Can neural networks applied to time series forecasting learn seasonal patterns: an empirical investigation

Artificial neural networks are increasingly being applied to time series forecasting, but with mixed results. There appears to be as many methods as there are studies. This research investigates whether prior statistical deseasonalising of data is necessary for producing accurate forecasts with neural networks, or whether the networks can adequately model seasonality. Neural networks trained with deseasonalised data from (T. Hill et al., 1992) were compared with neural networks developed without prior deseasonalisation. Both sets of neural networks produced forecasts for the 68 monthly time series from the M-competition (S. Makridakis et al., 1982). Results indicate that neural network forecasts from deseasonalised data were significantly more accurate than the forecasts produced by neural networks which modeled seasonality.<>