Time Series Analysis Spring 2015 Assignment 2 Solutions Kaiji Motegi Waseda University Reading: Chapter 5 of Enders (2010) Applied Econometric Time Series. Problem-1: In the midterm exam, we learned that a martingale difference sequence {yt } is always a white noise. The converse is not true, however. This problem constructs an example of white noise that is not a martingale difference sequence. Consider a bilinear process: i.i.d. ϵt ∼ (0, σ 2 ). yt = bϵt−1 yt−2 + ϵt , (a) Show that yt = b2 ϵt−1 ϵt−3 yt−4 + bϵt−1 ϵt−2 + ϵt . (b) Show that yt = b3 ϵt−1 ϵt−3 ϵt−5 yt−6 + b2 ϵt−1 ϵt−3 ϵt−4 + bϵt−1 ϵt−2 + ϵt . (c) Assume |b| < 1. Keep iterating to show that yt = ∞ ∑ bi ϵt−2i i=0 where ∏n j=1 zj ≡ z1 × z2 × · · · × zn . i ∏ ϵt−(2j−1) , (1) j=1 ∏0 j=1 ϵt−(2j−1) (d) Show that Et−1 [yt ] ≡ E[ yt | ϵt−1 , ϵt−2 , . . . ] = ∑∞ i=1 is understood as 1. bi ϵt−2i ∏i j=1 ϵt−(2j−1) . (Remark: This result indicates that {yt } is not a martingale difference sequence since Et−1 [yt ] ̸= 0.) (e) Using Eq. (1), show that E[yt ] = 0. (f) Assume that b2 σ 2 < 1. Show that E[yt2 ] = σ 2 /(1 − b2 σ 2 ). (Hint: Since {ϵt } is i.i.d., E[f (ϵt )g(ϵs )] = E[f (ϵt )]E[g(ϵs )] for any functions f (·) and g(·) whenever t ̸= s. For example, E[ϵ2t ϵs ] = E[ϵ2t ]E[ϵs ] = σ 2 × 0 = 0 whenever t ̸= s.) (g) Show that E[yt yt−j ] = 0 for all j ≥ 1. (Hint: Use the i.i.d. property of {ϵt } again.) (Remark: Parts (e), (f), and (g) indicate that {yt } is a white noise.) 1 Time Series Analysis Spring 2015 Assignment 2 Solutions Kaiji Motegi Waseda University Solution-1: (a) We have that yt = bϵt−1 (bϵt−3 yt−4 +ϵt−2 )+ϵt = b2 ϵt−1 ϵt−3 yt−4 +bϵt−1 ϵt−2 +ϵt . (b) Iterating back once more, we have that yt = b2 ϵt−1 ϵt−3 (bϵt−5 yt−6 + ϵt−4 )+bϵt−1 ϵt−2 +ϵt = b3 ϵt−1 ϵt−3 ϵt−5 yt−6 + b2 ϵt−1 ϵt−3 ϵt−4 + bϵt−1 ϵt−2 + ϵt . (c) Keep iterating to realize that yt = b0 ϵt−2×0 + b1 ϵt−2×1 (ϵt−1 ) + b2 ϵt−2×2 (ϵt−1 ϵt−3 ) + b3 ϵt−2×3 (ϵt−1 ϵt−3 ϵt−5 ) + . . . . Using the summation and product operators, this can be rewritten compactly as (1). (d) The first term of the right-hand side of Eq. (1), i.e. b0 ϵt−2×0 , contains ϵt . All other terms of the right-hand side contain past values of ϵ only (e.g. ϵt−1 ϵt−2 in the second term). Since Et−1 [ϵt ] = 0 and ϵt−1 , ϵt−2 , . . . are all known given the information set up to time t − 1, ∑ ∏i i we have that Et−1 [yt ] = ∞ i=1 b ϵt−2i j=1 ϵt−(2j−1) . (e) The unconditional expectation of each term of the right-hand side of Eq. (1) is zero due to the i.i.d. property of {ϵt }. Hence, E[yt ] = 0. (f) Consider E[yt2 ] = E[(ϵt + bϵt−1 ϵt−2 + b2 ϵt−1 ϵt−3 ϵt−4 + . . . )2 ]. The expectations of cross terms are all zeros due to the i.i.d. property of {ϵt }. Focusing on squared terms, we have that E[yt2 ] = σ 2 + b2 (σ 2 )2 + b4 (σ 2 )3 + · · · = σ 2 [1 + b2 σ 2 + b4 (σ 2 )2 + . . . ] = σ 2 /(1 − b2 σ 2 ). (g) Consider j = 1 first. We have that E[yt yt−1 ] = E[ (ϵt + bϵt−1 ϵt−2 + b2 ϵt−1 ϵt−3 ϵt−4 + . . . ) (ϵt−1 + bϵt−2 ϵt−3 + b2 ϵt−2 ϵt−4 ϵt−5 + . . . )]. The expectation of each term equals zero due to the i.i.d. property of {ϵt } (e.g. E[bϵt−1 ϵt−2 × ϵt−1 ] = E[bϵ2t−1 ϵt−2 ] = b × σ 2 × E[ϵt−2 ] = 0). The same argument holds for any j ≥ 2 as well. Problem-2: Consider bivariate time series {xt } with xt = [x1,t , x2,t ]′ . This problem has two similar goals. First, we construct an example where both {x1,t } and {x2,t } are covariance stationary but {xt } is not jointly covariance stationary. Second, keeping the same example, we show that a linear combination of individually covariance stationary series may not be covariance stationary. (a) Suppose that {bt } is independently and identically distributed mean-centered Bernoulli 2 Time Series Analysis Spring 2015 Assignment 2 Solutions random variables: bt = Kaiji Motegi Waseda University −1 with probability 0.5, 1 with probability 0.5. Show that E[bt ] = 0, E[b2t ] = 1, and E[bt bt−j ] = 0 for any j ≥ 1. (b) Assume that x1,2t = b3t and x1,2t+1 = b3t+1 for each t. Also assume that x2,2t = b3t and x2,2t+1 = b3t+2 for each t. These patterns can be summarized in the following table: t 1 2 3 4 5 6 7 ... x1,t b1 b3 b4 b6 b7 b9 b10 ... x2,t b2 b3 b5 b6 b8 b9 b11 ... Show that both {x1,t } and {x2,t } are covariance stationary. (c) Show that {xt } is not jointly covariance stationary. (Hint: Joint covariance stationarity requires that variance E[(xt − E[xt ])(xt − E[xt ])′ ] does not depend on t.) (Remark 1: As shown here, joint covariance stationarity is stronger than individual covariance stationarity.) (d) Consider a linear combination yt = β1 x1,t + β2 x2,t with β1 ̸= 0 and β2 ̸= 0. Show that {yt } is not covariance stationary. (Remark 2: We learned in class that any linear combination of jointly covariance stationary {xt } is covariance stationary. Problem-2 has shown that individual covariance stationarity does not guarantee the covariance stationarity of linear combination.) Solution-2: (a) E[bt ] = (−1) × 0.5 + 1 × 0.5 = 0. E[bt ] = (−1)2 × 0.5 + 12 × 0.5 = 1. E[bt bt−j ] = E[bt ]E[bt−j ] = 0 for any j ≥ 1 in view of the independence. (b) E[x1,t ] = 0, E[x21,t ] = 1, and E[x1,t x1,t−j ] = 0 for any j ≥ 1. {x1t } is therefore covariance stationary. Similarly, E[x2,t ] = 0, E[x22,t ] = 1, and E[x2,t x2,t−j ] = 0 for any j ≥ 1. {x2t } is therefore covariance stationary. 3 Time Series Analysis Spring 2015 Assignment 2 Solutions Kaiji Motegi Waseda University (c) E[(xt − E[xt ])(xt − E[xt ])′ ] is equal to I 2 if t is odd, while it is equal to a 2 × 2 matrix whose elements are all 1 if t is even. Thus, E[(xt − E[xt ])(xt − E[xt ])′ ] depends on time t. (d) We have that E[yt ] = 0 and E[(yt − E[yt ])2 ] = E[yt2 ] = β12 x21,t + 2β1 β2 x1,t x2,t + β22 x22,t . Hence, E[yt2 ] = β12 + β22 if t is odd, while E[yt2 ] = β12 + 2β1 β2 + β22 if t is even. Since β1 ̸= 0 and β2 ̸= 0, E[yt2 ] takes a different value for odd t and even t. Thus, {yt } is not covariance stationary. Problem-3: Consider a bivariate VAR(1) process: x1t 0.7 0.6 x1t−1 ϵ1t = + , x2t 0 0.8 x2t−1 ϵ2t | {z } | {z } | {z } | {z } ≡X t ≡A (2) ≡ϵt =X t−1 where {ϵt } is a joint white noise with 1 0.5 1 E[ϵt ϵ′t ] = = 0.5 1.25 0.5 | {z } | {z ≡Ω ≡L ′ 0 1 0 . 1 0.5 1 } | {z } =L′ (a) Show that the VAR(1) process appearing in Eq. (2) is covariance stationary. (b) Calculate VMA coefficients Θ0 , Θ1 , Θ2 , Θ3 , Θ4 , and Θ30 . (Recall xt = ∑∞ j=0 Θj η t−j , where {η(τL )} is a joint white noise with E[η(τL )η(τL )′ ] = I 2 .) (c) Implement the forecast error variance decomposition of {x1t } for horizons h = 1, 2, 3, 4, 30. (d) Implement the forecast error variance decomposition of {x2t } for horizons h = 1, 2, 3, 4, 30. Solution-3: (a) The eigenvalues of A are 0.6 and 0.7. Hence Eq. (2) is a covariance stationary VAR process. 4 Time Series Analysis Spring 2015 Assignment 2 Solutions Kaiji Motegi Waseda University (b) We have that Θj = Aj L for any j ≥ 0. Hence, 1 0 1 0.6 0.94 0.9 Θ0 = , Θ1 = , Θ2 = , 0.5 1 0.4 0.8 0.32 0.64 0.85 1.014 0.749 1.107 0.004 0.007 Θ3 = , Θ4 = , Θ30 = . 0.256 0.512 0.205 0.410 0.001 0.001 (c) See Table 1. Table 1: Forecast Error Variance Decomposition of {x1t } Share of η1 Share of η2 h=1 1 0 h=2 0.848 0.153 h=3 0.711 0.289 h=4 0.621 0.379 ... ... ... h = 30 0.445 0.555 (d) See Table 2. Table 2: Forecast Error Variance Decomposition of {x2t } Share of η1 Share of η2 h=1 0.2 0.8 h=2 0.2 0.8 h=3 0.2 0.8 5 h=4 0.2 0.8 ... ... ... h = 30 0.2 0.8