Combining Nonparametric and Optimal Linear Time Series Predictions

We introduce a semiparametric procedure for more efficient prediction of a strictly stationaryprocess admitting an ARMA representation. The procedure is based on the estimation of the ARMArepresentation, followed by a nonparametric regression where the ARMA residuals are used as explanatoryvariables. Compared to standard nonparametric regression methods, the number of explanatory variablescan be reduced because our approach exploits the linear dependence of the process. We establish consistencyand asymptotic normality results. A Monte Carlo study and an empirical application on stockindices suggest that significant gains can be achieved with our approach.