Prequential and cross-validated regression estimation (Q1275343)

In this paper nonparametric regression estimators based on neural networks are proposed. Several techniques of the given estimators are presented, namely: (a) rates of convergence for the integrated mean-square errorr in estimating regression functions using ``off-line'' or ``batch'' versions of the prequential and the cross-validated, (b) rates of convergence for the time-averaged expected prediction errors in using ``on-line'' versions of the prequential and cross-validated estimators. The performance of the proposed estimators is studied. It provides the very general remark that in practice prequential model selection is more reliable and useful than cross-validation. To summarize, prequential and cross-validated regression estimators are very attractive for practical applications. They require minimal input from the user and achieve the best tradeoff between the competing approximation error components and estimation error components.