Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Simulate α-Stable Random Variable and Estimate Stable Parameters Based on Market Data Author: Yiyang Yang Advisor: Pr. Xiaolin Li, Pr. Zari Rachev Department of Applied Mathematics and Statistics State University of New York at Stony Brook August 2012 Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Denition Characteristic Function Outline 1 Denition of α-Stable Random Variable Denition Characteristic Function 2 Weron's Algorithm for Generating Stable Random Variable Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable 3 Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Denition Characteristic Function Denition Given independent and identically distributed random variables X1 , X2 , · · · , X and X , then random variable X is said to follow an α-Stable distribution if there exists a positive constant C and a real number D such that the following relation holds: n n n X1 + X2 + · · · + X = C X + D n n n 1 where = denotes equality in distribution and constant C = n α determines the stability property. n When α = 2, it is the Gaussian case; when 0 < α < 2, have the non-Gaussian case. There are three special cases with a closed form p.d.f. the Gaussian case (α = 2) the Cauchy case (α = 1, β = 0) the Lévy case α = 21 , β = ±1 Yiyang Yang Simulate α-Stable Random Variable we Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Denition Characteristic Function α-Stable distribution does not have closed form density function and is expressed by characteristic function (form A): φstable (t ; α, σ, β, µ) = E e itX ( = where exp i µt − |σ t |α 1 − i β (signt ) tan πα 2 exp i µt − σ |t | 1 + i β π2 (signt ) ln |t | 1, signt = 0, − 1, t t t α 6= 1 α=1 >0 =0 <0 Four related parameters are: α: the index of stability or the shape parameter, α ∈ (0, 2) β : the skewness parameter, β ∈ [−1, 1] σ : the scale parameter, σ ∈ (0, +∞) µ: the location parameter, µ ∈ (−∞, +∞) Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Outline 1 Denition of α-Stable Random Variable Denition Characteristic Function 2 Weron's Algorithm for Generating Stable Random Variable Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable 3 Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable By transforming β, σ , there is an equivalent denition of α-stable distribution and characteristic function. Denition A random variable X is α-stable with parameters (α, β2 , σ2 , µ) if and only if its characteristic function (form B) is given by ln φ (t ) = ( i µt − σ2α |t |α exp −i β2 (signt ) π2 K (α) π i µt − σ2 |t | 2 + i β2 (signt ) ln |t | where K (α) = α − 1 + sign (1 − α) = ( α α−2 α 6= 1 α=1 α<1 ; α>1 and for α 6= 1 tan β2 π K (α) 2 = β tan πα 2 πα 21α , σ2 = σ 1 + β 2 tan2 ; 2 and for α = 1 β2 = β, σ2 = Yiyang Yang 2 π σ. Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Corollary Any two admissible parameter quadruples (α, β, σ, µ) and (α, β, σ 0 , µ0 ) uniquely determine real numbers a > 0 and b such that X d (α, β, σ, µ) = aX α, β, σ 0 , µ0 + b where σ a = , σ0 ( b = µ − µ0 σσ0 µ − µ0 σσ0 + σβ π2 ln σ σ0 α 6= 1 . α=1 X (α, β, 1, 0) and transform it to the general case X (α, β, σ, µ) = aX (α, β, 1, 0) + b with Consider the standard case a = σ, b = ( µ µ + σβ π2 ln σ α 6= 1 . α=1 There are three formulas regarding probability density function, cumulative density function and characteristic function f (−x , α, β) = f (x , α, −β) F (−x , α, β) = 1 − F (x , α, −β) φ (−t , α, β) = φ (t , α, −β) Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Assume f (x , α, β) and φ (t , α, β) are the density function and characteristic function (form B) of stable random variable X (α, β, 1, 0), then according to the inversion formula of characteristic function f (x , α, β) = = 1 ˆ ∞ 0 π 2 1 ˆ π 2 ∞ e −itx φ (t , α, β) dt −∞ e −itx φ (t , α, β) dt + ˆ ∞ 0 e itx φ (−t , α, β) dt . Since e −itx φ (t , α, β) = e itx φ (−t , α, β), then f (x , α, β) = Re 1 = 1 π Re ˆ π ∞ 0 ˆ ∞ 0 e −itx φ (t , α, β) dt e itx φ (−t , α, β) dt = Re 1 ˆ π ∞ 0 e itx φ (t , α, −β) dt . Finally, the integral form of the density function is f (x , α, β) = Re 1 ˆ ∞ 0 π in the case α 6= 1, and f (x , 1, β) = Re 1 π in the case α = 1. exp ˆ 0 π −itx − t α exp −i β K (α) dt 2 ∞ exp Yiyang Yang −itx − π 2 t − i β t ln t dt Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Theorem Let π K (α) , (α) = sign (1 − α) , γ0 = − β2 2 α 1 K (α) C (α, β2 ) = 1 − 1 + β (1 + (α)) , 4 α α 1−α cos (γ − α (γ − γ0 )) sin α (γ − γ0 ) Uα (γ, γ0 ) = cos γ cos γ π + β2 γ 1 π U1 (γ, β2 ) = 2 exp + β2 γ tan γ , cos γ β2 2 then the cumulative function of a standard stable distribution can be written as: F (x , α, β2 ) = C (α, β2 ) + F (x , 1, β2 ) = 1 π ˆ (α) π π 2 −π 2 exp ˆ π 2 exp γ0 − exp − Yiyang Yang α −x 1−α Uα (γ, γ0 ) d γ, α 6= 1, x β2 x >0 U1 (γ, β2 ) d γ, α = 1, β2 > 0. Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Theorem Let γ0 and Uα (γ, γ0 ) be dened as above, for α 6= 1, Sα (1, β2 , 0) random variable i for x > 0 1 π ˆ π 2 γ0 exp − α−1 x α Uα (γ, γ0 ) ( d (0 < X ≤ x ) P (X ≥ x ) P γ= γ0 < γ < π2 , X is a α<1 . α>1 Proof: 0 <α<1 F (x , α, β2 ) = P (X ≤ x) = = given that for 1 <α<2 α < 1, 1 1−β2 = 2 F (x , α, β2 ) = P (X 1 − β2 2 − β2 2 P (X + 1 ˆ π π 2 exp γ0 α−1 −x α Uα (γ, γ0 ) d γ + P (0 < X ≤ x ) ≤ 0). ≤ x) = 1 − 1 π ˆ π 2 γ0 exp −x α−1 α Uα (γ, γ0 ) d γ = 1 − P (X ≥ x ) . This completes the proof. Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Theorem Let γ0 be dened as above, γ be uniformly distributed on − π2 , π2 and W be an independent exponential randome variable with mean 1. Then sin α (γ − γ0 ) cos (γ − α (γ − γ0 )) X= 1 W (cos γ) α 1−α α (1) is Sα (1, β2 , 0) for α 6= 1. X= W cos γ + β2 γ tan γ − β2 log π 2 2 + β2 γ π is S1 (1, β2 , 0). Yiyang Yang Simulate α-Stable Random Variable (2) Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Proof. α 1−α 1−α α cos(γ−α(γ−γ0 )) 0) Assume a (γ) = sin α(γ−γ , then (1) is α(γ) cos γ cos γ When 0 < α < 1, (1) implies that X > 0 i γ > γ0 . Since 1−α α >0 W P (0 < X =P =P 0 < W α (γ) 1−α α W . ≤ x ) = P (0 < X ≤ x , γ > γ0 ) ≤ x , γ > γ0 ! =P W ≥ x α−1 a (γ) , γ > γ0 α α α ≥ x α−1 a (γ) P (γ > γ0 ) = Eγ exp −x α−1 a (γ) 1{γ>γ0 } ˆ = π 2 γ0 exp α −x α−1 a (γ) 1 π dγ Given a (γ) = Uα (γ, γ0 ) , we have proved X ∼ Sα (1, β2 , 0). And with the similar method, we can prove the case 1 the caseα <α<2 = 1. Yiyang Yang Simulate α-Stable Random Variable and Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable Algorithm: Generate a random variableU uniformly distributed on − π2 , π2 and an independent exponential random variable E with mean 1. For α 6= 1, compute sin (α (U + Bα,β )) cos (U − α (U + Bα,beta )) 1 E (cos (U )) α X = Sα,β arctan β tan πα 2 ) ( where Bα,β = α For α = 1, compute X , and Sα,β = 1 + β 2 tan2 πα 2 1 2α 1−α α , . π E cos U 2 π 2 = + β U tan U − β log π . π 2 2 + βU Generalize scale and location Y Yiyang Yang ( σ X + µ, = σ X + π2 βσ log σ + µ, α 6= 1 . α=1 Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Outline 1 Denition of α-Stable Random Variable Denition Characteristic Function 2 Weron's Algorithm for Generating Stable Random Variable Properties of Stable Law Integral Form of Stable Law Generation of Stable Random Variable 3 Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Given sample of observed data X = {x1 , x2 , · · · , xN } and assume it is directed by α-stable distribution.Then, dene the sample characteristic function φ̂X (u ) = N 1 X iuxj e . N j =1 By the law of large number, φ̂X (u ) is a consistent estimator of the charateristic function φX (u ). By simple transformation, we have for all α |φX (u )| = exp (−σ α |u |α ) . Thus − log |φX (u )| = σ α |u |α . Assume α 6= 1, choose two dierent nonzero values uk , k = 1, 2 − log φ̂X (uk ) = σ α |uk |α . Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Solve these two equations and get α̂, σ̂ log α̂ = u log log||φ̂(u21 )|| φ̂( ) log uu12 log |u1 | log − log φ̂ (u2 ) − log |u2 | log − log φ̂ (u1 ) log σ̂ = . log uu12 The estimation of β̂ and µ̂ based on the imaginary and real parts of the characteristic function Re Im πα , (φX (u )) = exp (− |σµ|α ) cos µu + |σ u |α β (signu ) tan 2 (φX (u )) = exp (− |σµ|α ) sin µu + |σ u |α β (signu ) tan πα Then, we have arctan (φX (u )) Re (φX (u )) Im = µu + |σ u |α β (signu ) tan Yiyang Yang πα 2 2 . Simulate α-Stable Random Variable . Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Based on α̂, σ̂ and two dierent nonzero values u , k = 3, 4, we can solve the system of equations to obtain estimation of β̂ and µ̂. k µ̂ = β̂ = u4α̂ arctan Im Re (φX (u3 )) (φX (u3 )) α̂ u3 u4 − u4 u3α̂ (φX (u3 )) (φX (u3 )) π α̂ σ̂ α̂ u4 arctan Im(φX (u4 )) − u3α̂ arctan Re (φX (u4 )) Im Re tan 2 Im(φX (u4 )) − u3 arctan Re (φ (u )) X 4 . u4 u3α̂ − u3 u4α̂ In this estimation, the values are u1 = 0.2, u2 = 0.8, u3 = 0.1 and u4 = 0.4 are proposed in the simulation study. Yiyang Yang Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Based on the equations log − log |φ (u )|2 = log 2σ α + α log |u | X πα Im (φ (u )) α = µu + σ α |u | β (signu ) tan . arctan Re (φ (u )) 2 X X By regression y = log − log |φ (u )|2 and w = log |u | in the model k X k k y = aw + b + k propose u = k πk 25 , k k k k = 1, 2, · · · , K , then α̂ = a, σ̂ = Yiyang Yang 1 e 2 b 1a . Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters By regression the model yl = Im(φX (ul )) 1 and Re (φX (ul )) ul arctan y = aw + b + l propose u = l The Method of Moments Time Regression Method πl 50 , l wl = |uul l| α signul in l l = 1, 2, · · · , L, then µ̂ = b , β̂ = a σ̂ α̂ tan π2α̂ . A very important step is normalization with initial estimation σ0 = x0.72 − x0.28 , µ0 = E [X0.25−0.75 ] 1.654 where x is the f sample quantile. Then the normalized data is f x = 0 j Yiyang Yang x − µ0 j σ0 . Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method Assume the estimated four paramters from regression of x are 0 j 0 0 0 0 α, β, σ, µ. Then the stable parameters of original data are α=α β=β 0 0 0 σ = σ σ0 ( µ= 0 µ σ0 + µ0 µ σ0 + µ0 − σβ π2 ln σ0 0 Yiyang Yang α 6= 1 α=1 Simulate α-Stable Random Variable Denition of α-Stable Random Variable Weron's Algorithm for Generating Stable Random Variable Estimation of Stable Distribution Parameters The Method of Moments Time Regression Method References J.M. Chambers, C.L. Mallows and B.W. Stuck, A Method for Simulating Stable Random Variables, Jounarl of the American Statistical Association 1976 Vol. 71, No. 354, pages 340-344 Rafel Weron, On the Chambers-Mallows-Stuck method for simulating skewed stable random variables, Probability Letters 28 (1996) 165-1771 Statistics & Rafel Weron, Correction to: On the Chambers-Mallows-Stuck method for simulating skewed stable random variables Ioannis A. Koutrouvelis, Regression-Type Estimation of the Parameters of Stable Laws, Journal of the American Statistical Association, December 1980 Svetlozar T. Rachev, Young Shin Kim, Michele Leonardo Bianchi, Frank J. Fabozzi, Financial Models with Levy Process and Volatility Clustering. Yiyang Yang Simulate α-Stable Random Variable