A Neural Network Model for Predicting Time Series with Interventions and a Comparative Analysis
M.D. Cubiles-de-la-Vega, R. Pino-Mejías, J.L. Moreno-Rebollo and J. Muñoz-García
A procedure for designing a multilayer perceptron for predicting time series with interventions is proposed. It is based on the generation, according to a rule emerging from an ARIMA model with interventions previously fitted, of a set of nonlinear forecasting models with interventions. Each model is approximated through a three-layered perceptron, selecting the one minimizing the Bayesian Information Criterion. The training of the multilayer perceptron is performed by three alternative learning rules, incorporating multiple repetitions, and the hidden layer size is computed by means of a grid search. A comparative analysis using time series from the Active Population Survey in Andalusia, Spain, shows a better performance of these neural network models over ARIMA models with interventions.
Backpropagation; backpropagation with momentum; Levenberg-Marquardt; mean squared error; mean absolute deviation; MATLAB; SAS; active population survey.