Forecasting with artificial neural network models

This paper contains a forecasting exercise on 30 time series, ranging on several fields, from economy to ecology. The statistical approach to artificial neural networks modelling developed by the author is compared to linear modelling and to other three well-known neural network modelling procedures: Information Criterion Pruning (ICP), Cross-Validation Pruning (CVP) and Bayesian Regularization Pruning (BRP). The findings are that 1) the linear models outperform the artificial neural network models and 2) albeit selecting and estimating much more parsimonious models, the statistical approach stands up well in comparison to other more sophisticated ANN models. JEL classification: C22, C45, C53 Acknowledgments: I am very grateful for the help I received fromMarcelo Medeiros, who developed the programs which produced forecasting results and comparisons for the linear, ICP, CVP and BRP models. Financial support from the Tore Browaldh Foundation and the Stockholm School of Economics is gratefully acknowledged.