in this con- text. 1864 H. T. P. Thao and T. H. Thao

3 Approximate Fractional Stochastic Volatil- ity

Instead of (1.1) and (1.2) we consider the following estimating problem: The system dynamics for volatility is given by

  H, dVt = −bVt dt + σdBt , 0 ≤ t ≤ T (3.3.1)

 2 V0 = V0 ∼ N(μ0,ν0 ).

And the observation dynamics is

 t Yt = λsds + Mt, 0 ≤ t ≤ T. (3.3.2) 0

H, By replacing Bt in (3.1) by its expression in (2.2) we have

   α dVt =[−bVt + σαϕt]dt +  σdWt. (3.3.3)

We can see that (3.3) can be splitted into two equations

,1 ,1 α dVt = −bVt dt +  σdWt, (3.3.4)

,2 ,2  dVt =[−bVt + σαϕt]dt. (3.3.5)

Where (3.4) is a traditional stochastic Langevin equation and (3.5) is in fact  an ordinary differential equation. And the solution Vt of (3.3) should be  ,1 ,2 Vt = Vt + Vt . So the state estimation problem is splitted into two problems as well

 Y ,1 Y ,2 Y E(Vt |Ft )=E(Vt |Ft )+E(Vt |Ft ).

Notation:

Y (i) The state estimation of a process Xt based on the information Ft given by an observation process Yt will be denoted by Xˆt or πt(X). More Y general, the notation πt(f) stands for the state estimation E(f(Xt)|Ft ), where f is a bounded and continuous function, f ∈ Cb(R). Recall that Y Ft = σ(Ys,s ≤ t) is the σ-algebra generated by all random variable Ys with 0 ≤ s ≤ t.

(ii) Suppose that the probability P is obtained from a probability Q by an absolutely continuous change of measure Q → P such that nt = Yt − t Estimating fractional stochastic volatility 1865

Y is a (Q, Ft )- martingale. A Bayes formula gives us

Y Y EQ(XtLt|Ft ) EP (Xt|Ft )= , EQ(Lt)

where  dP  t L E |F Y λ Y − λ ds. t = [dQ t ]= s s exp (1 s) 0≤s≤t 0

Denote by σ(Xt)orX˜t the unnormalized state estimation of Xt, given by

Y X˜t = σ(Xt)=EQ(LtXt|Ft ),

or more general

Y σ(f)=EQ(Ltf(Xt)|Ft ), for f ∈ Cb(R).

,1 4 Estimating Vt

,1 Y Theorem 4.1. (i) The state estimation πt(f)=E[f(Vt )|Ft ] for the sys- tem dynamics (3.4) with the point process observation (3.2) is given by

 t α 2 ,1  ,1 ( σ)  ,1 πt(f)=π0(f)+ πs[−bVt f (Vt )+ f (Vt )]ds 0 2b

 t −1 + πs (λ)[πs−(λf) − πs(f)πs−(λ)]dms, (4.4.1) 0  m Y − t λ ds, f ∈ C2 R . where t = t 0 s b ( )

,1 Y (ii) The unnormalized state estimation σt(f)=EQ[Ltf(Vt )|Ft ] is given by

 t α 2 ,1  ,1 ( σ)  ,1 σt(f)=σ0(f)+ σs[−bVt f (Vt )+ f (Vt )]ds 0 2b

 t + [σs−(λf) − σs(f)]dμs, (4.4.2) 0

2 where μt = Yt − t, f ∈ Cb (R). 1866 H. T. P. Thao and T. H. Thao

,1 Proof. It is known that the solution Vt is in fact an Ornstein-Uhlenbeck process that is a Markov process and its semigroup is defined by a family (Pt,t≥ 0) of operations on bounded Borelian function f.  ασ 2  −bt ( ) −2bt (Ptf)(x)= f[e x + 1 − e y]μ(dy), (4.4.3)

2b Ê where μ is a standard Gaussian measure on R

1 −y2 μ(dy)=√ exp ( )dy. 2π 2

,1 It is obvious that lim(Ptf)(x)=f(x) then Vt is a . The corre- t0 sponding infinitesimal operator At is given by

α 2 1  ( σ)  (Atf)(x) = lim (Ptf − f)(x)=−bxf (x)+ f (x). (4.4.4) t→0 t 2b

According to a result from [5], the state estimation πt(f) for a Fellerian system Y t λ ds M process from a point process observation t = 0 s + t is given by

 t  t −1 πt(f)=π0(f)+ πs(Af)ds + πs (λ)[πs−(λf) − πs(f)πs−(λ)]dms, 0 0  m Y − t π λ ds, π−1 where t = t 0 s( ) provided t = 0 and the unnormalized state estimation σt(f) is given by

 t  t σt(f)=σ0(f)+ σs(Af)ds + [σs−(λf) − σs(f)]dμs, 0 0 where μt = Yt − t. By replacing Af in these equations by its expression (4.4) we get equations (4.1) and (4.2) as required.

Corollary 4.1. By taking f = I (identification function) we have

 t  t ,1 ,1 ,1 1 ,1 ,1 Vt = V0 − bVs ds + [λsVs − λˆs−Vs ](dYs − λˆsds). (4.4.5) 0 0 λˆs

 t  t ,1 ,1 ,1 ,1 Vt := σt(I)=V0 − bVs ds + [λ˜s− − Vs ](dYs − ds). (4.4.6) 0 0 Estimating fractional stochastic volatility 1867

,2  5 Estimating Vt and Vt .

It is clear from (3.5) that

 t ,2 ,2 ,2  Vt = V0 + [−bVs + σαϕˆs]ds (5.5.1) 0 and  t ,2 ,2 ,2  Vt = V0 + [−bVs + σαϕ˜s]ds. (5.5.2) 0 Remark 5.1. These equations have explicit solutions as follows:

 t ,2 ,2 −bt −b(t−s)  Vt = V0 e − σα e ϕˆsds (5.5.3) 0 and  t ,2 ,2 −bt −b(t−s)  Vt = V0 e − σα e ϕ˜sds. (5.5.4) 0 Combining (4.5), (4.6) and (5.1), (5.2) we get Corollary 5.1.

 ,1 ,2 Vt = Vt + Vt (5.5.5)  t  t  t V  − bV ds σαϕ ds 1 λV ,1 − λˆ V,1 dY − λˆ ds = 0 s + s + [ s s s− s ]( s s ) 0 0 0 λs (5.5.6)

 ,1 ,2 Vt = Vt + Vt (5.5.7)  t  t  t    ,1 = V0 − b Vs ds + σα ϕsds + [λ˜s− − Vs ](dYs − ds), (5.5.8) 0 0 0

,1 ,2  ,1 ,2  where we assume that V0 + V0 = V0 and V0 + V0 = V0 .

 6 Convergence of Vt to the solution Vt

 Assume that Vt and Vt are solutions of (1.1) and (3.1), as in [5] we can establish

 2 Proposition 6.1. Vt converges to Vt in L (Ω) uniformly with respect to t ∈ [0,T]. 1868 H. T. P. Thao and T. H. Thao

Proof. We have

 t   H H, Vt − Vt = −b (Vs − Vs )ds + σ(Bt − Bt ) (6.6.1) 0 so

 t   H H, ||Vt − Vt || ≤ ||b (Vs − Vs )ds|| + σ||Bt − Bt || (6.6.2) 0

It follows from (2.1) that

 t   α+ 1 ||Vt − Vt || ≤ b ||Vs − Vs ||ds + σC(α) 2 , (6.6.3) 0 where ||.|| stands for the norm in L2(Ω). Applying Gronwall’s Lemma to (6.3) we will get

 α+ 1 bt ||Vt − Vt || ≤ σC(α) 2 e . (6.6.4)

And we see that

 α+1 bT sup ||Vt − Vt || ≤ σC(α) e . (6.6.5) 0≤t≤T

 2 So Vt −→ Vt in L (Ω) uniformly with respect to t ∈ [0,T].

7 Estimating volatility Vt

Proposition 7.1. The state estimation Vˆt of the stochastic volatility Vt is the 2  ,1 ,2 L -limit of Vˆt = Vt + Vt as  → 0 and the convergence is uniform with respect to t ∈ [0,T].

Proof. The assertion of this proposition follows from a result on the conver-  L2 2 gence of condition expectation (refer, for instance, to [3]). If Vt −→ Vt ∈ L  Y Y as  → 0 then E(Vt |Ft ) −→ E(Vt|Ft )as → 0. And the uniform conver- gence is deduced from the estimation (6.5).

References

[1] E. Al`os, O.Mazet, D.Niralart, with respect to Frac- tional Motion with Hurst Paramater less than 1/2, Stochastic Processes and Application, 86 (2006), 121-139. Estimating fractional stochastic volatility 1869

[2] M. Lo`eve, , D.Van Nostrand Company, Third Edition, 1963.

[3] D. Revuz and M.Yor, Continuous Martingales and Brownian Motion, Springer, 3rd, 1998.

[4] A. N. Shiryaev, Essentials of Stochastic Facts, Models, Theory, Book of 834 pages, World Scientific, 1999.

[5] T. H. Thao, Optimal State Estimation of A Markov Process From Point Process Observations, Annales Scientifiques de l’Universit´e Blaise Pascal, Clermont Ferrand II, Fasc. 9, 1991, 1-10.

[6] T. H. Thao and T. T. Nguyen, Fractal Langevin EquationVietnam Jour- nal of Mathematics, Vol. 30, No.1, 2002, 89-96.

[7] T. H. Thao, An Approximate Approach to Fractional Analysis for Fi- nance, Nonlinear Analysis: Real World Applications, Vol. 7, 2006, 124- 132.

Received: April, 2012