On Estimating Conditional Mean-Squared Prediction Error in Autoregressive Models

From MaRDI portal





The authors consider the \(p\)\,th-order autoregressive (AR) process \(x_{t}=\beta_{1}x_{t-1}+\ldots+\beta_{p}x_{t-p}+\varepsilon_{t}\), where \(\varepsilon_{t}\) are independent and identically \(N(0,\sigma^2)\) distributed, and the polynomial \(1-\beta_{1}z-\ldots-\beta_{p}z^{p}\) has no zeros inside or on the unit disk. The conditional mean-squared prediction error (CMSPE) is given by \[ E\{(x_{n+1}-\widehat x_{n+1})^2\mid \,x_{1},\ldots,x_{n}\}=\sigma^2+\{{\mathbf x}'_{n}(p) (\widehat \beta_{n}-\beta)\}^2. \] The final prediction error (FPE) in estimating CMSPE is given by \(\text{FPE}= (N+p) (N-p)^{-1} \widehat\sigma^2_{n}\), where \[ \widehat\sigma ^2_{n}= N^{-1} \sum_{j=p}^{n-1} \{x_{j+1}- {\mathbf x}'_{j} \widehat \beta _{n}\}^2,\;{\mathbf x}_{j}=(x_{j}, \ldots,x_{j-p+ 1}),\;N=n-p. \] The authors propose an estimator of CMSPE, \(\widetilde V\), in the higher-order AR model, and obtain the asymptotic correlation of FPE and \(\widetilde V\), with the CMSPE. It is shown that the limit of the normalized correlation of \(\widetilde V\) with CMSPE is larger than that of FPE with CMSPE. The main conclusion is that from the MSE point of view, \(\widetilde V\) is the best choice among a family of asymptotically unbiased estimators of CMSPE.











This page was built for publication: On Estimating Conditional Mean-Squared Prediction Error in Autoregressive Models

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4455658)