Joint Parameter Estimation of the Ornstein-Uhlenbeck SDE driven by Fractional Brownian Motion

Transcription

1 Joint Parameter Estimation of the Ornstein-Uhlenbeck SDE driven by Fractional Brownian Motion Luis Barboza October 23, 2012 Department of Statistics, Purdue University () Probability Seminar 1 / 59

2 Introduction Main Objective: To study the GMM (Generalized Method of Moments) joint estimator of the drift and memory parameters of the Ornstein-Uhlenbeck SDE driven by fbm. Joint work with Prof. Frederi Viens (Department of Statistics, Purdue University). Department of Statistics, Purdue University () Probability Seminar 2 / 59

3 Outline 1 Preliminaries 2 Joint estimation of Gaussian stationary processes 3 fou Case 4 Simulation Department of Statistics, Purdue University () Probability Seminar 3 / 59

4 Outline 1 Preliminaries 2 Joint estimation of Gaussian stationary processes 3 fou Case 4 Simulation Department of Statistics, Purdue University () Probability Seminar 4 / 59

5 Fractional Brownian Motion The fractional Brownian motion (fbm) with Hurst parameter H (0, 1) is the centered Gaussian process B H t, continuous a.s. with: Cov(B H t, B H s ) = 1 2 ( t 2H + s 2H t s 2H ), t, s R. Some properties: Self-Similarity: for each a > 0 Bat H d = a H Bt H. It admits an integral representation with respect to the standard Brownian motion over a finite interval: B H t = t 0 K H (t, s)dw (s) Department of Statistics, Purdue University () Probability Seminar 5 / 59

6 Fractional Gaussian noise (fgn) Definition: N H t := B H t B H t α where α > 0. Gaussian and stationary process. Long-memory behavior of the increments when H > 1 2 that n ρ(n) = ). Ergodicity. (in the sense Department of Statistics, Purdue University () Probability Seminar 6 / 59

7 Fractional Gaussian noise (fgn) Autocovariance function: ρ θ (t) = 1 [ t + α 2H + t α 2H 2 t 2H] 2 Spectral density: f θ (t) = 2c H (1 cos t) t 1 2H where c H = sin(πh)γ(2h+1) 2π. (Beran, 1994). Department of Statistics, Purdue University () Probability Seminar 7 / 59

8 Estimation of H (fbm and fgn) Classical methods: R/S statistic (Hurst 1951), Variance plot (Heuristic method). Robinson (1992,1995): Semiparametric Gaussian estimation. (spectral information) MLE Methods: Whittle s estimator. Department of Statistics, Purdue University () Probability Seminar 8 / 59

9 Estimation of H Methods using variations (filters): Coeurjolly (2001): consistent estimator of H (0, 1) based on the asymptotic behavior of discrete variations of the fbm. Asymptotic normality for H < 3/4. Tudor and Viens (2008): proved consistency and asymptotics of the second-order variational estimator (convergence to Rosenblatt random variable when H > 3/4) using Malliavin calculus. Department of Statistics, Purdue University () Probability Seminar 9 / 59

10 Ornstein-Uhlenbeck SDE driven by fbm Cheridito (2003) Take λ, σ > 0 and ζ a.s bounded r.v. The Langevin equation: t X t = ζ λ X s ds + σbt H, t 0 0 has as an unique strong solution and it is called the fractional Ornstein-Uhlenbeck process: ( t ) ζ X t = e λt ζ + σ e λu dbu H, t T 0 and this integral exists in the Riemann-Stieltjes sense. Department of Statistics, Purdue University () Probability Seminar 10 / 59

11 Properties Cheridito (2003): Stationary solution (fou process): t X t = σ e λ(t u) dbu H, t > 0 Autocovariance function (Pipiras and Taqqu, 2000): ρ θ (t) = 2σ 2 c H 0 cos(tx) x 1 2H λ 2 + x 2 dx where c H = Γ(2H+1) sin(πh) 2π. Department of Statistics, Purdue University () Probability Seminar 11 / 59

12 Properties Cheridito (2003): X t has long memory when H > 1 2, due to the following approximation when x is large: X t is ergodic. ρ θ (x) = H(2H 1) x 2H 2 + O(x 2H 4 ). λ X t is not self-similar, but it exhibits asymptotic selfsimilarity (Bonami and Estrade, 2003): f θ (x) = c H x 1 2H + O( x 3 2H ). Department of Statistics, Purdue University () Probability Seminar 12 / 59

13 Estimation of λ given H (OU-fBm) MLE estimators: Kleptsyna and Le Breton (2002): MLE estimator based on Girsanov formula for fbm. Strong consistency when H > 1/2. Tudor and Viens (2006): Extended the K&L result to more general drift conditions. Strong consistency of MLE estimator when H < 1 2 using Malliavin calculus. They work with the non-stationary case. Department of Statistics, Purdue University () Probability Seminar 13 / 59

14 Estimation of λ given H (Least-Squares methods) Hu and Nualart (2010) Estimate of λ which is strongly consistent for H 1 2. ( 1 λ T = HΓ(2H)T T λ T is asymptotically normal if H ( 1 2, 3 4 ) 0 ) 1 0 Xt 2 2H dt The proofs of these results rely mostly on Malliavin calculus techniques. Department of Statistics, Purdue University () Probability Seminar 14 / 59

15 Estimation of H and λ Methods based on variations: Biermé et al (2011): For fixed T, they use the results in Biermé and Richard (2006) to prove consistency and asymptotic normality of the joint estimator of (H, σ 2 ) for any stationary gaussian process with asymptotic self-similarity (Infill-asymptotics case). Brouste and Iacus (2012): For T and α 0 they proved consistency and asymp. normality when 1 2 < H < 3 4 for the pair (H, σ 2 ) (non-stationary case). Department of Statistics, Purdue University () Probability Seminar 15 / 59

16 Outline 1 Preliminaries 2 Joint estimation of Gaussian stationary processes 3 fou Case 4 Simulation Department of Statistics, Purdue University () Probability Seminar 16 / 59

17 Preliminaries X t : real-valued centered gaussian stationary process with spectral density f θ0 (x). f θ (x): continuous function with respect to x, continuously differentiable with respect to θ. θ belongs a compact set Θ R p. Bochner s theorem: ρ θ (s) := Cov(X t+s, X t ) = R cos(sx)f θ (x)dx Department of Statistics, Purdue University () Probability Seminar 17 / 59

18 Preliminaries If ρ θ (s) is a continuous function of s, then the process X t is ergodic. Assumption 1 Take α > 0 and L a positive integer. Then there exists k {0, 1,..., L} such that ρ θ (αk) is an injective function of θ. Department of Statistics, Purdue University () Probability Seminar 18 / 59

19 Preliminaries Recall: a(l) := (a 0 (l),..., a L (l)) is a discrete filter of length L + 1 and order l for L Z + and l {0,..., L} if L a k (l)k p = 0 for 0 p l 1 k=0 L a k (l)k p 0 if p = l. k=0 Examples: finite-difference filters, Daubechies filters (wavelets). Assume that we can choose L filters with orders l i {1,..., L} for i = 1,..., L and a extra filter with l 0 = 0. Department of Statistics, Purdue University () Probability Seminar 19 / 59

20 Preliminaries Define the filtered process of order l i and step size f > 0 at t 0 as: ϕ i (X t ) := L a q (l i )X t f q q=0 Its expected value is: V i (θ 0 ) := E[ϕ i (X t ) 2 ] = L k=0 b k(l i )ρ θ0 ( f k). Define the set of moment equations by: where g(x t, θ) := (g 0 (X t, θ),..., g L (X t, θ)) g i (X t, θ) = ϕ i (X t ) 2 V i (θ), for 0 i L. Department of Statistics, Purdue University () Probability Seminar 20 / 59

21 Preliminaries Assume that we have observed the stationary process X t at times 0 = t 0 < t 1 < < t N 1 < t N = T and α := t i t i 1 > 0 (fixed). Assume there exists a sequence of symmetric positive-definite random matrices { N } such that  N p A, and A > 0. Define: ĝ N (θ) := 1 N L+1 N i=l g(x t i, θ). (sample moments) ˆQ N (θ) := ĝ N (θ)  N ĝ N (θ). Q 0 (θ) := E[g(X t, θ)] AE[g(X t, θ)]. Department of Statistics, Purdue University () Probability Seminar 21 / 59

22 GMM estimation Define the GMM estimator of θ 0 : ˆθ N : = argmin θ Θ ˆQ N (θ) ( ) T ( ) 1 N 1 N = argmin θ Θ g(x ti, θ) Â N g(x ti, θ). N N i=1 i=1 Department of Statistics, Purdue University () Probability Seminar 22 / 59

23 Consistency Lemma 1 Under the above assumptions: (i) (ii) if and only if θ = θ 0. sup ˆQ N (θ) Q 0 (θ) a.s 0. θ Θ Q 0 (θ) = 0. Department of Statistics, Purdue University () Probability Seminar 23 / 59

24 Consistency Key aspects of proof: Ergodicity of X t Continuity of ρ θ ( ) over the compact set Θ. Injectivity of: ρ θ (α 0) ρ θ (α) :=. ρ θ (α L) Results of Newey and McFadden, (GMM) Department of Statistics, Purdue University () Probability Seminar 24 / 59

25 Consistency We have all the conditions to apply: Theorem 1 (Newey and McFadden, 1994) Under the above assumptions, it holds that: ˆθ N a.s. θ 0. Department of Statistics, Purdue University () Probability Seminar 25 / 59

26 Asymptotic Normality of the sample moments Denote: Ĝ N (θ) := θ ĝ N (θ) G(θ) := E[ θ g(x t, θ)] Assumption 2 For fixed α > 0, assume that the (L + 1) p matrix θ ρ θ (α) is a full-column rank matrix for any θ Θ. We can prove that: Ĝ N (θ) = G(θ) = θ V (θ), where V (θ) = (V 0 (θ),..., V L (θ)). Department of Statistics, Purdue University () Probability Seminar 26 / 59

27 Asymptotic Normality of the sample moments And based on Biermé et al (2011) article, let us assume: Assumption 3 (Biermé et al s condition) Assume that for any l, l {l 0,..., l L } we have that: R(u l, l ) := min (1, u l+l ) ( ) u + 2πp f θ0 L 2 (( π, π)) α p Z Department of Statistics, Purdue University () Probability Seminar 27 / 59

28 Asymptotic Normality of the sample moments Lemma 2 Under Assumptions 1-3 NĝN (θ 0 ) d N(0, Ω) where π Ω ij = 2α 2 P ali [cos u] 2 P alj [cos u] 2 f θ0 (u/α) 2 du π and f θ0 (u) := p Z f θ 0 (u + 2πp). Note: P al (x) := L k=0 a k(l)x k. Department of Statistics, Purdue University () Probability Seminar 28 / 59

29 Sketch of proof Let V D (θ 0 ) := Diag (1/V j (θ 0 )) j {0,...,L}. We scale the vector g(x t, θ 0 ) as follows: N N ( NVD (θ 0 )ĝ N (θ 0 ) = H2 (Z lj,t N L + 1 i ) ) j {0,...,L} where Z lj,t i := ϕ j (X ti ) Vj (θ 0 ) and H 2( ) is the 2nd-order Hermite process. Use the vector-valued version of the Breuer-Major theorem with spectral-information conditions (Biermé et al, 2011) to deduce the asymptotic behavior of the previous sum. i=l Department of Statistics, Purdue University () Probability Seminar 29 / 59

30 Asymptotic behavior of the error Let ε N := ˆθ N θ 0. Using the mean value theorem: ε N := ˆθ N θ 0 = ψ N ( θ N, ˆθ N ) ĝ N (θ 0 ) where ψ N ( θ N, ˆθ N ) := [G(ˆθ N ) Â N G( θ N )] 1 G(ˆθ N ) Â N. Also note that: ψ N ( θ N, ˆθ N ) p [G(θ) AG(θ)] 1 G(θ) A then we can bound E[ ψ N ( θ N, ˆθ N ) 4p ] for any p > 0 Department of Statistics, Purdue University () Probability Seminar 30 / 59

31 Asymptotic behavior of the error X t can be represented as a Wiener-Ito integral with respect to the standard Brownian motion: where A k (x θ 0 ) := cos(αkx) f θ0 (x). X tk = I 1 (A k ( θ 0 )) Using the multiplication rule of Wiener integrals: (ĝ N (θ)) i = I 2 [B i,j ( θ)] where B i,j ( θ) is a kernel depending on the filter a. Department of Statistics, Purdue University () Probability Seminar 31 / 59

32 Asymptotic behavior of the error We already proved a CLT for ĝ N (θ 0 ), hence, for all i: E[ (ĝ N (θ 0 )) i 2 ] = O(N 1 ) Let p > 0, we can use the equivalence of L p -norms of a fixed Weiner chaos to get: By Borel-Cantelli, we conclude: E[ ĝ N (θ 0 ) 4p ] < τ p,l N 2p Department of Statistics, Purdue University () Probability Seminar 32 / 59

33 Asymptotic behavior of the error Corollary 1 Under Assumptions 1-3 we have: (i) E[ ˆθ N θ 0 2 ] = O(N 1 ) (ii) N γ ˆθ N θ 0 a.s. 0 for any γ < 1 2. Department of Statistics, Purdue University () Probability Seminar 33 / 59

34 Asymptotic behavior of the error And we can generalize the previous result as follows: Theorem 2 Let ˆθ N the GMM estimator of θ 0. Assume there exists a diagonal (L + 1) (L + 1) matrix D N (θ 0 ) such that: E[D (i,i) N (θ 0) (ĝ N (θ 0 )) i 2 ] = O(1) for all i {0,..., L}. Then: ( ) (i) E[ ˆθ N θ 0 2 ] = O max 0 i L {D (i,i) N (θ 0)} (ii) If max 0 i L {D (i,i) N (θ 0)} = f (N) N, for f (N) = o(n) and ν > 0 then for ν any γ < ν 2 : N γ ε N a.s. 0. Department of Statistics, Purdue University () Probability Seminar 34 / 59

35 Asymptotic Normality Theorem 3 Let X t a Gaussian stationary process with parameter θ 0. If ˆθ N is the GMM estimator of θ 0, then under Assumptions 1-3, it holds that: N(ˆθ N θ 0 ) d N(0, C(θ 0 )ΩC(θ 0 ) ) where C(θ 0 ) = [G(θ 0 ) AG(θ 0 )] 1 G(θ 0 ) A. Key ideas of the proof (Newey and McFadden, 1994; Hansen, 1982): Linear behavior of ε N = ˆθ N θ 0. Slutsky s theorem. Department of Statistics, Purdue University () Probability Seminar 35 / 59

36 Efficiency The GMM asymptotic variance is minimized by taking A = Ω 1. There are numerical techniques to compute ÂN such that p  N Ω 1, for example (Hansen et al., 1996): Two-step estimators. Iterative estimator. Continuous-updating estimator. Department of Statistics, Purdue University () Probability Seminar 36 / 59

37 Outline 1 Preliminaries 2 Joint estimation of Gaussian stationary processes 3 fou Case 4 Simulation Department of Statistics, Purdue University () Probability Seminar 37 / 59

38 Preliminaries Assume that there exist a closed rectangle Θ R 2 such that θ = (H, λ) Θ, and assume that σ = 1. ρ θ (t) and its partial derivatives are continuous functions of θ. The Assumptions 1 and 2 are difficult to confirm due to the analytical complexity of the covariance function ρ θ (t). we decided to check these two assumptions at least locally, by checking numerically that: ] det [ ρθ (0) H ρ θ (α) H for different values of θ and α. ρ θ (0) λ ρ θ (α) λ 0 The injectivity approximately holds for 0 < λ < 5 and H > 0.3. Department of Statistics, Purdue University () Probability Seminar 38 / 59

39 Preliminaries Lemma 3 Let X t be the stationary fou process with parameters θ = (H, λ). The Assumption 3 (Biermé et al s condition) holds under the following two cases: Case 1 If l + l > 1 then it holds for all H (0, 1). Case 2 If l + l 1 then it holds if H (0, 3 4 ). Conclusion: Assumption 3 is not valid for the first component of ĝ N if H 3 4. Department of Statistics, Purdue University () Probability Seminar 39 / 59

40 Asymptotic Normality of the sample errors Using the previous lemma, for H < 3 4 : NĝN (θ) d N(0, Λ(θ)) where the covariance matrix Λ(θ) has entries: π Λ ij (θ) =2cH 2 α4h P ali [cos u] 2 Palj [cos u] 2 u + 2πp 1 2H 2 π (u + 2πp) 2 + (λα) 2 du p Z }{{} K i,j (α) Department of Statistics, Purdue University () Probability Seminar 40 / 59

41 Asymptotic Normality of the sample errors Lemma 4 Assume that X t is a stationary fou process with parameters θ = (H, λ) where H 3 4. Then: (i) If H = 3 4 : N log N (ĝ d N(θ)) 0 N(0, 2α 1 cθ 2 ) where c θ = H(2H 1) λ = 3 8λ. (ii) If H > 3 4, (ĝ N(θ)) 0 does not converge to a normal law, or even a second-chaos law. However, [ E N 2 2H (ĝ N (θ)) 0 2] = O(1). Department of Statistics, Purdue University () Probability Seminar 41 / 59

42 Sketch of proof Denote: ĝ N,0 (θ) = (ĝ N (θ)) 0. For H = 3 4, as N : [ ] N E log N ĝ N,0(θ) 2 2α 1 cθ 2 := c 1 where c θ := H(2H 1) λ. For H > 3 4 : E[ N 2 2H ĝ N,0 (θ) 2 ] 2α 4H 4 c 2 θ (2H 1)(4H 3) := c 2 Department of Statistics, Purdue University () Probability Seminar 42 / 59

43 Sketch of proof N c If F N = 1 log N ĝn,0(θ) if H = 3 4 N 4 4H c 2 ĝ N,0 (θ) if H > 3 4 then we need to prove where H = L 2 (( π, π)). DF N 2 H This result is valid only if H = 3 4. L 2 (Ω) 2 Hence we can use Nualart and Ortiz-Latorre (2008) to conclude asymptotic normality of F N. Department of Statistics, Purdue University () Probability Seminar 43 / 59

44 Sketch of proof In the case H > 3 4 we can use the same criteria to conclude that F N does not have a normal limit in law. Furthermore, F N has this kernel: I N (r, s) := N1 2H c2 N j=1 (A j A j )(r, s) It can be proved that I N (r, s) is not a Cauchy sequence in H 2. Then F N does not have a 2nd-chaos limit. Department of Statistics, Purdue University () Probability Seminar 44 / 59

45 Asymptotic behavior of the error Proposition 1 Let X t be a fou process with parameters θ = (H, λ). Then the GMM estimate ˆθ N satisfies: (i) (ii) As N goes to infinity: O(N ( 1 ) ) if H (0, 3 4 ) E[ ˆθ N θ 2 ] = O log N N if H = 3 4 O(N 4H 4 ) if H ( 3 4, 1) N γ ˆθ N θ a.s. 0 for all γ < 1 2 (if H 3 4 ). Otherwise for all γ < 2 2H, if H > 3 4. Department of Statistics, Purdue University () Probability Seminar 45 / 59

46 Asymptotics of ˆθ N : Proposition 2 Let X t be the stationary fou process with parameters θ = (H, λ). Then, for any positive-definite matrix A, the GMM estimator of θ is consistent for any H (0, 1) and: If H (0, 3 4 ): N(ˆθ N θ) d N(0, C(θ)ΛC(θ) ) where C(θ) = [G(θ) AG(θ)] 1 G(θ) A and Λ = 2c 2 H α4h K(α). If H = 3 4, then N log N (ˆθ N θ) d N where (C(θ)) 1 is the first column of C(θ). ( ) 0, 2c2 θ α (C(θ)) 1(C(θ)) 1 Department of Statistics, Purdue University () Probability Seminar 46 / 59

47 Asymptotics of ˆθ N : If H > 3 4, the GMM estimator does not converge to a multivariate normal law or even a second-chaos law. Department of Statistics, Purdue University () Probability Seminar 47 / 59

48 Outline 1 Preliminaries 2 Joint estimation of Gaussian stationary processes 3 fou Case 4 Simulation Department of Statistics, Purdue University () Probability Seminar 48 / 59

49 Numerical Details Approximation of ρ θ (t) = 2σ 2 c H 0 cos(tx) x1 2H λ 2 +x 2 dx: Filon s Method (integration of oscillatory integrands) 4-5 precision digits. Simulation of fbm: Davies and Harte (1987). Approximation of fbm-ou process: the stationary fbm-ou satisfies: ti+1 X ti+1 = e λ X ti + σ e λ(t i u) dbu H t i where X 0 N(0, ρ θ (0)). For comparison purposes we use the yuima R package of Brouste and Iacus (2012). Finite-difference filters. Department of Statistics, Purdue University () Probability Seminar 49 / 59

50 Asymptotic behavior of N(ˆθ N θ) Figure: Asymptotic variance of N(Ĥ N 0.62) Department of Statistics, Purdue University () Probability Seminar 50 / 59

51 Asymptotic behavior of N(ˆθ N θ) Figure: Asymptotic variance of N(ˆλ N 0.8) Department of Statistics, Purdue University () Probability Seminar 51 / 59

52 Numerical Details (H, λ) = (0.37, 1.45) N = 1000, α = 1. L = 3 (filters maximum order) Number of repetitions in the sampling dist.: 100. Efficient matrix estimation: Continuous-updating estimator. Department of Statistics, Purdue University () Probability Seminar 52 / 59

53 Sampling distribution of ĤN Department of Statistics, Purdue University () Probability Seminar 53 / 59

54 Sampling distribution of ˆλ N Department of Statistics, Purdue University () Probability Seminar 54 / 59

55 Some statistics... Empirical MSE of ˆθ N Asymp. variance of ĤN (empirical) Asymp. variance of Ĥ N (theoretical) Asymp. variance of ˆλ N (empirical) Asymp. variance of ˆλ N (theoretical) Department of Statistics, Purdue University () Probability Seminar 55 / 59

56 Sampling distribution of ĤN (H, λ) = (0.8, 2) with 4 filters. Department of Statistics, Purdue University () Probability Seminar 56 / 59

57 Sampling distribution of ˆλ N (H, λ) = (0.8, 2) with 4 filters. Department of Statistics, Purdue University () Probability Seminar 57 / 59

58 Future Work More accurate simulation of the fou process? Asymptotic law of ˆθ N when H > 3 4? (fou case) Department of Statistics, Purdue University () Probability Seminar 58 / 59

59 Thanks! Department of Statistics, Purdue University () Probability Seminar 59 / 59