Integrated Mixture of Local Experts Model for Forecasting
Rodrigo Arnaldo Scarpel, Armando Zeferino Milioni

Abstract
The estimation and usage of real-valued functions for forecasting is a central problem in applied statistics. There are several approaches to deal with such problem as the least-squares method, neural networks and the mixture of local experts model (MLEM). MLEM is built following four stages: (a) partition the input space into regions; (b) for each region train different models; (c) find the best model for each region (local expert); and (d) implement a composition of the local experts that will decide how to weight the local experts output. In this paper we integrate the parameters estimation for the partition of the input space and for training of the local experts, as a way to improve the performance of both the fitting of the models and their usage in forecasting. In order to illustrate the usefulness of the integrated approach, some applications to real datasets are shown.

Full Text: PDF