Segmenting Multiple Time Series by Contemporaneous Linear Transformation: PCA for time series Qiwei Yao Department of Statistics, London School of Economics q.yao@lse.ac.uk Joint work with Jinyuan Chang Bin Guo University of Melbourne Peking University – p.1 Goal of study: a high-dim TS → several uncorrelated lower-dim TS Methodology: PCA for time series • Transformation via an eigenanalysis • Permutation – maximum cross correlations – FDR based on multiple tests Real data illustration Asymptotic properties in 3 settings: p fixed, p = o(nc ), log p = o(nc ) Simulation Segmenting multiple volatility processes – p.2 Goal: For p × 1 weakly stationary time series yt , search for a contemporaneous linear transformation: yt = Axt , or xt = Byt (i.e. B = A−1 ) such that xt = (1) (1) xt .. . (q) xt , (i) (j) Cov(xt , xs ) = 0 ∀ i 6= j and t, s. (q) Hence, xt , · · · , xt can be modelled separately! – p.3 Goal: For p × 1 weakly stationary time series yt , search for a contemporaneous linear transformation: yt = Axt , or xt = Byt (i.e. B = A−1 ) such that xt = (1) xt .. . (q) xt (1) , (i) (j) Cov(xt , xs ) = 0 ∀ i 6= j and t, s. (q) Hence, xt , · · · , xt can be modelled separately! • realistic? RealData • how to find B and xt ? – p.3 Observations: y1 , · · · , yn from a p×1 weakly stationary time series Assumption: yt = Axt , and xt = (1) xt .. . (q) xt , (i) (j) Cov(xt , xs ) = 0 ∀ i 6= j and t, s. Without loss of generality: Var(yt ) = Var(xt ) = Ip , and thus A′ A = Ip , i.e. A is orthogonal, xt = A′ yt Goal: estimate A = (A1 , · · · , Aq ), or more precisely, M(A1 ), · · · , M(Aq ), as (j) x bt b ′ yt , j = 1, · · · , q . =A j Note. (A, xt ) can be replaced by (AH, H′ xt ) for any H = diag(H1 , · · · , Hq ) with H′j Hj = Ipj – p.4 Step 1: Transformation via eigenanalysis Notation: Σy (k) = Cov(yt+k , yt ), Wy = Wx = k0 X Σx (k) = Cov(xt+k , xt ), Σy (k)Σy (k)′ , k=0 k0 X k0 X k0 X Σx (k)Σx (k)′ . k=0 Σy (k)Σy (k)′ = Ip + k=1 Σx (k)Σx (k)′ = Ip + k=1 – p.5 Step 1: Transformation via eigenanalysis Notation: Σy (k) = Cov(yt+k , yt ), Wy = Wx = k0 X Σx (k) = Cov(xt+k , xt ), Σy (k)Σy (k)′ , k=0 k0 X k0 X k0 X Σx (k)Σx (k)′ . Σy (k)Σy (k)′ = Ip + k=1 Σx (k)Σx (k)′ = Ip + k=0 As Σy (k) = AΣx (k)A′ , k=1 Wy = AWx A′ . – p.5 Step 1: Transformation via eigenanalysis Notation: Σy (k) = Cov(yt+k , yt ), Wy = Wx = k0 X Σx (k) = Cov(xt+k , xt ), Σy (k)Σy (k)′ , k=0 k0 X k0 X k0 X Σx (k)Σx (k)′ . Σy (k)Σy (k)′ = Ip + k=1 Σx (k)Σx (k)′ = Ip + k=0 As Σy (k) = AΣx (k)A′ , k=1 Wy = AWx A′ . Eigenanalysis: Wx Γx = Γx D, columns of Γx are eigenvectors of Wx with the eigenvalues in diagonal matrix D. Wy AΓx = AWx A′ AΓx = AWx Γx = AΓx D Thus Γy = AΓx , and Γ′y yt = Γ′x A′ yt = Γ′x xt . – p.5 zt ≡ Γ′y yt (= Γ′x xt ) is the required transformation upto a permutation – p.6 zt ≡ Γ′y yt (= Γ′x xt ) is the required transformation upto a permutation Note Wx = diag(Wx,1 , · · · , Wx,q ) Proposition 1. (i) Γx can be taken with the same block-diagonal structure as Wx . (ii) Any Γx is a column-permutation of Γx described in (i), provided λ(Wx,i ) 6= λ(Wx,j ) for any i 6= j. – p.6 zt ≡ Γ′y yt (= Γ′x xt ) is the required transformation upto a permutation Note Wx = diag(Wx,1 , · · · , Wx,q ) Proposition 1. (i) Γx can be taken with the same block-diagonal structure as Wx . (ii) Any Γx is a column-permutation of Γx described in (i), provided λ(Wx,i ) 6= λ(Wx,j ) for any i 6= j. For Γx specified in (i), Γ′x xt is segmented exactly the same as xt – p.6 zt ≡ Γ′y yt (= Γ′x xt ) is the required transformation upto a permutation Note Wx = diag(Wx,1 , · · · , Wx,q ) Proposition 1. (i) Γx can be taken with the same block-diagonal structure as Wx . (ii) Any Γx is a column-permutation of Γx described in (i), provided λ(Wx,i ) 6= λ(Wx,j ) for any i 6= j. For Γx specified in (i), Γ′x xt is segmented exactly the same as xt Let c y = Ip + W k0 X k=1 b y (k)Σ b y (k)′ , Σ cyΓ by = Γ b y D. b W ′ b bt zt to obtain x zt = Γy yt – require permute components of b Then b – p.6 Permutation ′ b Goal: put the connected components of b zt = Γy yt together. Visual examination of CCF if p is not large! Two component series of b zt is connected if the multiple null hypothesis H0 : ρ(k) = 0 for any k = 0, ±1, ±2, . . . , ±m is rejected, where ρ(k) is cross correlation between two series at lag k. Permutation is performed as follows: i. Start with p groups: each containing one component of b zt . ii. Combine the two groups together if one connected pair are split over the two groups. iii. Repeat Step ii above until all connected components are within one group. – p.7 Permutation Method I: Max CCF Put b zt = (zb1,t , · · · , zbp,t )′ . Let ρbi,j (h) be the sample CCF of (zbi,t , zbj,t ) at lag h, and bn (i, j) = max |ρbi,j (h)|, L |h|≤m bn (i, j). reject H0 for the pair (zbi,t , zbj,t ) for large values of L – p.8 Permutation Method I: Max CCF Put b zt = (zb1,t , · · · , zbp,t )′ . Let ρbi,j (h) be the sample CCF of (zbi,t , zbj,t ) at lag h, and bn (i, j) = max |ρbi,j (h)|, L |h|≤m bn (i, j). reject H0 for the pair (zbi,t , zbj,t ) for large values of L bn (i, j), 1 ≤ i < j ≤ p, in the descending order: Line up L Define b1 ≥ · · · ≥ L b p0 , L p0 = p(p − 1)/2. bj /L bj+1 , rb = arg max L 1≤j<c0 p0 c0 ∈ (0, 1). b1 , · · · , L brb Reject H0 for the pairs corresponding to L – p.8 Graph representation: Let vertexes Vb = {1, · · · , p} stand for the ′ b components of b zt = Γ yt , and bn = {edge connecting i and j : zbi,t , zbj,t are connected}. E Let V = {1, · · · , p} stand for the components of zt = Γ′ yt , and E = {edge connecting i and j : max −m≤h≤m |Corr(zi,t+h , zj,t )| > 0 }. – p.9 Graph representation: Let vertexes Vb = {1, · · · , p} stand for the ′ b components of b zt = Γ yt , and bn = {edge connecting i and j : zbi,t , zbj,t are connected}. E Let V = {1, · · · , p} stand for the components of zt = Γ′ yt , and E = {edge connecting i and j : max |Corr(zi,t+h , zj,t )| > 0 }. Let ̟ = mini6=j min |λ(Wx,i ) − λ(Wx,j )|, ⇒ max max |Corr(zi,t+h , zj,t )| > 0, 1≤i<j≤p |h|≤m −m≤h≤m min max |Corr(zi,t+h , zj,t )| = 0, 1≤i<j≤p |h|≤m b j + δn )/(L b j+1 + δn ). rb = arg max (L 1≤j<p0 Proposition 2. As n → ∞, δn → 0, 1 c ̟ kWy − Wy k2 = o(δn ), and bn = E) → 1. log p = o(nα ). Then P (E – p.9 Prewhitening To make CCF for different pairs comparable, prewhiten each ′ b component series of b zt = Γy yt separately first. – p.10 Prewhitening To make CCF for different pairs comparable, prewhiten each ′ b component series of b zt = Γy yt separately first. (i) If ρi,j (h) = 0, ρbi,j (h) ∼ N (0, 1/n) asymptotically, provided at least one of xi,t and xj,t is white noise. (ii) For h 6= k , ρbi,j (h), ρbi,j (k) are asymptotically independent, and Cov{ρbi,j (h), ρbi,j (k)} = oP (1/n), provided both xi,t and xj,t are white noise. Brockwell & Davis (1996, Corollary 7.3.1). – p.10 Prewhitening To make CCF for different pairs comparable, prewhiten each ′ b component series of b zt = Γy yt separately first. (i) If ρi,j (h) = 0, ρbi,j (h) ∼ N (0, 1/n) asymptotically, provided at least one of xi,t and xj,t is white noise. (ii) For h 6= k , ρbi,j (h), ρbi,j (k) are asymptotically independent, and Cov{ρbi,j (h), ρbi,j (k)} = oP (1/n), provided both xi,t and xj,t are white noise. Brockwell & Davis (1996, Corollary 7.3.1). In practice, we filter out the autocorrelation for each component series of b zt by fitting an AR with the order determined by AIC and not greater than 5. – p.10 Permutation Method II: FDR based on multiple tests To fix the idea, let ξt , ηt be two WN, ρ(k) = Corr(ξt+k , ηt ) = 0, ρb(k) = n−k X t=1 ¯ t − η̄) (ξt+k − ξ)(η n .n X t=1 (ξt − ξ̄) 2 n X t=1 (ηt − η̄) 2 o1/2 . Asymptotically ρb(k) ∼ N (0, 1/n), and ρb(k), ρb(h) are independent. – p.11 Permutation Method II: FDR based on multiple tests To fix the idea, let ξt , ηt be two WN, ρ(k) = Corr(ξt+k , ηt ) = 0, ρb(k) = n−k X t=1 ¯ t − η̄) (ξt+k − ξ)(η n .n X t=1 (ξt − ξ̄) 2 n X t=1 (ηt − η̄) 2 o1/2 . Asymptotically ρb(k) ∼ N (0, 1/n), and ρb(k), ρb(h) are independent. √ The P -value for testing ρ(k) = 0 (simple) is pk = 2Φ(− n|b ρ(k)|) – p.11 Permutation Method II: FDR based on multiple tests To fix the idea, let ξt , ηt be two WN, ρ(k) = Corr(ξt+k , ηt ) = 0, ρb(k) = n−k X t=1 ¯ t − η̄) (ξt+k − ξ)(η n .n X t=1 (ξt − ξ̄) 2 n X t=1 (ηt − η̄) 2 o1/2 . Asymptotically ρb(k) ∼ N (0, 1/n), and ρb(k), ρb(h) are independent. √ The P -value for testing ρ(k) = 0 (simple) is pk = 2Φ(− n|b ρ(k)|) Let p(1) ≤ · · · ≤ p(2m+1) be the order statistics of {pk , |k| ≤ m}. Simes (1986): For H0 : ρ(k) = 0, ∀ |k| ≤ m, a multiple test rejects H0 at the level α if p(j) ≤ jα/(2m + 1) for at least one 1 ≤ j ≤ 2m + 1 – p.11 Permutation Method II: FDR based on multiple tests To fix the idea, let ξt , ηt be two WN, ρ(k) = Corr(ξt+k , ηt ) = 0, ρb(k) = n−k X t=1 ¯ t − η̄) (ξt+k − ξ)(η n .n X t=1 (ξt − ξ̄) 2 n X t=1 (ηt − η̄) 2 o1/2 . Asymptotically ρb(k) ∼ N (0, 1/n), and ρb(k), ρb(h) are independent. √ The P -value for testing ρ(k) = 0 (simple) is pk = 2Φ(− n|b ρ(k)|) Let p(1) ≤ · · · ≤ p(2m+1) be the order statistics of {pk , |k| ≤ m}. Simes (1986): For H0 : ρ(k) = 0, ∀ |k| ≤ m, a multiple test rejects H0 at the level α if p(j) ≤ jα/(2m + 1) for at least one 1 ≤ j ≤ 2m + 1 The P -value for the multiple test is P = min{α > 0 : p(j) ≤ jα/(2m + 1) for some 1 ≤ j ≤ 2m + 1} = min 1≤j≤2m+1 p(j) (2m + 1)/j. – p.11 b ′ yt , we test multiple hypothesis For each pair components of b zt = Γ y H0 , obtaining P -value Pi,j for 1 ≤ i < j ≤ q. Arranging those P -values in ascending order: P(1) ≤ P(2) ≤ · · · ≤ P(p0 ) , p0 = p(p − 1)/2 FDR: For a given small β ∈ (0, 1), let db = max{k : 1 ≤ k ≤ p0 , P(k) ≤ kβ/p0 }, and rejects the hypothesis H0 for the db pairs of the components of zt corresponding to the P -values P(1) , · · · , P(d) b – p.12 b ′ yt , we test multiple hypothesis For each pair components of b zt = Γ y H0 , obtaining P -value Pi,j for 1 ≤ i < j ≤ q. Arranging those P -values in ascending order: P(1) ≤ P(2) ≤ · · · ≤ P(p0 ) , p0 = p(p − 1)/2 FDR: For a given small β ∈ (0, 1), let db = max{k : 1 ≤ k ≤ p0 , P(k) ≤ kβ/p0 }, and rejects the hypothesis H0 for the db pairs of the components of zt corresponding to the P -values P(1) , · · · , P(d) b (i) The P -values pk (|k| < m) are asymptotically independent (ii) The P -values Pi,j (1 ≤< j ≤ p) are not independent, causing difficulties in choosing β in FDR. (iii) Ranking the pairwise dependences among the components of b zt . – p.12 Real data examples Example 1. Monthly temperatures in 1954 - 1998 in p = 7 cities in Eastern China. 32.5 Dongtai • Shanghai • 30.5 31.0 31.5 Hefei • Huoshan • Anqing • Hangzhow • 30.0 Latitude 32.0 Nanjing • 116 118 120 122 Longitude – p.13 0 15 30 Time series plots of monthly temperatures in 7 cities 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 0 15 30 0 15 30 0 10 25 0 15 30 0 15 30 0 15 30 1954 – p.14 CCF of monthly temperatures in 7 cities −5 −10 −5 0 −15 −10 −5 −15 −10 −5 5 1.0 1.0 −1.0 0.0 15 −15 −10 −15 −10 0 15 −5 −5 5 5 10 15 0 5 10 15 10 15 1.0 Y5 & Y7 0 5 10 15 0 5 10 15 Y6 & Y7 0 5 10 15 0 5 Y7 & Y6 0 15 1.0 0 Y6 0 10 Y4 & Y7 Y5 & Y6 10 5 1.0 −1.0 0.0 10 −1.0 0.0 15 Y7 & Y5 0 −1.0 0.0 1.0 −1.0 0.0 1.0 5 1.0 0 0 Y3 & Y7 1.0 −1.0 0.0 0 Y6 & Y5 0 15 −1.0 0.0 −5 10 1.0 10 15 1.0 −10 5 Y4 & Y6 −1.0 0.0 5 Y7 & Y4 0 15 1.0 −15 0 Y5 1.0 −15 0 Y6 & Y4 0 10 10 Y2 & Y7 −1.0 0.0 −5 5 5 10 15 Y7 1.0 −10 −1.0 0.0 1.0 1.0 −1.0 0.0 15 1.0 −15 0 0 Y3 & Y6 −1.0 0.0 0 15 Y4 & Y5 10 15 1.0 −5 10 1.0 5 10 Y2 & Y6 1.0 −1.0 0.0 15 −1.0 0.0 0 Y7 & Y3 0 10 5 −15 −10 −5 0 −1.0 0.0 −10 5 Y6 & Y3 0 −1.0 0.0 1.0 −1.0 0.0 1.0 1.0 −10 5 Y5 & Y4 1.0 −15 0 0 Y3 & Y5 1.0 −15 0 Y4 0 15 −1.0 0.0 −5 15 Y1 & Y7 1.0 −10 10 1.0 −5 10 −1.0 0.0 0 1.0 0 −1.0 0.0 −5 −10 5 1.0 −5 Y7 & Y2 1.0 −10 15 1.0 −15 Y7 & Y1 −15 10 1.0 −10 5 Y5 & Y3 1.0 0 −1.0 0.0 −5 −15 Y6 & Y2 1.0 −10 5 0 Y2 & Y5 −1.0 0.0 0 15 −1.0 0.0 −5 0 Y4 & Y3 −1.0 0.0 −15 Y6 & Y1 −15 0 Y1 & Y6 1.0 −10 10 Y3 & Y4 1.0 −15 5 −1.0 0.0 0 1.0 0 −1.0 0.0 −5 15 −1.0 0.0 −5 Y5 & Y2 1.0 −10 10 −1.0 0.0 −10 −1.0 0.0 0 Y5 & Y1 −15 5 1.0 −15 0 Y3 1.0 −5 −1.0 0.0 −10 0 −1.0 0.0 15 Y1 & Y5 Y2 & Y4 −1.0 0.0 10 Y4 & Y2 1.0 −1.0 0.0 −15 15 −1.0 0.0 5 −1.0 0.0 0 10 1.0 0 1.0 −5 −1.0 0.0 −10 5 Y3 & Y2 1.0 −1.0 0.0 −15 0 Y2 & Y3 −1.0 0.0 0 Y4 & Y1 −1.0 0.0 15 1.0 −5 −1.0 0.0 −10 Y3 & Y1 −1.0 0.0 10 −1.0 0.0 5 Y2 1.0 −1.0 0.0 −15 Y1 & Y4 1.0 0 −1.0 0.0 15 −1.0 0.0 10 −1.0 0.0 5 −1.0 0.0 −1.0 0.0 0 Y2 & Y1 −1.0 0.0 Y1 & Y3 1.0 Y1 & Y2 1.0 Y1 0 5 10 15 – p.15 b t bt = By Transformation: x 0.244 −0.066 0.0187 −0.050 −0.313 −0.154 0.200 −0.703 0.324 −0.617 0.189 0.633 0.499 −0.323 0.375 1.544 −1.615 0.170 −2.266 0.126 1.596 b B = 3.025 −1.381 −0.787 −1.691 −0.212 1.188 −0.165 −0.197 −1.820 −1.416 3.269 .301 −1.438 1.299 −0.584 −0.354 0.847 −1.262 −0.218 −0.151 1.831 1.869 −0.742 0.034 0.501 0.492 −2.533 0.339 ′ b y (0)−1/2 . b b Note. B = Γy Σ n = 540, p = 7 Used k0 = 5 (in defining Wy ), hardly changed for 2 ≤ k0 ≤ 36. – p.16 −4 −2 0 Time series plots for transformed monthly temperatures 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 1954 1959 1964 1969 1974 1979 1984 1989 1994 1999 −5 −2 1 0 2 4 6 −4 2 6 0 4 −4 −1 1 −1 1 1954 – p.17 CCF for transformed monthly temperatures −5 −10 −5 0 −15 −10 −5 −15 −10 −5 5 1.0 1.0 −1.0 0.0 15 −15 −10 Segmentation: {1, 2, 3}, {4}, {5}, {6}, {7} −15 −10 0 15 −5 −5 5 5 10 15 0 5 10 15 10 15 1.0 X5 & X7 0 5 10 15 0 5 10 15 X6 & X7 0 5 10 15 0 5 X7 & X6 0 15 1.0 0 X6 0 10 X4 & X7 X5 & X6 10 5 1.0 −1.0 0.0 10 −1.0 0.0 15 X7 & X5 0 −1.0 0.0 1.0 −1.0 0.0 1.0 5 1.0 0 0 X3 & X7 1.0 −1.0 0.0 0 X6 & X5 0 15 −1.0 0.0 −5 10 1.0 10 15 1.0 −10 5 X4 & X6 −1.0 0.0 5 X7 & X4 0 15 1.0 −15 0 X5 1.0 −15 0 X6 & X4 0 10 10 X2 & X7 −1.0 0.0 −5 5 5 10 15 X7 1.0 −10 −1.0 0.0 1.0 1.0 −1.0 0.0 15 1.0 −15 0 0 X3 & X6 −1.0 0.0 0 15 X4 & X5 10 15 1.0 −5 10 1.0 5 10 X2 & X6 1.0 −1.0 0.0 15 −1.0 0.0 0 X7 & X3 0 10 5 −15 −10 −5 0 −1.0 0.0 −10 5 X6 & X3 0 −1.0 0.0 1.0 −1.0 0.0 1.0 1.0 −10 5 X5 & X4 1.0 −15 0 0 X3 & X5 1.0 −15 0 X4 0 15 −1.0 0.0 −5 15 X1 & X7 1.0 −10 10 1.0 −5 10 −1.0 0.0 0 1.0 0 −1.0 0.0 −5 −10 5 1.0 −5 X7 & X2 1.0 −10 15 1.0 −15 X7 & X1 −15 10 1.0 −10 5 X5 & X3 1.0 0 −1.0 0.0 −5 −15 X6 & X2 1.0 −10 5 0 X2 & X5 −1.0 0.0 0 15 −1.0 0.0 −5 0 X4 & X3 −1.0 0.0 −15 X6 & X1 −15 0 X1 & X6 1.0 −10 10 X3 & X4 1.0 −15 5 −1.0 0.0 0 1.0 0 −1.0 0.0 −5 15 −1.0 0.0 −5 X5 & X2 1.0 −10 10 −1.0 0.0 −10 −1.0 0.0 0 X5 & X1 −15 5 1.0 −15 0 X3 1.0 −5 −1.0 0.0 −10 0 −1.0 0.0 15 X1 & X5 X2 & X4 −1.0 0.0 10 X4 & X2 1.0 −1.0 0.0 −15 15 −1.0 0.0 5 −1.0 0.0 0 10 1.0 0 1.0 −5 −1.0 0.0 −10 5 X3 & X2 1.0 −1.0 0.0 −15 0 X2 & X3 −1.0 0.0 0 X4 & X1 −1.0 0.0 15 1.0 −5 −1.0 0.0 −10 X3 & X1 −1.0 0.0 10 −1.0 0.0 5 X2 1.0 −1.0 0.0 −15 X1 & X4 1.0 0 −1.0 0.0 15 −1.0 0.0 10 −1.0 0.0 5 −1.0 0.0 −1.0 0.0 0 X2 & X1 −1.0 0.0 X1 & X3 1.0 X1 & X2 1.0 X1 0 5 10 15 – p.18 • Visual examination of CCF: {1, 2, 3}, {4}, {5}, {6} and {7} • Permutation based on max-CCF: the same grouping with 2 ≤ m ≤ 30 • Permutation based on FDR: the same grouping with 2 ≤ m ≤ 30 and β ∈ [0.001%, 1%]. 7-dim time series → 5 uncorrelated time series Methodology – p.19 Post-Sample Forecast Forecasting based on segmentation: fit each subseries of xt with a VAR model, forecast xt based on the fitted models, the b −1 xt . forecasts for yt are obtained via yt = B Compare with the forecasts based on fitting a VAR and a restricted VAR (RVAR) directly to yt . Implementation: using VAR in R-package vars For each of the last 24 values (i.e. the monthly temperatures in 1997-1998), we use the data upto the previous month for the fittings. We calculate MSE of one-step-ahead forecasts, two-step-ahead forecasts (by plug-in) for each of 7 cities. One-step MSE Two-step MSE VAR 1.669(2.355) 1.815(2.372) 1.677(2.324) 1.829(2.398) RVAR Segmentation 1.381(1.888) 1.543(1.874) – p.20 Example 2. 8 monthly US Industrial Production indices in 1947-1993 published by the US Federal Reserve. 8 indices: total index, manufacturing index, durable manufacturing, nondurable manufacturing, mining, utilities, products, materials. Nonstationary trends: difference each series – p.21 −3 0 2 8 differenced monthly US Industrial Indices 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 −4 0 4 −2 0 2 −5 5 −5 5 −2 0 2 −3 0 3 −3 0 2 1947 – p.22 CCF of 8 differenced monthly US Industrial Indices −15 −5 −15 −5 0 −15 −5 −15 0 −15 −15 0 −5 −5 −5 0 5 −15 15 −15 10 15 −15 0 5 10 0 5 −5 0 −5 −5 0 5 5 0 5 −15 −15 −0.2 0.6 −0.2 0.6 −0.2 0.6 15 0 10 15 10 0 5 10 15 0 5 −5 0 0 0 −5 0 5 5 5 5 5 15 0 5 −15 15 10 15 10 15 10 15 10 15 Y7 & Y8 10 15 0 5 Y8 & Y7 0 10 Y6 & Y8 Y7 0 15 Y5 & Y8 15 10 10 Y4 & Y8 15 10 5 Y3 & Y8 Y6 & Y7 Y8 & Y6 0 15 Y5 & Y7 Y7 & Y6 0 10 Y4 & Y7 15 10 5 −0.2 0.6 0 Y6 Y8 & Y5 0 15 Y5 & Y6 Y7 & Y5 0 0 Y4 & Y6 Y6 & Y5 0 10 0 Y2 & Y8 −0.2 0.6 5 15 Y3 & Y7 −0.2 0.6 0 10 −0.2 0.6 −5 −0.2 0.6 −0.2 0.6 −0.2 0.6 15 10 15 −0.2 0.6 5 10 −0.2 0.6 0 5 5 Y2 & Y7 Y3 & Y6 Y5 Y8 & Y4 0 −0.2 0.6 −0.2 0.6 15 Y7 & Y4 Y8 & Y3 0 10 Y6 & Y4 Y7 & Y3 0 −15 0 Y4 & Y5 Y5 & Y4 0 10 0 −0.2 0.6 −5 5 5 15 −0.2 0.6 −15 0 15 −0.2 0.6 0 Y6 & Y3 0 0 10 −0.2 0.6 −5 15 Y4 −0.2 0.6 −5 10 10 −0.2 0.6 5 5 −0.2 0.6 0 −0.2 0.6 −15 0 5 Y2 & Y6 Y3 & Y5 −0.2 0.6 15 Y5 & Y3 0 −0.2 0.6 −0.2 0.6 −0.2 0.6 10 0 Y1 & Y8 −5 10 15 Y8 0 −0.2 0.6 0 −15 15 Y3 & Y4 −0.2 0.6 0 10 15 −0.2 0.6 −5 5 10 Y2 & Y5 Y1 & Y7 −0.2 0.6 −15 0 5 −0.2 0.6 −5 5 0 Y1 & Y6 −0.2 0.6 −15 0 15 −0.2 0.6 −5 15 −0.2 0.6 −5 10 −0.2 0.6 5 10 Y2 & Y4 Y4 & Y3 Y8 & Y2 −0.2 0.6 −0.2 0.6 −5 0 5 −0.2 0.6 0 Y8 & Y1 −15 0 Y7 & Y2 −0.2 0.6 −0.2 0.6 −5 −15 0 Y1 & Y5 −0.2 0.6 0 Y7 & Y1 −15 −5 Y6 & Y2 −0.2 0.6 −0.2 0.6 −5 −15 15 −0.2 0.6 0 Y6 & Y1 −15 −15 10 Y3 Y5 & Y2 −0.2 0.6 −0.2 0.6 −5 15 −0.2 0.6 0 Y5 & Y1 −15 10 Y4 & Y2 −0.2 0.6 −0.2 0.6 −5 5 −0.2 0.6 0 Y4 & Y1 −15 0 5 Y3 & Y2 −0.2 0.6 −0.2 0.6 −5 0 Y2 & Y3 −0.2 0.6 0 Y3 & Y1 −15 15 −0.2 0.6 −5 −0.2 0.6 −0.2 0.6 −15 10 Y2 −0.2 0.6 5 Y1 & Y4 −0.2 0.6 0 −0.2 0.6 15 −0.2 0.6 10 Y2 & Y1 −0.2 0.6 5 Y1 & Y3 −0.2 0.6 0 Y1 & Y2 −0.2 0.6 −0.2 0.6 Y1 0 5 10 15 – p.23 5.012 10.391 −6.247 1.162 b = B 6.172 0.868 3.455 0.902 b t bt = By Transformation: x −1.154 −0.472 −0.880 −0.082 −0.247 −2.69 8.022 −3.981 −3.142 0.186 0.019 −6.949 11.879 −4.8845 −4.0436 0.289 −0.011 2.557 −6.219 3.163 1.725 0.074 −0.823 0.646 −4.116 2.958 1.887 0.010 0.111 −2.542 1.023 −2.946 −4.615 −0.271 −0.354 3.972 −2.744 5.557 3.165 0.753 0.725 −2.331 −2.933 −1.750 −0.123 0.191 −0.265 3.759 −1.463 −4.203 0.243 −0.010 . −3.961 1.902 −1.777 0.987 – p.24 −2 2 Transformed differenced monthly US Industrial Indices 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 1947 1952 1957 1962 1967 1972 1977 1982 1987 1992 −4 0 4 −4 0 4 −4 0 3 −4 0 −4 0 4 −3 0 3 −3 0 1947 – p.25 CCF of transformed differenced monthly US Industrial Indices −15 −5 15 −15 −15 10 −5 −5 5 15 −5 0 0.6 −0.4 0.6 10 15 Segmentation: {1, 2, 3}, {4, 8}, {5}, {6}, {7} −5 5 0 5 10 15 10 15 0.6 0 5 10 15 0 5 10 15 0.6 X6 & X8 0 5 10 15 0 5 10 15 0.6 X7 & X8 0 5 10 15 0 5 X8 & X7 0 15 X5 & X8 10 15 X8 0.6 −15 10 X4 & X8 0.6 −15 0 X7 X8 & X6 0 15 X6 & X7 10 5 0.6 5 X7 & X6 0 10 0.6 0 0 X5 & X7 15 15 0.6 0 X6 0 5 −0.4 0 −0.4 5 10 X3 & X8 0.6 0 5 0.6 −5 15 −0.4 10 −0.4 −15 10 −0.4 15 5 X4 & X7 −0.4 10 −0.4 0.6 15 X5 & X6 X8 & X5 0 −0.4 0.6 5 0 X2 & X8 0.6 0 X7 & X5 0 10 0.6 5 15 −0.4 15 −0.4 0 0 X4 & X6 −0.4 10 10 0.6 5 −0.4 0 0.6 −5 −0.4 0.6 −0.4 0.6 5 5 X3 & X7 −0.4 0 15 0.6 −5 10 0.6 0 0 X2 & X7 0.6 −0.4 15 0.6 −15 5 X6 & X5 X8 & X4 0 10 X1 & X8 −15 −5 0 −0.4 −5 5 0.6 −15 0 X5 0 15 0.6 −5 10 −0.4 −15 0 0.6 −15 5 X3 & X6 −0.4 15 X7 & X4 0 15 −0.4 10 −0.4 0 10 0.6 −5 5 −0.4 5 0.6 −15 0 0.6 0 0 X4 & X5 −0.4 −5 −0.4 0.6 0.6 −0.4 15 X6 & X4 X8 & X3 0 10 X1 & X7 X2 & X6 0.6 5 −0.4 0 0.6 −15 15 X3 & X5 −0.4 0 10 −0.4 −5 15 −0.4 −5 X7 & X3 0 10 0.6 −15 5 X5 & X4 0.6 −15 5 0.6 −5 0 −0.4 0 0.6 −15 −0.4 0.6 −0.4 0.6 −5 X1 & X6 X2 & X5 0.6 −15 0 X4 −0.4 0 15 0.6 15 0.6 −5 −0.4 −15 0.6 0 −0.4 −5 10 X6 & X3 X8 & X2 0.6 −15 5 −0.4 0 0.6 0 −0.4 −5 X8 & X1 0 0.6 −5 −0.4 −15 10 X3 & X4 X5 & X3 X7 & X2 0.6 −15 15 −0.4 0 0.6 0 −0.4 −5 X7 & X1 10 0.6 −5 −0.4 −15 5 X4 & X3 X6 & X2 0.6 X6 & X1 −15 5 −0.4 0 0.6 0 −0.4 −5 0 −0.4 −5 X1 & X5 X2 & X4 0.6 −15 0 X3 X5 & X2 0.6 −15 15 −0.4 15 0.6 0 −0.4 −5 10 0.6 10 −0.4 5 X4 & X2 X5 & X1 −0.4 0 −0.4 0 0.6 −0.4 −15 5 X2 & X3 0.6 −5 −0.4 0.6 −0.4 −15 0 X3 & X2 X4 & X1 −0.4 15 0.6 0 −0.4 −5 X3 & X1 −0.4 10 −0.4 5 X2 0.6 −0.4 −15 X1 & X4 0.6 0 −0.4 15 −0.4 10 −0.4 5 −0.4 −0.4 0 X2 & X1 −0.4 X1 & X3 0.6 X1 & X2 0.6 X1 0 5 10 15 – p.26 • Visual examining CCF: {1, 2, 3}, {4, 8} • Permutation with max-CCF 1 ≤ m ≤ 20: {1, 3} • Permutation with FDR m = 20 and β ∈ [10−6 , 0.01], or m = 5 and β ∈ [10−6 , 0.001]: {1, 3} m = 5 and β = 0.005: m = 5 and β = 0.01: {1, 2, 3}, {4, 8} {1, 2, 3, 5, 6, 7} and {4, 8}. Two recommended groupings: seven groups: {1, 3} five groups: {1, 2, 3}, {4, 8} Methodology – p.27 Post-sample forecast Forecast 24 monthly indices in Jan 1992 – Dec 1993. Using the segmentation: {1, 3} and other six single element groups Miss some small but significant cross correlations One-step MSE Two-step MSE VAR 0.615(1.349) 1.168(2.129) RVAR 0.606(1.293) 1.159(2.285) 0.588(1.341) 1.154(2.312) Segmentation – p.28 Example 3. Weekly notified measles cases in 7 cities in England (i.e. London, Bristol, Liverpool, Manchester, Newcastle, Birmingham and Sheffield) in 1948-1965, before the advent of vaccination. All the 7 series show biennial cycles, which is the major driving force for the cross correlations among different cities. n = 937, p = 7. – p.29 0 2000 Weekly recorded cases of measles in 7 cities in England in 1948-1965 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 0 400 800 0 1500 0 300 600 0 400 0 400 800 0 400 1950 – p.30 CCF of weekly recorded cases of measles in 7 cities 20 −5 −5 0 −20 −15 −10 −20 −15 −10 0.6 −0.2 0.6 15 20 10 15 0 −5 −5 10 15 5 10 −20 −15 −10 5 20 −5 −5 10 15 20 0 5 10 15 20 0.6 Y4 & Y7 5 10 15 20 0 5 10 15 20 0.6 Y5 & Y7 5 10 15 20 0 5 10 15 20 Y6 & Y7 0.6 Y6 0 5 10 15 20 0 5 Y7 & Y6 0 20 −0.2 0 0 15 0.6 5 Y5 & Y6 15 10 Y3 & Y7 −0.2 0 Y7 & Y5 −20 −15 −10 0 Y4 & Y6 20 20 −0.2 0 Y6 & Y5 0 20 0.6 0 0 15 0.6 5 15 −0.2 −5 10 Y3 & Y6 20 10 0.6 5 0.6 5 5 −0.2 0 Y5 Y7 & Y4 0 10 0 Y2 & Y7 −0.2 0 Y6 & Y4 20 −0.2 0 20 15 0.6 5 Y4 & Y5 15 10 10 15 20 15 20 Y7 0.6 10 5 −0.2 0.6 −0.2 0 0.6 5 0 Y2 & Y6 −0.2 15 0.6 −20 −15 −10 −0.2 0.6 −0.2 0.6 −0.2 0.6 −0.2 0.6 10 0.6 −20 −15 −10 20 Y3 & Y5 −0.2 5 −20 −15 −10 Y7 & Y3 0 20 0.6 0 0 0.6 −5 15 −0.2 −5 15 −0.2 0 10 Y5 & Y4 −0.2 −20 −15 −10 5 −0.2 0 0.6 −5 0.6 0 −5 10 Y2 & Y5 Y4 −0.2 −20 −15 −10 0 Y6 & Y3 −0.2 0.6 −5 −20 −15 −10 Y7 & Y2 −0.2 −20 −15 −10 −20 −15 −10 0 0.6 0 20 0.6 −5 0 Y5 & Y3 −0.2 0.6 −5 15 −0.2 −20 −15 −10 Y7 & Y1 10 5 −0.2 0 Y6 & Y2 −0.2 −20 −15 −10 5 0 Y3 & Y4 0.6 −5 0.6 0 Y6 & Y1 20 Y4 & Y3 −0.2 0.6 −5 0 Y5 & Y2 −0.2 −20 −15 −10 15 −0.2 −20 −15 −10 20 0.6 0 0.6 0 Y5 & Y1 10 0.6 −5 −0.2 0.6 −5 5 Y3 Y4 & Y2 −0.2 −20 −15 −10 0 −0.2 −20 −15 −10 15 −0.2 20 10 Y2 & Y4 0.6 15 0.6 0 5 −0.2 10 0 0.6 5 −0.2 0.6 −5 Y4 & Y1 20 0.6 0 Y3 & Y2 −0.2 −20 −15 −10 15 −0.2 0 Y3 & Y1 10 −0.2 0.6 −5 5 Y2 & Y3 −0.2 0.6 −0.2 −20 −15 −10 0 0.6 20 Y1 & Y7 −0.2 15 Y2 0.6 10 −0.2 5 0.6 0 Y1 & Y6 −0.2 20 −0.2 15 0.6 10 Y2 & Y1 Y1 & Y5 −0.2 5 −0.2 −0.2 −0.2 0 Y1 & Y4 0.6 Y1 & Y3 0.6 Y1 & Y2 0.6 Y1 −20 −15 −10 −5 0 0 5 10 – p.31 b = B b t bt = By Transformation: x −4.898e4 3.357e3 −3.315e04 −6.455e3 2.337e3 1.151e3 7.328e4 2.85e4 −9.569e6 −2.189e3 1.842e3 1.457e3 −5.780e5 5.420e3 −5.247e3 5.878e4 −2.674e3 −1.238e3 −1.766e3 3.654e3 3.066e3 2.492e3 2.780e3 8.571e4 −1.466e3 −7.337e4 −5.896e3 3.663e3 6.633e3 3.472e3 −2.981e4 −8.716e4 6.393e8 −2.327e3 5.365e3 −9.475e4 −7.620e4 −3.338e3 1.471e3 2.099e3 −1.318e2 4.259e3 −1.047e3 1.067e3 6.280e3 2.356e3 −4.668e3 8.629e3 6.581e4 Code: aek = a × 10−k – p.32 0 2 4 6 Transformed weekly recorded cases of measles in 7 cities 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 1950 1955 1960 1965 −6 0 4 −4 0 4 −6 −2 2 −4 0 4 −4 0 4 −4 0 1950 – p.33 CCF of transformed weekly recorded cases of measles in 7 cities 20 −5 −5 0 −20 −15 −10 −20 −15 −10 0.5 −0.5 0.5 15 20 10 15 0 −5 −5 10 15 5 10 −20 −15 −10 5 20 −5 −5 10 15 20 0 5 10 15 20 0.5 X4 & X7 5 10 15 20 0 5 10 15 20 0.5 X5 & X7 5 10 15 20 0 5 10 15 20 X6 & X7 0.5 X6 0 5 10 15 20 0 5 X7 & X6 0 20 −0.5 0 0 15 0.5 5 X5 & X6 15 10 X3 & X7 −0.5 0 X7 & X5 −20 −15 −10 0 X4 & X6 20 20 −0.5 0 X6 & X5 0 20 0.5 0 0 15 0.5 5 15 −0.5 −5 10 X3 & X6 20 10 0.5 5 0.5 5 5 −0.5 0 X5 X7 & X4 0 10 0 X2 & X7 −0.5 0 X6 & X4 20 −0.5 0 20 15 0.5 5 X4 & X5 15 10 10 15 20 15 20 X7 0.5 10 5 −0.5 0.5 −0.5 0 0.5 5 0 X2 & X6 −0.5 15 0.5 −20 −15 −10 −0.5 0.5 −0.5 0.5 −0.5 0.5 −0.5 0.5 10 0.5 −20 −15 −10 20 X3 & X5 −0.5 5 −20 −15 −10 X7 & X3 0 20 0.5 0 0 0.5 −5 15 −0.5 −5 15 −0.5 0 10 X5 & X4 −0.5 −20 −15 −10 5 −0.5 0 0.5 −5 0.5 0 −5 10 X2 & X5 X4 −0.5 −20 −15 −10 0 X6 & X3 −0.5 0.5 −5 −20 −15 −10 X7 & X2 −0.5 −20 −15 −10 −20 −15 −10 0 0.5 0 20 0.5 −5 0 X5 & X3 −0.5 0.5 −5 15 −0.5 −20 −15 −10 X7 & X1 10 5 −0.5 0 X6 & X2 −0.5 −20 −15 −10 5 0 X3 & X4 0.5 −5 0.5 0 X6 & X1 20 X4 & X3 −0.5 0.5 −5 0 X5 & X2 −0.5 −20 −15 −10 15 −0.5 −20 −15 −10 20 0.5 0 0.5 0 X5 & X1 10 0.5 −5 −0.5 0.5 −5 5 X3 X4 & X2 −0.5 −20 −15 −10 0 −0.5 −20 −15 −10 15 −0.5 20 10 X2 & X4 0.5 15 0.5 0 5 −0.5 10 0 0.5 5 −0.5 0.5 −5 X4 & X1 20 0.5 0 X3 & X2 −0.5 −20 −15 −10 15 −0.5 0 X3 & X1 10 −0.5 0.5 −5 5 X2 & X3 −0.5 0.5 −0.5 −20 −15 −10 0 0.5 20 X1 & X7 −0.5 15 X2 0.5 10 −0.5 5 0.5 0 X1 & X6 −0.5 20 −0.5 15 0.5 10 X2 & X1 X1 & X5 −0.5 5 −0.5 −0.5 −0.5 0 X1 & X4 0.5 X1 & X3 0.5 X1 & X2 0.5 X1 −20 −15 −10 −5 0 0 5 10 – p.34 CCF of prewhitened transformed weekly recorded cases of measles 20 −5 −5 0 −20 −15 −10 −20 −15 −10 1.0 −0.2 0.4 1.0 15 20 10 15 0 −5 −5 10 15 5 10 −20 −15 −10 5 20 −5 −5 10 15 20 0 5 10 15 20 1.0 X4 & X7 5 10 15 20 0 5 10 15 20 1.0 X5 & X7 5 10 15 20 0 5 10 15 20 X6 & X7 1.0 X6 0 5 10 15 20 0 5 X7 & X6 0 20 −0.2 0.4 0 0 15 1.0 5 X5 & X6 15 10 X3 & X7 −0.2 0.4 0 X7 & X5 −20 −15 −10 0 X4 & X6 20 20 −0.2 0.4 0 X6 & X5 0 20 1.0 0 0 15 1.0 5 15 −0.2 0.4 −5 10 X3 & X6 20 10 1.0 5 1.0 5 5 −0.2 0.4 0 X5 X7 & X4 0 10 0 X2 & X7 −0.2 0.4 0 X6 & X4 20 −0.2 0.4 0 20 15 1.0 5 X4 & X5 15 10 10 15 20 15 20 X7 1.0 10 5 −0.2 0.4 1.0 −0.2 0.4 0 1.0 5 0 X2 & X6 −0.2 0.4 15 1.0 −20 −15 −10 −0.2 0.4 1.0 −0.2 0.4 1.0 −0.2 0.4 1.0 −0.2 0.4 1.0 10 1.0 −20 −15 −10 20 X3 & X5 −0.2 0.4 5 −20 −15 −10 X7 & X3 0 20 1.0 0 0 1.0 −5 15 −0.2 0.4 −5 15 −0.2 0.4 0 10 X5 & X4 −0.2 0.4 −20 −15 −10 5 −0.2 0.4 0 1.0 −5 1.0 0 −5 10 X2 & X5 X4 −0.2 0.4 −20 −15 −10 0 X6 & X3 −0.2 0.4 1.0 −5 −20 −15 −10 X7 & X2 −0.2 0.4 −20 −15 −10 −20 −15 −10 0 1.0 0 20 1.0 −5 0 X5 & X3 −0.2 0.4 1.0 −5 15 −0.2 0.4 −20 −15 −10 X7 & X1 10 5 −0.2 0.4 0 X6 & X2 −0.2 0.4 −20 −15 −10 5 0 X3 & X4 1.0 −5 1.0 0 X6 & X1 20 X4 & X3 −0.2 0.4 1.0 −5 0 X5 & X2 −0.2 0.4 −20 −15 −10 15 −0.2 0.4 −20 −15 −10 20 1.0 0 1.0 0 X5 & X1 10 1.0 −5 −0.2 0.4 1.0 −5 5 X3 X4 & X2 −0.2 0.4 −20 −15 −10 0 −0.2 0.4 −20 −15 −10 15 −0.2 0.4 20 10 X2 & X4 1.0 15 1.0 0 5 −0.2 0.4 10 0 1.0 5 −0.2 0.4 1.0 −5 X4 & X1 20 1.0 0 X3 & X2 −0.2 0.4 −20 −15 −10 15 −0.2 0.4 0 X3 & X1 10 −0.2 0.4 1.0 −5 5 X2 & X3 −0.2 0.4 1.0 −0.2 0.4 −20 −15 −10 0 1.0 20 X1 & X7 −0.2 0.4 15 X2 1.0 10 −0.2 0.4 5 1.0 0 X1 & X6 −0.2 0.4 20 −0.2 0.4 15 1.0 10 X2 & X1 X1 & X5 −0.2 0.4 5 −0.2 0.4 −0.2 0.4 −0.2 0.4 0 X1 & X4 1.0 X1 & X3 1.0 X1 & X2 1.0 X1 −20 −15 −10 −5 0 0 5 10 – p.35 Note. When none of the component series are WN, the √ confidence bounds ±1.96/ n could be misleading. Segmentation assumption is invalid for this example! (b) Ratios of maximum cross correlations 1.00 0.08 1.05 0.12 1.10 0.16 1.15 1.20 0.20 (a) Max cross correlations between pairs 5 10 15 20 5 10 15 20 (a) The maximum cross correlations, plotted in descending order, among each of the ( 72 ) = 21 pairs component series of the transformed and prewhitened measles series. The maximization was taken over the lags between -20 to 20. (b) The ratios of two successive correlations in (a). – p.36 The segmentations determined by different numbers of connected pairs for the transformed measles series from 7 cities in England. No. of connected pairs No. of groups 1 6 {4, 5}, {1}, {2}, {3}, {6}, {7} 2 5 {1, 2}, {4, 5}, {3}, {6}, {7} 3 4 {1, 2, 3} , {4, 5}, {6}, {7} 4 3 {1, 2, 3, 7}, {4, 5}, {6} 5 2 {1, 2, 3, 6, 7}, {4, 5} 6 1 {1, · · · , 7} Segmentation – p.37 Post-sample forecasting Forecast the notified measles cases in the last 14 weeks of the period for all 7 cities Using the segmentation with four groups: {1, 2, 3}, {4, 5}, {6} and {7} One-step MSE Two-step MSE VAR 503.408(1124.213) 719.499(2249.986) RVAR 574.582(1432.217) 846.141(2462.019) Segmentation 472.106(1088.17) 654.843(1807.502) – p.38 Example 4. Daily impressions of 32 keywords for an air ticket booking website powered by Baidu (www.baidu.com) over 130 days. ⇒ n = 130, p = 32 Data have been coded to protect the confidentiality. Advertisements are accessed by keywords search from a search engine. Each display of an advertisement is counted as one impression. Applying the proposed transformation with m = 10 (or 1 ≤ m ≤ 15), the transformed 32 time series are segmented into 31 groups with the only non-single-element group {10, 13}. – p.39 −0.5 2.0 Time series of 8 randomly selected daily impressions 20 40 60 80 100 120 1 20 40 60 80 100 120 1 20 40 60 80 100 120 1 20 40 60 80 100 120 1 20 40 60 80 100 120 1 20 40 60 80 100 120 1 20 40 60 80 100 120 1 20 40 60 80 100 120 −0.5 1.0 0.5 2.5 −1 1 −0.5 1.5 −1.0 0.0 0.5 2.0 0.5 2.5 1 – p.40 CCF of 8 randomly selected daily impressions 0 −5 0 0 −5 −15 −10 −5 −15 −10 −5 −15 −10 −5 −15 −10 −5 5 10 −15 −10 −5 15 −10 −5 −15 −10 −5 0.8 −0.2 0.8 0.8 −0.2 0.8 −0.2 15 5 10 5 10 −15 −10 −5 −15 −10 −5 0.8 10 15 0 15 5 10 15 0.8 Y21 & Y31 5 10 15 0 5 10 15 Y23 & Y31 0 5 10 15 0 Y24 5 10 15 Y24 & Y31 0 5 10 15 0 5 Y31 & Y24 0 10 0.8 5 Y23 & Y24 0 5 Y18 & Y31 −0.2 0 Y31 & Y23 0 0 Y21 & Y24 15 15 −0.2 0 15 10 −0.2 0.8 −0.2 10 Y24 & Y23 Y31 & Y21 0 15 0.8 0 0 10 0.8 5 Y23 0 5 Y18 & Y24 −0.2 0 Y24 & Y21 −15 0 Y21 & Y23 Y23 & Y21 0 15 5 Y17 & Y31 −0.2 0 0 0.8 0 0 10 Y18 & Y23 15 15 −0.2 0 5 15 0.8 −5 0 10 −0.2 −10 −0.2 0.8 0.8 −0.2 0.8 −0.2 10 10 0.8 −5 Y31 & Y18 0 15 5 Y17 & Y24 0.8 5 0 −0.2 −10 15 −0.2 0 Y24 & Y18 Y31 & Y17 0 −0.2 0.8 0.8 −0.2 0.8 −0.2 −15 0 10 0.8 −15 0 0.8 −10 −0.2 0.8 −0.2 0.8 −0.2 0.8 −0.2 0.8 −0.2 0.8 −0.2 −5 5 Y21 0.8 −10 −0.2 −15 15 10 Y17 & Y23 Y18 & Y21 Y23 & Y18 0.8 −10 0.8 −5 0 −0.2 −15 −0.2 0.8 −10 −15 Y31 & Y9 −0.2 −15 −5 Y24 & Y17 0.8 −5 Y31 & Y1 10 5 5 Y9 & Y31 10 15 Y31 0.8 0 5 0 0 −0.2 −5 −0.2 0.8 −10 −10 0.8 −10 0 Y23 & Y17 Y24 & Y9 −0.2 −15 −15 0 Y21 & Y18 −0.2 −15 15 15 0.8 0 Y24 & Y1 10 10 −0.2 0 15 5 0.8 −5 0.8 −5 0 0.8 −10 −0.2 0.8 −10 −5 −0.2 −15 Y23 & Y9 −0.2 −15 −10 10 0 −0.2 0 Y23 & Y1 5 Y18 Y21 & Y17 0.8 −5 −15 5 Y17 & Y21 0.8 0 0 Y1 & Y31 Y9 & Y24 −0.2 −5 0 15 0.8 −10 −0.2 0.8 −10 15 0.8 −15 Y21 & Y9 −0.2 −15 10 10 −0.2 0 Y21 & Y1 15 −0.2 −5 5 −0.2 0.8 −10 0 Y18 & Y17 −0.2 0.8 −0.2 −15 10 5 0.8 0 5 0 Y9 & Y23 −0.2 −5 Y18 & Y9 15 0.8 −10 0 Y17 & Y18 0.8 −15 Y18 & Y1 10 −0.2 0 15 5 Y9 & Y21 0.8 −5 10 −0.2 0.8 −10 5 Y17 −0.2 0.8 −0.2 −15 0 Y17 & Y9 0 −0.2 15 15 0.8 10 10 −0.2 5 5 Y9 & Y18 0.8 0 Y17 & Y1 0 0.8 0 15 −0.2 −5 10 −0.2 0.8 −10 5 Y9 & Y17 −0.2 0.8 −0.2 −15 0 0.8 15 Y1 & Y24 −0.2 10 Y9 −0.2 5 0.8 0 Y1 & Y23 −0.2 15 0.8 10 Y9 & Y1 Y1 & Y21 −0.2 5 −0.2 −0.2 −0.2 0 Y1 & Y18 0.8 Y1 & Y17 0.8 Y1 & Y9 0.8 Y1 −15 −10 −5 0 0 5 10 15 – p.41 CCF of 8 prewhitened transformed daily impressions (6 randomly selected) 0 −5 0 0 −5 −15 −10 −5 −15 −10 −5 −5 −15 −10 −5 −15 −10 −5 15 −15 10 15 −10 −5 −15 −10 −5 −15 −10 −5 0.8 0.8 −0.2 0.8 −0.2 0 10 5 10 −15 −10 −5 −15 −10 −5 0.8 0 10 15 0 5 10 15 X15 & X28 5 10 15 0 5 10 15 X23 & X28 5 10 15 0 X24 5 10 15 X24 & X28 0 5 10 15 0 5 X28 & X24 0 15 0.8 0 0 10 0.8 5 X23 & X24 15 5 X11 & X28 −0.2 0 X28 & X23 0 15 X15 & X24 15 15 −0.2 0 X24 & X23 0 10 0.8 5 10 −0.2 0.8 −0.2 15 X23 0 5 X11 & X24 −0.2 0 X28 & X15 0 10 5 0.8 5 0 X15 & X23 X24 & X15 0 15 0.8 5 0 X1 & X28 −0.2 0 X23 & X15 0 10 X11 & X23 0.8 0 −0.2 0.8 −0.2 0.8 0.8 −0.2 0.8 −0.2 10 15 −0.2 0 5 15 0.8 −10 0 0.8 5 10 0.8 −5 X28 & X11 0 15 5 X1 & X24 −0.2 0 X24 & X11 0 10 0 −0.2 −10 15 10 −0.2 −15 X28 & X1 0 −0.2 0.8 0.8 −0.2 0.8 −0.2 −15 0 0.8 −10 −0.2 0.8 −0.2 0.8 −0.2 0.8 −0.2 0.8 −0.2 0.8 −0.2 −5 5 X15 0.8 −10 −0.2 −15 15 10 X1 & X23 X11 & X15 X23 & X11 0.8 −10 0.8 −5 0 −0.2 −15 −0.2 0.8 −10 −15 X28 & X13 −0.2 −15 −5 X24 & X1 0.8 −5 X28 & X10 10 5 5 X13 & X28 10 15 X28 0.8 0 5 0 0 −0.2 −5 −0.2 0.8 −10 −10 0.8 −10 0 X23 & X1 X24 & X13 −0.2 −15 −15 0 X15 & X11 −0.2 −15 15 15 0.8 0 X24 & X10 10 10 −0.2 0 15 5 0.8 −5 0.8 −5 0 0.8 −10 −0.2 0.8 −10 −5 −0.2 −15 X23 & X13 −0.2 −15 −10 10 0 −0.2 0 X23 & X10 5 X11 X15 & X1 0.8 −5 −15 5 X1 & X15 −0.2 0 0 X10 & X28 X13 & X24 0.8 −5 0 15 −0.2 −10 −0.2 0.8 −10 15 0.8 −15 X15 & X13 −0.2 −15 10 10 0.8 0 X15 & X10 15 −0.2 −5 5 −0.2 0.8 −10 0 X11 & X1 −0.2 0.8 −0.2 −15 10 5 −0.2 0 5 0 X13 & X23 0.8 −5 X11 & X13 15 −0.2 −10 0 X1 & X11 0.8 −15 X11 & X10 10 0.8 0 15 5 X13 & X15 0.8 −5 10 −0.2 0.8 −10 5 X1 −0.2 0.8 −0.2 −15 0 X1 & X13 0 0.8 15 15 −0.2 10 10 0.8 5 5 X13 & X11 0.8 0 X1 & X10 0 −0.2 0 15 −0.2 −5 10 −0.2 0.8 −10 5 X13 & X1 −0.2 0.8 −0.2 −15 0 X10 & X24 −0.2 15 0.8 10 X13 −0.2 5 0.8 0 X10 & X23 −0.2 15 0.8 10 X13 & X10 X10 & X15 −0.2 5 −0.2 −0.2 −0.2 0 X10 & X11 0.8 X10 & X1 0.8 X10 & X13 0.8 X10 −15 −10 −5 0 0 5 10 15 – p.42 Example 5. Daily sales of a clothing brand in 25 provinces in China in 1 January 2008 – 16 December 2012. n = 1812, p = 25 Annual pattern: peak in February Strong periodicity component with the period 7. The 25 transformed series are segmented into 24 groups with {15, 16} as one group. Permutation is performed using the max-CCF with 14 ≤ m ≤ 30. – p.43 7 9 11 Time series of daily sales of a clothing brand in 8 provinces 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 7.5 9.5 7.0 9.0 7 9 7.0 9.0 8.0 9.5 7.5 9.5 8.0 9.5 2008 – p.44 CCF of daily sales of a clothing brand in 8 provinces −20 −10 −10 20 20 −10 −10 10 10 0.8 0.8 0.0 0.8 0.0 10 20 0 10 20 0 −10 5 10 20 0 5 20 10 20 Y7 & Y8 0.8 Y7 0 5 10 20 0 5 Y8 & Y7 0 10 Y6 & Y8 0.8 5 Y8 & Y6 −20 20 0.0 0 0 10 0.8 5 0.8 −10 20 Y5 & Y8 0.0 −20 5 Y6 & Y7 20 10 Y4 & Y8 0.8 5 5 0.0 0 Y7 & Y6 0 0 0.0 0 0 20 Y5 & Y7 20 20 0.8 5 Y6 0 10 0.8 5 Y8 & Y5 −20 5 0.0 0 10 0.0 0 0.8 −20 0.0 0.8 0.8 0.0 10 5 Y3 & Y8 0.8 5 0.8 −10 0 Y4 & Y7 0.0 −20 20 0.8 0 Y5 & Y6 20 10 0.0 0 Y7 & Y5 0 20 20 0.0 10 5 Y4 & Y6 Y6 & Y5 0 10 10 Y2 & Y8 0.0 5 0.8 5 Y8 & Y4 −20 0.0 0.8 0.8 0 5 Y3 & Y7 0.8 10 0 10 20 Y8 0.8 −10 0 0.0 5 20 0.0 −20 20 0.0 0 0 10 0.8 20 10 Y2 & Y7 0.0 10 5 0.8 −10 5 Y5 0 0 Y3 & Y6 0.8 −10 20 0.0 0.8 0.0 0 0.8 5 Y7 & Y4 0 20 0.0 0 0.8 −20 0 10 0.8 20 10 Y2 & Y6 0.0 0.8 10 5 Y4 & Y5 0.8 −10 0.0 0.8 0.0 0.8 0.8 0.8 0.0 0 0.0 −20 0 0.8 0 20 0.0 −20 5 Y6 & Y4 0.0 −10 5 Y8 & Y3 0.8 −20 10 0.0 −10 0 Y3 & Y5 0.8 0 20 Y1 & Y8 0.0 0 0 0.0 −10 0.8 −10 20 Y5 & Y4 0.0 0.8 0.0 −20 0 Y7 & Y3 0.0 0.8 0 0 0.8 −20 Y8 & Y2 0.0 −10 −10 0.8 0 10 Y4 0.0 −10 5 Y6 & Y3 0.8 0.8 0.8 0 Y8 & Y1 −20 −20 Y7 & Y2 0.0 −10 −20 0 0.0 −20 Y7 & Y1 −20 0 0.0 −10 5 Y5 & Y3 0.8 −20 0 20 10 Y2 & Y5 0.0 0 Y6 & Y2 0.0 −10 10 0.8 −10 0.0 0.8 0 Y6 & Y1 −20 5 Y5 & Y2 0.0 −10 0 5 Y3 & Y4 0.0 −20 Y5 & Y1 −20 0 Y4 & Y3 0.8 0 20 0.8 0 0.0 0.8 −10 10 0.0 −10 Y4 & Y2 0.0 −20 5 Y3 0.8 −20 0 0.0 0 0.0 0.8 0 Y4 & Y1 20 0.8 20 10 Y2 & Y4 0.0 10 5 0.8 5 Y3 & Y2 0.0 −10 0 0.8 0 Y3 & Y1 −20 20 0.0 0 10 0.0 0.8 −10 5 Y2 & Y3 0.0 0.8 0.0 −20 0.0 0 Y1 & Y7 0.0 20 Y2 0.8 10 0.0 5 0.8 0 Y1 & Y6 0.0 20 0.8 10 Y2 & Y1 Y1 & Y5 0.0 5 0.0 0.0 0.0 0 Y1 & Y4 0.8 Y1 & Y3 0.8 Y1 & Y2 0.8 Y1 −20 −10 0 0 5 10 20 – p.45 CCF of 8 transformed daily sales of a clothing brand −10 10 20 −10 −20 −10 10 20 0.6 −0.2 0.6 10 20 20 −20 −10 −20 −10 0 5 5 5 10 20 0 5 20 10 20 10 20 0.6 X20 & X23 0 5 10 20 0 5 10 20 0.6 X21 & X23 0 5 10 20 0 5 X23 & X21 0 10 0.6 0 X21 0 20 X19 & X23 −0.2 10 10 0.6 5 0.6 5 5 X3 & X23 X20 & X21 −0.2 0 0 X19 & X21 X23 & X20 0 20 −0.2 5 0 0.6 0 X21 & X20 0 10 −0.2 5 0.6 0 20 X2 & X23 −0.2 5 10 10 20 X23 0.6 −20 0 X20 0 20 0.6 −10 10 0.6 0 5 X3 & X21 0.6 −20 5 −0.2 20 −0.2 0.6 0.6 −0.2 20 −0.2 10 0 X19 & X20 X23 & X19 0 10 0 X16 & X23 0.6 5 −0.2 0 −0.2 5 20 X2 & X21 0.6 0 10 −20 −10 0 −0.2 −20 20 −0.2 20 X21 & X19 0 −0.2 0.6 −0.2 0.6 10 5 0.6 −10 10 0.6 5 −0.2 0 0 X3 & X20 0.6 −20 5 0.6 −0.2 20 −0.2 0 X23 & X3 0 10 X15 & X23 X16 & X21 0.6 −10 20 −0.2 −10 5 0.6 −20 0 X20 & X19 0.6 −20 0 −0.2 0 10 X2 & X20 −0.2 −10 X21 & X3 0 20 0.6 −20 5 0.6 −10 10 X19 −0.2 0 −0.2 0.6 0.6 −0.2 −0.2 20 0.6 −20 5 0.6 10 0 X3 & X19 −0.2 −10 0 0.6 5 X23 & X2 0 20 −0.2 0 X21 & X2 0 10 X15 & X21 X16 & X20 −0.2 −10 5 X20 & X3 0.6 −20 0 0.6 −20 20 X2 & X19 −0.2 0 10 0.6 −10 20 −0.2 −10 0.6 −20 10 0.6 −20 5 X19 & X3 −0.2 0 −0.2 0.6 −0.2 0.6 0 X15 & X20 X16 & X19 0.6 −10 0 X3 0.6 −10 5 X20 & X2 −0.2 −20 20 0.6 20 −0.2 0 0.6 0 −0.2 −10 −20 X23 & X16 0.6 −20 10 0.6 −10 0 X19 & X2 −0.2 −20 10 X2 & X3 −0.2 0 X23 & X15 20 −0.2 0 0.6 −10 −0.2 −20 5 X21 & X16 0.6 X21 & X15 10 0.6 −10 −0.2 −20 5 X3 & X2 0.6 0 −0.2 −10 0 X20 & X16 0.6 X20 & X15 −20 5 −0.2 0 X15 & X19 X16 & X3 0.6 −10 −0.2 −20 0 X2 0.6 0 −0.2 −10 0 X19 & X16 0.6 −20 20 −0.2 20 0.6 0 −0.2 −10 X19 & X15 −0.2 10 X3 & X16 0.6 −0.2 −20 10 0.6 5 0.6 0 −0.2 −10 5 X2 & X16 0.6 −0.2 −20 0 X16 & X2 −0.2 0 X3 & X15 −0.2 20 0.6 0 −0.2 −10 X2 & X15 −0.2 10 −0.2 5 X16 0.6 −0.2 −20 X15 & X3 0.6 0 −0.2 20 −0.2 10 −0.2 5 −0.2 −0.2 0 X16 & X15 −0.2 X15 & X2 0.6 X15 & X16 0.6 X15 0 5 10 20 – p.46 CCF of 8 prewhitened transformed daily sales of a clothing brand −20 −10 −10 −20 −10 −20 −20 −10 −10 0 −20 −10 −20 −10 0 −20 −10 5 10 0 5 10 0 −20 −10 −20 −10 20 −20 −10 0 5 10 20 0 5 10 20 0 0 5 10 5 20 0 5 −20 −10 20 −20 −10 −0.2 0.6 20 0 10 0 5 10 20 0 5 10 20 0 5 0 5 10 5 5 20 0 5 −20 −10 20 10 20 10 20 10 20 X20 & X23 20 0 5 10 20 X21 & X23 20 0 5 X23 & X21 0 10 X19 & X23 X21 0 20 X3 & X23 X20 & X21 X23 & X20 0 10 10 X2 & X23 X19 & X21 X21 & X20 0 0 X3 & X21 X20 0 20 −0.2 0.6 10 −0.2 0.6 5 10 −0.2 0.6 −0.2 0.6 5 5 X2 & X21 X19 & X20 X23 & X19 0 0 X3 & X20 X21 & X19 0 20 −0.2 0.6 0 −0.2 0.6 −0.2 0.6 −0.2 0.6 0 X20 & X19 X23 & X3 0 20 X19 X21 & X3 0 −0.2 0.6 −0.2 0.6 −0.2 0.6 20 X20 & X3 X23 & X2 0 10 10 X3 & X19 0 X16 & X23 −0.2 0.6 0 X21 & X2 0 5 20 −0.2 0.6 −20 0 X19 & X3 X20 & X2 0 20 10 10 −0.2 0.6 5 5 −0.2 0.6 0 0 5 X2 & X20 −0.2 0.6 0 20 X2 & X19 X3 0 X16 & X21 −0.2 0.6 −10 10 10 20 −0.2 0.6 −20 −0.2 0.6 −0.2 0.6 −0.2 0.6 −0.2 0.6 5 X19 & X2 0 5 10 X16 & X20 X15 & X23 10 20 X23 0 −0.2 0.6 0 −10 0 5 −0.2 0.6 −10 0 0 X15 & X21 −0.2 0.6 −20 20 20 −0.2 0.6 −20 20 X2 & X3 −0.2 0.6 0 10 10 −0.2 0.6 −10 5 5 −0.2 0.6 −20 0 −0.2 0.6 −10 10 0 X16 & X19 −0.2 0.6 −20 20 X3 & X2 X23 & X16 −0.2 0.6 −0.2 0.6 −10 5 20 X15 & X20 −0.2 0.6 0 X23 & X15 −20 0 X21 & X16 −0.2 0.6 −0.2 0.6 −10 −10 10 10 −0.2 0.6 0 X21 & X15 −20 0 X20 & X16 −0.2 0.6 −0.2 0.6 −10 −20 5 −0.2 0.6 0 X20 & X15 −20 5 X19 & X16 −0.2 0.6 −0.2 0.6 −10 −10 0 X15 & X19 X16 & X3 −0.2 0.6 −20 20 X2 −0.2 0.6 0 X19 & X15 −20 0 X3 & X16 −0.2 0.6 −0.2 0.6 −10 20 −0.2 0.6 0 X3 & X15 −20 10 10 −0.2 0.6 5 X2 & X16 −0.2 0.6 −0.2 0.6 −10 5 −0.2 0.6 0 −0.2 0.6 0 X2 & X15 −20 0 X16 & X2 −0.2 0.6 −10 20 X16 −0.2 0.6 −0.2 0.6 −20 10 −0.2 0.6 5 X15 & X3 −0.2 0.6 0 −0.2 0.6 20 −0.2 0.6 10 X16 & X15 −0.2 0.6 5 X15 & X2 −0.2 0.6 0 X15 & X16 −0.2 0.6 −0.2 0.6 X15 0 5 10 20 – p.47 Post-sample forecasting Forecast the daily sales for the last two weeks in each of the 25 provinces. Segmentation: 24 groups with {15, 16} as one group One-step MSE Two-step MSE univariate AR 0.208(1.255) 0.194(1.250) VAR 0.295(2.830) 0.301(2.930) 0.293(2.817) 0.296(2.962) RVAR 0.153(0.361) 0.163(0.341) Segmentation Cross correlations between provinces are useful information. However they cannot be used via direct VAR! – p.48 Asymptotic theory For p × r matrices H1 and H2 and H′1 H1 = H′2 H2 = Ir , let r 1 D(M(H1 ), M(H2 )) = 1 − tr(H1 H′1 H2 H′2 ). r Then D(M(H1 ), M(H2 )) ∈ [0, 1]. It equals 0 iff M(H1 ) = M(H2 ), and 1 iff M(H1 ) ⊥ M(H2 ). yt is assumed to be weakly stationary and α-mixing, i.e. αk ≡ sup i sup ∞ i , B∈Fi+k A∈F−∞ |P (A ∩ B) − P (A)P (B)| → 0, as k → ∞, where Fij = σ(yt : i ≤ t ≤ j). – p.49 p fixed C1. For some constant γ > 2, sup max E(|yi,t − Eyi,t |2γ ) < ∞. t C2. P∞ 1−2/γ k=1 αk 1≤i≤p < ∞. Theorem 1. Let conditions C1, C2 hold, ̟ be positive and p be b = (A b 1, · · · , A b q ) of which the fixed. Then there exists an A b y , such that columns are a permutation of the columns of Γ b j ), M(Aj )) = Op (n−1/2 ). max D(M(A 1≤j≤q – p.50 p = o(nc ) for some c > 0 C3. (Sparsity of A = (ai,j )) For some constant ι ∈ [0, 1), p p X X max |ai,j |ι ≤ s1 and max |ai,j |ι ≤ s2 , 1≤j≤p i=1 1≤i≤p j=1 where s1 and s2 are positive constants which may diverge together with p. C4. For some positive constants l > 2 and τ > 0, sup max P (|yi,t − µi | > x) = O(x−2(l+τ ) ) as x → ∞. t 1≤i≤p C5. αk = O(k −l(l+τ )/(2τ ) ) as k → ∞. Remark. As Σy (k) = AΣx (k)A′ , the sparsity of A and the maximum block size of Σx (k) provide a measure for the sparsity of Σy (k). – p.51 Let Smax be the maximum block size in Σx (k), and ρj = min min |λ(Wx,i ) − λ(Wx,j )|, j = 1, · · · , q, 1≤i≤q i6=j δ = s1 s2 max k=1,...,k0 kΣx (k)kι∞ , κ1 = min k=1,...,k0 kΣx (k)k2 , κ2 = max k=1,...,k0 kΣx (k)k2 . b y (k) = (b Thresholding CCVF: Let Σ σi,j (k)) be CCVF. Define e y (k) = σ Σ bi,j (k)I{|b σi,j (k)| ≥ u} , u = M p2/l n−1/2 . Theorem 2. Let conditions C3-C5 hold, minj ρj > 0 and p = o(nl/4 ). b = (A b 1, · · · , A b q ) of which the columns are Then there exists an A b y , such that a permutation of the columns of Γ b j ), M(Aj )) max ρj D(M(A 1≤j≤q n Op κ2 (p4/l n−1 )(1−ι)/2 Smax δ , κ−1 (p4/l n−1 )(1−ι)/2 Smax δ = O(1) 1 = 4/l −1 1−ι 2 −1 δ −1 = O(1). Op (p n ) Smax δ 2 , κ2 (p4/l n−1 )−(1−ι)/2 Smax – p.52 Remarks (i) Theorem 2 gives the uniform convergence rate for b j ), M(Aj )). The smaller ρj is, more difficult the ρj D(M(A estimation for M(Aj ) is. (ii) The smaller ι, s1 , s2 and Smax are, more sparse Σy (k) is, and the faster the convergences are. (iii) Similar results can be obtained for the cases with log p = o(nc ) by assuming the sub-Gaussianity for yt and exponential decay rates for α-mixing coefficients. – p.53 Simulation Let A = (A1 , · · · , Aq ), Aj is p × pj , P j pj = p, and P b b b b bj = p. A = (A1 , · · · , Aqb), Aj is p × pbj , jp Correct segmentation: qb = q, pbj = pj , and b j )) = min D(M(Aj ), M(A b i )), D(M(Aj ), M(A 1≤i≤q j = 1, · · · , q. for H′1 H1 = Ir1 and H′2 H2 = Ir2 , d(M(H1 ), M(H2 )) = 1 − 1/2 1 ′ ′ tr(H1 H1 H2 H2 ) . min (r1 , r2 ) b j ) is an estimator for Incomplete segmentation: qb < q, and each M(A the linear space spanned by one, or more than one Ai volatility – p.54 No. of replications: 500 times for each setting. Let A = (aij ), aij ∼ U (−3, 3) independently. Example 6. Let p = 6, yt = Axt , and (1) (2) (3) xi,t = ηt+i−1 for i = 1, 2, 3, xj,t = ηt+j−4 for j = 4, 5, x6,t = ηt . where (1) = 0.5ηt−1 + 0.3ηt−2 + et − 0.9et−1 + 0.3et−2 + 1.2et−3 + 1.3et−4 , (2) = −0.4ηt−1 + 0.5ηt−2 + et + et−1 − 0.8et−2 + 1.5et−3 , (3) = 0.9ηt−1 + et , ηt ηt ηt (1) (1) (2) (1) (3) (2) (2) (1) (2) (1) (2) (1) (2) (1) (1) (2) (3) (3) and et , et , et are indep N (0, 1). Thus q = 3, i.e. 3 segmented subseries with 3, 2, 1 elements. – p.55 Relative frequencies of correct and incomplete segmentations n 100 200 300 400 500 1000 1500 2000 2500 3000 Correct .350 .564 .660 .712 .800 .896 .906 .920 .932 .945 Incomplete .508 .386 .326 .282 .192 .104 .094 .080 .068 .055 P b i )) (with correct segmentations only) D(M(A ), M( A i 1≤i≤3 0.0 0.2 0.4 0.6 0.8 1.0 Boxplots of 1 3 100 200 300 400 Sample size 500 1000 1500 – p.56 CCF of yt (one instance) −10 0 0 −10 −10 0 10 15 20 0 1.0 −1.0 1.0 −1.0 0.0 15 20 −20 −15 5 −10 0 5 0 −10 10 15 20 0 5 0 20 10 15 20 10 15 20 0.0 1.0 Y5 & Y6 0 5 10 15 20 0 5 Y6 & Y5 −5 15 1.0 5 Y5 −5 10 0.0 0 1.0 −15 0 Y4 & Y6 0.0 −20 20 1.0 10 1.0 5 15 Y3 & Y6 0.0 0 10 0.0 5 Y6 & Y4 −5 20 −1.0 0 Y5 & Y4 −5 15 Y4 & Y5 1.0 −15 0.0 1.0 1.0 0.0 −1.0 20 0.0 −20 10 1.0 15 1.0 −15 5 0.0 10 0.0 −20 0 −1.0 −5 Y6 & Y3 −5 −1.0 0.0 1.0 1.0 0.0 5 5 Y3 & Y5 1.0 0 0 Y2 & Y6 −1.0 0 20 0.0 −10 20 10 15 20 15 20 Y6 1.0 −5 15 1.0 −15 1.0 −15 10 Y4 0.0 −20 5 0.0 −20 15 1.0 −10 1.0 0 −1.0 −5 20 1.0 −15 0 Y5 & Y3 0.0 1.0 −10 15 10 Y2 & Y5 0.0 0 5 −20 −15 −10 −5 0 −1.0 −5 Y6 & Y2 0.0 −15 10 0 Y3 & Y4 0.0 −20 Y6 & Y1 −20 5 Y1 & Y6 0.0 −10 −1.0 0 20 1.0 −15 1.0 −5 −1.0 −10 15 0.0 −20 0.0 1.0 −15 0 Y5 & Y2 0.0 −20 10 Y4 & Y3 −1.0 0 Y5 & Y1 20 −1.0 0 15 −1.0 −5 1.0 −5 −1.0 −10 5 Y3 0.0 1.0 −15 0 1.0 −10 10 −1.0 20 0.0 −15 5 −1.0 15 −1.0 −20 0 Y2 & Y4 −1.0 10 Y4 & Y2 0.0 −20 20 1.0 5 1.0 0 15 0.0 0 0.0 −5 −1.0 −10 10 Y3 & Y2 1.0 −15 5 −1.0 0 −1.0 −5 0.0 −1.0 −20 0 Y2 & Y3 1.0 −10 −1.0 −15 Y4 & Y1 −1.0 20 0.0 1.0 0.0 −1.0 −20 Y3 & Y1 −1.0 15 Y2 −1.0 10 −1.0 5 Y1 & Y5 0.0 1.0 0.0 0 −1.0 20 Y1 & Y4 −1.0 15 −1.0 10 −1.0 0.0 5 −1.0 0.0 −1.0 0 Y2 & Y1 −1.0 Y1 & Y3 1.0 Y1 & Y2 1.0 Y1 0 5 10 – p.57 −10 0 0 −10 −10 0 Segmentation: {1, 4, 6}, {2, 5}, {3} 0 10 15 20 1.0 1.0 20 −20 −15 5 −10 0 5 0 −10 0 20 10 15 20 1.0 5 10 15 20 0 5 10 15 20 0.0 1.0 Z5 & Z6 0 5 10 15 20 0 5 Z6 & Z5 −5 15 0.0 0 Z5 −5 10 Z4 & Z6 1.0 −15 0 1.0 15 0.0 −20 20 0.0 10 1.0 5 15 Z3 & Z6 0.0 0 10 0.0 −1.0 20 −1.0 5 Z6 & Z4 −5 −1.0 0.0 1.0 −1.0 1.0 −1.0 0.0 0 Z5 & Z4 −5 15 Z4 & Z5 1.0 −15 10 1.0 20 0.0 −20 5 −1.0 −5 Z6 & Z3 −5 15 1.0 −15 0 0.0 10 0.0 −20 0.0 1.0 1.0 0.0 5 5 Z3 & Z5 1.0 0 0 Z2 & Z6 −1.0 0 20 0.0 −10 20 10 15 20 15 20 Z6 1.0 −5 15 1.0 −15 1.0 −15 10 Z4 0.0 −20 5 0.0 −20 15 1.0 −10 1.0 0 −1.0 −5 20 1.0 −15 0 Z5 & Z3 0.0 1.0 −10 15 10 Z2 & Z5 0.0 0 5 −20 −15 −10 −5 0 −1.0 −5 Z6 & Z2 0.0 −15 10 0 Z3 & Z4 0.0 −20 Z6 & Z1 −20 5 Z1 & Z6 0.0 −10 −1.0 0 20 1.0 −15 1.0 −5 −1.0 −10 15 0.0 −20 0.0 1.0 −15 0 Z5 & Z2 0.0 −20 10 Z4 & Z3 −1.0 0 Z5 & Z1 20 −1.0 0 15 −1.0 −5 1.0 −5 −1.0 −10 5 Z3 0.0 1.0 −15 0 1.0 −10 10 −1.0 20 0.0 −15 5 −1.0 15 −1.0 −20 0 Z1 & Z5 Z2 & Z4 −1.0 10 Z4 & Z2 0.0 −20 20 1.0 5 1.0 0 15 0.0 0 0.0 −5 −1.0 −10 10 Z3 & Z2 1.0 −15 5 −1.0 0 −1.0 −5 0.0 −1.0 −20 0 Z2 & Z3 1.0 −10 −1.0 −15 Z4 & Z1 −1.0 20 0.0 1.0 0.0 −1.0 −20 Z3 & Z1 −1.0 15 Z2 −1.0 10 −1.0 5 0.0 1.0 0.0 0 −1.0 20 Z1 & Z4 −1.0 15 −1.0 10 −1.0 0.0 5 −1.0 0.0 −1.0 0 Z2 & Z1 −1.0 Z1 & Z3 1.0 Z1 & Z2 1.0 Z1 bt (one instance) CCF of x 0 5 10 – p.58 Example 7. Let p = 20, q = 5, (p1 , · · · , p5 ) = (6, 5, 4, 3, 2) xt is defined similarly as in Example 6. Relative frequencies of correct and incomplete segmentations n 400 500 1000 1500 2000 2500 3000 Correct 0.058 0.118 0.492 0.726 0.862 0.902 0.940 Incomplete 0.516 0.672 0.460 0.258 0.130 0.096 0.060 1 5 P 1≤i≤5 b i )) (with correct segmentations only) D(M(Ai ), M(A 0.0 0.2 0.4 0.6 0.8 1.0 Boxplots of 400 500 1000 1500 Sample size 2000 2500 3000 – p.59 Segmenting multiple volatility processes Let Ft = σ(yt , yt−1 , · · · ), E(yt |Ft−1 ) = 0, Assumption: Var(yt |Ft−1 ) = Σy (t). yt = Axt , Var(xt |Ft−1 ) = diag Σ1 (t), · · · , Σq (t) . Let Var(yt ) = Var(xt ) = Ip , then A is orthogonal. Let Bt−1 be a π -class and σ(Bt−1 ) = Ft−1 . put X X [E{xt x′t I(B)}]2 . [E{yt yt′ I(B)}]2 , Wx = Wy = B∈Bt−1 B∈Bt−1 For any B ∈ Bt−1 , E{xt x′t I(B)} = E[I(B)E{xt x′t |Ft−1 }] = E[I(B)diag(Σ1 (t), · · · , Σq (t))] is a block diagonal matrix, so is Wx . – p.60 Since Wy = AWx A′ , the proposed method continues to apply. In practice we estimate Wy by cy = W k0 XX B∈B k=1 n X 2 1 ′ yt yt I(yt−k ∈ B) , n−k t=k+1 where B may consist of {u ∈ Rp : kuk ≤ kyt k} for t = 1, · · · , n. – p.61 Example 8. Consider the daily returns in 2 Jan 2002 – 10 July 2008 of six stocks: Bank of America, Dell, JPMorgan, FedEx, McDonald and American International Group. n = 1642, p = 6 −0.227 −0.203 0.022 b B= −0.583 0.804 0.144 −0.093 0.031 0.550 0.348 −0.041 −0.562 0.201 0.073 −0.059 0.158 0.054 −0.068 0.436 −0.549 0.005 0.096 −0.129 −0.068 −0.012 0.668 −0.099 −0.409 −0.033 0.008 0.233 −0.012 −0.582 0.131 0.098 −0.028 Segmentation for transformed series: CUC of Fan, Wang & Y (2008) – p.62 −10 0 5 Daily returns of 6 stocks 2003 2004 2005 2006 2007 2008 2002 2003 2004 2005 2006 2007 2008 2002 2003 2004 2005 2006 2007 2008 2002 2003 2004 2005 2006 2007 2008 2002 2003 2004 2005 2006 2007 2008 2002 2003 2004 2005 2006 2007 2008 −10 0 5 −15 −5 5 −15 −5 5 −20 −5 5 15 −15 −5 5 2002 – p.63 CCF of squared returns 0.8 0.0 0 20 15 0 −20 −15 −10 0.8 10 15 20 0 5 0 10 15 20 0.4 0.8 Y4^2 & Y6^2 5 10 15 20 0 5 10 15 20 0.0 0.4 0.8 Y5^2 & Y6^2 0 5 10 15 20 0 5 10 15 20 Y6^2 0.8 Y6^2 & Y5^2 −5 20 0.0 0 0 15 0.4 5 Y5^2 −5 10 Y3^2 & Y6^2 Y4^2 & Y5^2 0.8 −20 −15 −10 5 0.0 0 20 20 0.8 20 0.8 10 15 0.0 15 0.4 5 10 0.4 0.8 0.4 10 0.8 15 5 Y2^2 & Y6^2 0.0 10 0.4 0.8 0 20 0.4 0.8 5 Y6^2 & Y4^2 −5 15 Y3^2 & Y5^2 0.8 0 0 0.0 −20 −15 −10 5 Y5^2 & Y4^2 −5 10 0.0 0 0.0 −20 −15 −10 0.4 0.8 0.0 0.4 0.8 0.8 0.4 20 0.4 0 0.4 0.8 0 15 0.0 −5 Y6^2 & Y3^2 −5 10 Y4^2 0.8 0 0.0 −20 −15 −10 5 0.0 0 0.0 −5 5 Y2^2 & Y5^2 0.4 0.8 20 0.4 0.8 −20 −15 −10 0.4 0.8 0.4 0.0 0 15 −20 −15 −10 Y6^2 & Y2^2 −5 10 Y5^2 & Y3^2 0.0 0 5 0.8 0 0 Y3^2 & Y4^2 0.4 −5 0.4 0.8 0.4 −5 0 0.0 −20 −15 −10 Y6^2 & Y1^2 −20 −15 −10 0 Y5^2 & Y2^2 0.0 −20 −15 −10 20 Y4^2 & Y3^2 0.8 0 15 0.0 0 0.0 −5 Y5^2 & Y1^2 10 0.4 0.8 −5 0.4 0.8 0.4 0.0 −20 −15 −10 5 Y3^2 Y4^2 & Y2^2 20 0.0 0 0.0 −20 −15 −10 15 0.4 20 10 Y2^2 & Y4^2 0.0 15 5 0.8 10 0.4 0.8 0.4 0.0 0 0 0.8 5 Y3^2 & Y2^2 −5 20 0.0 0 Y4^2 & Y1^2 15 0.4 0 10 0.4 0.8 0.0 −5 Y3^2 & Y1^2 −20 −15 −10 5 Y2^2 & Y3^2 0.4 0.8 0.4 0.0 −20 −15 −10 0.4 0 0.0 20 0.0 15 Y2^2 0.8 10 0.4 5 0.0 0.4 0 Y1^2 & Y6^2 0.0 20 0.8 15 0.4 10 Y2^2 & Y1^2 Y1^2 & Y5^2 0.0 5 0.0 0.4 0.0 0.4 0.0 0 Y1^2 & Y4^2 0.8 Y1^2 & Y3^2 0.8 Y1^2 & Y2^2 0.8 Y1^2 −20 −15 −10 −5 0 0 5 10 15 20 – p.64 CCF of residuals from fitted GARCH(1,1) for each return series 20 0.8 0.0 0.8 0.4 5 20 −20 −15 −10 −20 −15 −10 15 20 0 5 0 10 15 20 0.4 0.8 Y4 & Y6 5 10 15 20 0 5 10 15 20 0.0 0.4 0.8 Y5 & Y6 0 5 10 15 20 0 5 Y6 & Y5 −5 20 0.0 0 0 15 0.8 10 Y5 −5 10 0.4 5 0.8 15 20 Y3 & Y6 0.4 10 0.8 0 0 Y4 & Y5 0.0 −5 20 0.0 5 15 0.0 0 Y6 & Y4 0.8 −20 −15 −10 15 0.8 15 0.8 0 0 10 0.4 10 10 Y2 & Y6 0.0 5 5 Y3 & Y5 0.4 0.8 −5 0 0.0 5 Y5 & Y4 0.0 −20 −15 −10 20 0.8 0 0.4 0 15 0.4 20 0.0 −5 10 0.0 15 Y4 0.0 0 0.4 0.8 0.0 0.4 0.8 0.8 0.4 0 0.4 0.8 −5 10 0.8 20 Y6 & Y3 0.0 −20 −15 −10 5 0.0 15 −20 −15 −10 0 5 Y2 & Y5 0.4 0.8 0.0 10 0.4 0.8 −5 0.4 0.8 0.4 0 5 Y5 & Y3 0.0 −20 −15 −10 0 Y3 & Y4 0.8 0 Y6 & Y2 −5 0 0.0 −5 0.4 0.8 0.4 0.0 0 0.0 −20 −15 −10 0 Y5 & Y2 −5 20 0.4 0.8 −20 −15 −10 Y6 & Y1 15 Y4 & Y3 0.0 0 10 0.4 0.8 0 0.4 0.8 0.4 −5 Y5 & Y1 −20 −15 −10 5 Y3 −5 20 0.0 0 0.0 −20 −15 −10 15 10 15 20 15 20 Y6 0.8 20 10 Y2 & Y4 0.8 15 Y4 & Y2 0.0 −20 −15 −10 5 0.4 10 0.4 0.8 0.4 0.0 0 0 0.8 5 Y3 & Y2 −5 20 0.0 0 Y4 & Y1 15 0.0 0 10 0.4 0.8 0.0 −5 Y3 & Y1 −20 −15 −10 5 Y2 & Y3 0.4 0.8 0.4 0.0 −20 −15 −10 0.4 0 0.4 20 0.0 15 Y2 0.8 10 0.4 5 0.0 0.4 0 Y1 & Y6 0.0 20 0.8 15 0.4 10 Y2 & Y1 Y1 & Y5 0.0 5 0.0 0.4 0.0 0.4 0.0 0 Y1 & Y4 0.8 Y1 & Y3 0.8 Y1 & Y2 0.8 Y1 −20 −15 −10 −5 0 0 5 10 – p.65 CCF of squared transformed returns 0.8 0.0 0 20 15 0 −20 −15 −10 0.8 10 15 20 0 5 0 10 15 20 0.4 0.8 X4^2 & X6^2 5 10 15 20 0 5 10 15 20 0.0 0.4 0.8 X5^2 & X6^2 0 5 10 15 20 0 5 10 15 20 X6^2 0.8 X6^2 & X5^2 −5 20 0.0 0 0 15 0.4 5 X5^2 −5 10 X3^2 & X6^2 X4^2 & X5^2 0.8 −20 −15 −10 5 0.0 0 20 20 0.8 20 0.8 10 15 0.0 15 0.4 5 10 0.4 0.8 0.4 10 0.8 15 5 X2^2 & X6^2 0.0 10 0.4 0.8 0 20 0.4 0.8 5 X6^2 & X4^2 −5 15 X3^2 & X5^2 0.8 0 0 0.0 −20 −15 −10 5 X5^2 & X4^2 −5 10 0.0 0 0.0 −20 −15 −10 0.4 0.8 0.0 0.4 0.8 0.8 0.4 20 0.4 0 0.4 0.8 0 15 0.0 −5 X6^2 & X3^2 −5 10 X4^2 0.8 0 0.0 −20 −15 −10 5 0.0 0 0.0 −5 5 X2^2 & X5^2 0.4 0.8 20 0.4 0.8 −20 −15 −10 0.4 0.8 0.4 0.0 0 15 −20 −15 −10 X6^2 & X2^2 −5 10 X5^2 & X3^2 0.0 0 5 0.8 0 0 X3^2 & X4^2 0.4 −5 0.4 0.8 0.4 −5 0 0.0 −20 −15 −10 X6^2 & X1^2 −20 −15 −10 0 X5^2 & X2^2 0.0 −20 −15 −10 20 X4^2 & X3^2 0.8 0 15 0.0 0 0.0 −5 X5^2 & X1^2 10 0.4 0.8 −5 0.4 0.8 0.4 0.0 −20 −15 −10 5 X3^2 X4^2 & X2^2 20 0.0 0 0.0 −20 −15 −10 15 0.4 20 10 X2^2 & X4^2 0.0 15 5 0.8 10 0.4 0.8 0.4 0.0 0 0 0.8 5 X3^2 & X2^2 −5 20 0.0 0 X4^2 & X1^2 15 0.4 0 10 0.4 0.8 0.0 −5 X3^2 & X1^2 −20 −15 −10 5 X2^2 & X3^2 0.4 0.8 0.4 0.0 −20 −15 −10 0.4 0 0.0 20 0.0 15 X2^2 0.8 10 0.4 5 0.0 0.4 0 X1^2 & X6^2 0.0 20 0.8 15 0.4 10 X2^2 & X1^2 X1^2 & X5^2 0.0 5 0.0 0.4 0.0 0.4 0.0 0 X1^2 & X4^2 0.8 X1^2 & X3^2 0.8 X1^2 & X2^2 0.8 X1^2 −20 −15 −10 −5 0 0 5 10 15 20 – p.66 CCF of residuals from transformed return series 20 0.8 0.0 0.8 0.4 5 20 −20 −15 −10 −20 −15 −10 15 20 0 5 0 10 15 20 0.4 0.8 X4 & X6 5 10 15 20 0 5 10 15 20 0.0 0.4 0.8 X5 & X6 0 5 10 15 20 0 5 X6 & X5 −5 20 0.0 0 0 15 0.8 10 X5 −5 10 0.4 5 0.8 15 20 X3 & X6 0.4 10 0.8 0 0 X4 & X5 0.0 −5 20 0.0 5 15 0.0 0 X6 & X4 0.8 −20 −15 −10 15 0.8 15 0.8 0 0 10 0.4 10 10 X2 & X6 0.0 5 5 X3 & X5 0.4 0.8 −5 0 0.0 5 X5 & X4 0.0 −20 −15 −10 20 0.8 0 0.4 0 15 0.4 20 0.0 −5 10 0.0 15 X4 0.0 0 0.4 0.8 0.0 0.4 0.8 0.8 0.4 0 0.4 0.8 −5 10 0.8 20 X6 & X3 0.0 −20 −15 −10 5 0.0 15 −20 −15 −10 0 5 X2 & X5 0.4 0.8 0.0 10 0.4 0.8 −5 0.4 0.8 0.4 0 5 X5 & X3 0.0 −20 −15 −10 0 X3 & X4 0.8 0 X6 & X2 −5 0 0.0 −5 0.4 0.8 0.4 0.0 0 0.0 −20 −15 −10 0 X5 & X2 −5 20 0.4 0.8 −20 −15 −10 X6 & X1 15 X4 & X3 0.0 0 10 0.4 0.8 0 0.4 0.8 0.4 −5 X5 & X1 −20 −15 −10 5 X3 −5 20 0.0 0 0.0 −20 −15 −10 15 10 15 20 15 20 X6 0.8 20 10 X2 & X4 0.8 15 X4 & X2 0.0 −20 −15 −10 5 0.4 10 0.4 0.8 0.4 0.0 0 0 0.8 5 X3 & X2 −5 20 0.0 0 X4 & X1 15 0.0 0 10 0.4 0.8 0.0 −5 X3 & X1 −20 −15 −10 5 X2 & X3 0.4 0.8 0.4 0.0 −20 −15 −10 0.4 0 0.4 20 0.0 15 X2 0.8 10 0.4 5 0.0 0.4 0 X1 & X6 0.0 20 0.8 15 0.4 10 X2 & X1 X1 & X5 0.0 5 0.0 0.4 0.0 0.4 0.0 0 X1 & X4 0.8 X1 & X3 0.8 X1 & X2 0.8 X1 −20 −15 −10 −5 0 0 5 10 – p.67 Conclusions • A new version of PCA for time series: finding a latent segmentation via a contemporaneous linear transformation • When the segmentation does not exist, provide effective approximations by ignoring negligible, though significant, correlations P • Why works? Wy = k Σy (k)Σy (k)′ = AWx A′ , tr(Wy ) = p XX k i,j=1 ρyij (k)2 = tr(Wx ) = p XX k i,j=1 ρxij (k)2 = XX k ρxii (k)2 i provided all components of xt are uncorrelated across all time lags Components of xt are predictable, due to stronger ACF! – p.68 • When latend segmentation does not exist, Wy = X use zt = Γ′y yt , Σy (k)Σy (k)′ = Γy DΓ′y . k Hence Σz (k) = Γ′y Σy (k)Γy , and therefore Wz = X Σz (k)Σz (k)′ = Γ′y Wy Γy = D, k i.e. Wz is a diagonal matrix. Note. Σz (k) are unlikely to be diagonal, though the off-diagonal elements tend to be small. – p.69