Time Series Analysis 3. ADL Model Autoregressive Distributed Lag Model Autoregressive: p lags of dependent variable Yt distributed lag: q lags of additional regressor Xt ⇒ ADL(p, q) Yt = β0 + β1Yt−1 + ... + βpYt−p + δ1Xt−1 + ... + δq Xt−q + ut • β(L)Yt = β0 + δ(L)Xt−1 with lag-polynomials defined by β(L) = 1 − β1L − ... − βpLp δ(L) = δ1 + δ2L + ... + δq Lq−1 • k additional predictors: ADL(p, q1, ..., qk ) β(L)Yt = β0 + δ1(L)X1,t−1 + δ2(L)X2,t−1 + ... + δk (L)Xk,t−1 1 Time Series Analysis 3. ADL Model Autoregressive Distributed Lag Model • Model Assumptions (1) E[ut|Yt−1, Yt−2, ..., X1t−1, X1t−1, ..., Xkt−1, Xkt−2, ...] = 0 (2)(a) (Yt, X1t, ..., Xkt) are (strictly) stationary (b) (Yt, X1t, ..., Xkt) are ergodic, i.e. (3) (Yt, X1t, ..., Xkt) and (Yt−j , X1t−j , ..., Xkt−j ) become independent for j→∞ Yt and X1t, ..., Xkt have nonzero, finite fourth moments (4) no perfect multicollinearity ⇒ OLS regression theory applies 2 Time Series Analysis 3. ADL Model Autoregressive Distributed Lag Model (1) implies o Cov(ut, Yt−j ) = 0, Cov(ut, Xit−j ) = 0 ∀j > 0 and i = 1, ..., k o Cov(ut, ut−j ) = 0 ∀j > 0 : ut’s are not serially correlated ⇒ no further lags of Yt, Xit’s needed (3) assures that variance estimators are consistent 3 Time Series Analysis 3. ADL Model Granger Causality (“predictability”) • Asks whether forecast of Yt can be improved by considering lags of Xt • Ωt: information set in period t MSFE (Yt+h|Ωt): MSFE of forecasting Yt+h given Ωt Xt Granger causes Yt: Gr Xt → Yt if MSFE (Yt+h|Ωt) < MSFE (Yt+h|Ωt/{Xs|s ≤ t})) for at least one forecast horizon h > 0 4 Time Series Analysis 3. ADL Model Granger Causality (“predictability”) Gr • if Xt 9 Yt, then Xt has no predictive content for Yt Gr Gr • Test: H0 : Xt 9 Yt vs. H1 : Xt → Yt = b F -Test on significance of lags of Xt in ADL(p, q) model Gr Example: H0 : Unemploymentt 9 ∆Inft ADL(4, 4) model: F -Test = 10.45 ⇒ p-value < 0.0001 ⇒ Reject H0 5 Time Series Analysis 3. ADL Model Forecast Uncertainty • Forecast uncertainty: unknown uT +1 + estimation error • RMSFE usual measure → standard error of forecast • Example ADL(1, 1): Yt = β0 + β1Yt−1 + δXt−1 + ut ŶT +1|T = β̂0 + β̂1YT + δ̂1XT ûT +1|T = YT +1 − ŶT +1|T = uT +1 + [(β̂0 − β0) + (β̂1 − β1)YT + (δ̂1 − δ1)XT ] MSFE = E[(YT +1 − ŶT +1|T )2] = σu2 + Var [(β̂0 − β0) + (β̂1 − β1)YT + (δ̂1 − δ1)XT ] | {z } γ̂ because uT +1 is uncorrelated with second term 6 Time Series Analysis 3. ADL Model Forecast Uncertainty • How to estimate Var(γ̂)? Problem: dependence of β̂0, β̂1 and δ̂1 on YT , XT o Pseudo out-of-sample forecasts o Estimate conditional variance Var(γ̂|YT , XT ) (as for models with nonstochastic regressors) o Estimate asymptotic variance lim T · Var(γ̂) T →∞ 7 Time Series Analysis 3. ADL Model Standard Error of Forecast: ADL(1,1) Var(γ̂|YT , XT ) = σu2 zT (Z 0Z)−1zT0 with zT = (1 YT XT ) and 1 Y1 X1 .. .. Z = .. 1 YT XT MSFE = σu2 (1 + zT (Z 0Z)−1zT0 ) | {z } Inflation factor q RMSFE = σu (1 + zT (Z 0Z)−1zT0 ) is standard error of forecast q \ = σ̂u (1 + zT (Z 0Z)−1z 0 ) is estimated standard error of forecast RMSFE T 8 Time Series Analysis 3. ADL Model Forecast Interval • ”Confidence interval for forecast” • If ut is normally distributed, then (asymptotical) 95% forecast interval for ŶT +1|T is given by \ ŶT +1|T ± 1.96 · RMSFE • If infinitely many FI’s are computed, then 95% of the FI estimates contain the true value YT +1 9