Prequential and cross-validated mixture regression estimation

Abstract

Prequential model selection and cross-validation are data-driven methodologies for selecting a single «best» model from a collection of competing models. In contrast, we propose prequential and cross-validated mixtures which are suitably weighted combinations of all the rival models under consideration. We empirically study prequential and cross-validated mixtures (both based on neural networks) for regression estimation. © 1998 IEEE.