Time series forecasting by evolutionary neural networks

Abstract

This chapter presents a hybrid Evolutionary Computation/Neural Network combination for time series prediction. Neural networks are innate candidates for the forecasting domain due to advantages such as nonlinear learning and noise tolerance. However, the search for the ideal network structure is a complex and crucial task. Under this context, Evolutionary Computation, guided by the Bayesian Information Criterion, makes a promising global search approach for feature and model selection. A set of ten time series, from different domains, were used to evaluate this strategy, comparing it with a heuristic model selection, as well as with conventional forecasting methods (e.g., Holt-Winters and Box-Jenkins methodology).