Neural Networks Versus Time Series Methods: A Forecasting Exercise1

This paper presents an empirical exercise in economic forecast using traditional time series methods, such as ARIMA and unobservable components models (UCM), and artificial neural networks (ANN). We use monthly gross industrial output data for the state of Rio Grande do Sul (Brazil) to perform a comparative exercise and access the relative performance of the different forecasting methods. The results show that ANN forecast more accurately than ARIMA models, but the comparison with UCM is not quite straightforward. The UCM is found to produce better one step ahead forecast than the ANN, but the performance of the ANN for larger forecasts horizons shows that, specially once a proper modeling methodology has been established, it may be a valuable tool to economic forecasting.