STAT520 Hwk2 Solution Courtesy of Wei Li Q2.1 Solution: Since {Xn } is a stationary, E(Xn ) = µ E(Xn Xn+h ) = γ(h) + µ2 = ρ(h)γ(0) + µ2 . 2 To minimize E (Xn+h − aXn − b) , we need to set ∂ 2 E (Xn+h − aXn − b) ⇒ a(γ(0) + µ2 ) + bµ = ρ(h)γ(0) + µ2 ∂a ∂ 2 0= E (Xn+h − aXn − b) ⇒ b = µ − aµ. ∂b 0= Solve the equations, we have a = ρ(h), b = µ(1 − ρ(h)). Therefore, a = ρ(h) and b = µ(1 − ρ(h)) minimize the MSE. The estimator X̂n+h = aXn + b is thence the best estimator. Q2.2 Solution: E(Xt ) = E(A cos(ωt) + B sin(ωt)) = E(A) cos(ωt) + E(B) sin(ωt) = 0 γ(h) = Cov(Xt , Xt+h ) = E(Xt Xt+h ) = E(A2 cos(ωt) cos(ωt + ωh) + AB cos(ωt) sin(ωt + ωh) + AB sin(ωt) cos(ωt + ωh) + B 2 sin(ωt) sin(ωt + ωh)) = cos(ωt) cos(ωt + ωh) + sin(ωt) sin(ωt + ωh) = cos(ωh) Therefore, {Xt } is a stationary series with mean zero and ACVF cos(ωh). Since κ(h) = cos(ωh) is the ACVF of a stationary time series, it follows from theorem 2.1.1 that cos(ωh) is non-negative definite. Remark: To show the non-negativity, it’s easier to construct a process with acvf being the given function. Q2.4 Solution: Use the result as above, we may construct the stationary series 1 (a) Xt = Z cos(πt), where Z ∼ W N (0, 1), t = 0, ±1, ±2... (b) Xt = A+B cos( π2 t)+C sin( π2 t)+D cos( π4 t)+E sin( π4 t), t = 0, ±1, ±2... where A, B, C, D, E are iid N (0, 1) (c) Xt = √25 Zt + √15 Zt−1 , t = 0, ±1, ±2... Zt ∼ W N (0, 1) Q2.8 Answer: Only need to show {Xt } non-stationary. Suppose otherwise, then γ = V ar(Xt ) = φ2 V ar(Xt−1 ) + σ 2 = φ2 γ + σ 2 = γ + σ 2 Note σ 2 > 0, then we must have γ = V ar(Xt ) = ∞. This fact contradicts its stationarity. Remark: This is a convenient way to show there is no stationary solution. If you follow the hint in the textbook, you still need to show V ar(Xt ) = ∞ eventually. QA 1. Solution: R̂101 = E(Z101 + 0.2Z100 |Z100 ) = E(Z101 + 0.2 × 0.01) = 0.002 R̂102 = E(Z102 + 0.2Z101 |Z100 ) = E(Z102 ) + 0.2E(Z101 ) = 0. Clearly, by definition the forecast error ê101 = Z101 and the forecast error ê102 = Z102 + 0.2Z101 . Thus, the standard deviation of the forecast errors are sd(ê101 ) = 0.025 sd(ê102 ) = 0.025 × √ 1.04 = 0.0255 the autocorrelations are cov(Rt , Rt+1 ) = 0.1923 V ar(Rt ) cov(Rt , Rt+2 ) ρ(2) = = 0. V ar(Rt ) ρ(1) = QA 2. Solution: Similarly, plug in, we have R̂101 = 0.014 R̂102 = 0.008 sd(ê101 ) = 0.1414 sd(ê102 ) = 0.1414 ρ(1) = 0 ρ(2) = 0.2 E(Rt ) = 0.0125 V ar(Rt ) = 0.0208 2