ST414 – Spectral Analysis of Time Series Data Lecture 6 18 February 2014 Last Time • • • • Introduction to bivariate processes The cross-covariance function The cross-spectrum Coherence analysis 2 Today’s Objectives • • • • Asymptotics for the spectral density matrix Estimating VAR parameters Causality The VAR spectral density matrix 3 Nonparametric Estimation Given a bivariate time series (π1 t , π2 t )′, construct the bivariate DFTs π ππ = (ππ1 π , ππ2 π )′, where π πππ ππ = π −1/2 ππ π‘ exp(−π2πππ π‘) π‘=1 4 Nonparametric Estimation The periodogram matrix is π° ππ = π ππ π ππ ∗ 5 Nonparametric Estimation The smoothed periodogram matrix is π ππ 1 = 2π + 1 π π=−π π π°(ππ + ) π 6 Nonparametric Estimation More generally, π π ππ = π=−π where βπ ≥ 0, |π|≤π βπ π βπ π°(ππ + ) π = 1, and βπ = β−π 7 Nonparametric Estimation From here, an estimate of the squared coherence is 2 πππ π = |πππ π |2 ππ π ππ π 8 Nonparametric Estimation Let Z(t) be bivariate Gaussian white noise (0, Σ), Σ non-singular, and let π°(π) be the periodogram matrix of Z(t). Then as π → ∞, π°(π) converges in distribution to a random matrix ππ ∗ , where π~π πΆ (0, Σ). 9 Nonparametric Estimation πΈ π π πΆππ£ πππ π , πππ π 2 π β π=−π π → π ππ →π π 0 if π ≠ π π ππ π π if π = π ≠ 0, 0.5 10 Nonparametric Estimation |πππ π | 2 π β π=−π π ~π΄π( πππ π , 1 − πππ π π‘ππβ−1 ( πππ π ) 2 π β π=−π π 2 /2) ~π΄π(π‘ππβ−1 ( πππ π ), 1/2) 11 Nonparametric Estimation Alternatively, if we use a uniform kernel for 2 smoothing, under π»0 : πππ π = 0, we have 2 πππ π πΉ= 2π 2 1 − πππ π πΉ~πΉ(2, 2 2π + 1 − 2) 12 SOI Example 13 SOI Example 14 Sales Example 15 Sales Example 16 Sales Example 17 Sales Example 18 Sales Example 19 Sales Example 20 Sales Example 21 Sales Example 22 Sales Example −2.682 β 2π −4.173 β 2π 23 Least Squares Estimation Consider the K-variate VAR(p) model: π π‘ = Φ1 π π‘ − 1 + β― Φπ π π‘ − π + π π‘ Assume “presamples” π 0 , π −1 , … , π −π + 1 are available. 24 Least Squares Estimation Rewrite as π = π΅π + π. The least squares estimator is π΅ = ππ′ ππ ′ −1 Moreover, π΄ = π −1 (π − π΅π)(π − π΅π)′ 25 Causality If Y(t) can be predicted more efficiently if the past values of the process X(t) is taken into account, then X(t) Granger-causes Y(t). If Y(t+1) can be predicted more efficiently if X(t+1) is taken into account, then X(t) instantaneously-causes Y(t). 26 Causality π(π‘) = π(π‘) π π=1 π11,π π12,π π21,π π22,π π(π‘ − π) π1 (π‘) + π(π‘ − π) π2 (π‘) X(t) does not Granger-cause Y(t) if π21,π = 0 for all k. 27 Causality π11 π12 π΄= π π 21 22 Y(t) does not instantaneously-cause X(t) if π21 = π12 = 0. 28 The VAR Spectrum Recall the spectral density for an AR(p) is π2 π π = , 2 |Φ(exp −π2ππ )| where π ππ π§ π Φ π§ =1− π=1 29 The VAR Spectrum The spectral density matrix for a VAR(p) is π π = Φ(exp −π2ππ )−1 Σ[Φ(exp −π2ππ )∗ ]−1 , where π Φπ π§ π Φ π§ = πΌπ − π=1 30 Sales Example VAR(2) for Sales/Lead data: 0.280 − 0.730 Φ1 = 0.028 − 0.516 0.205 − 2.177 Φ2 = −0.011 − 0.153 1.431 − 0.022 Σ= −0.022 0.077 31 Sales Example Causality tests: π»0 : Sales does not Granger-cause Lead p = 0.2669 π»0 : Lead does not Granger-cause Sales p = 1.806 x 10-8 π»0 : No instantaneous causality p = 0.4221 32 Sales Example 33 Sales Example 34 Sales Example 35