On Assessing Prediction Error in Autoregressive Models (Q4956035)

For a Gaussian AR(1) stationary time series \(Y(t)=\vartheta Y(t-1)+\eta(t)\) the problem of one-step-ahead prediction is considered. A standard predictor for \(Y(t)\), say \(\hat\vartheta_n Y(t-1)\), is used with \( \hat\vartheta_n=\sum_{t=2}^n Y(t)Y(t-1)/\sum_{t=1}^{n-1}Y^2(t) \). The conditional (with respect to the observed data) mean square error \(V\) is used as a measure of the prediction performance. The authors demonstrate analytically and by simulations that the estimator of \(V\) proposed by \textit{P. Zhang} and \textit{P. Shaman} [Trans. Am. Math. Soc. 347, No. 2, 627-637 (1995; Zbl 0814.62062)] has bad performance. They propose a new estimator \[ \bar V=n(n-1)^{-1}\hat s_n^2+n^{-1}(1-\hat\vartheta_n^2)Y^2(n), \] \[ \text{where}\quad \hat s_n^2=(n-1)^{-1}\sum_{t=2}^n(Y(t)-\hat\vartheta_n Y(t+1))^2. \] Results of simulations demonstrate good performance of the \(\bar V\) estimate.