Time Series Forecasting with PSO-Optimized Neural Networks
Daniel Alba-Cuéllar, Ángel Eduardo Muñoz Zavala, Arturo Hernández-Aguirre, Eunice Esther Ponce de León Sentí, Elva Diaz-Diaz · 2014
In this paper, we propose a new methodology to forecast values for univariate time series datasets, based on a Feed Forward Neural Network (FFNN) ensemble. Each ensemble element is trained with the Particle Swarm Optimization (PSO) algorithm, this ensemble produces a final sequence of time series forecasts via a bootstrapping procedure. Our proposed methodology is compared against Auto-Regressive Integrated Moving Average (ARIMA) models. This experiment gives us a good idea of how effective soft computing techniques can be in the field of time series modeling. The results obtained show empirically that our proposed methodology is robust and produces useful forecast error bounds that provide a clear picture of a time series' future movements.