Time series forecasting using neural networks: Should the data be deseasonalized first?

This research investigates whether prior statistical deseasonalization of d ata is necessary to produce more accurate neural network forecasts. Neural networks trained with deseasonalized data from Hill et al. (1996) were comp ared with neural networks estimated without prior deseasonalization. Both s ets of neural networks produced forecasts for the 68 monthly time series fr om the M-competition (Makridakis et al., 1982). Results indicate that when there was seasonality in the time series, forecasts from neural networks es timated on deseasonalized data were significantly more accurate than the fo recasts produced by neural networks that were estimated using data which we re not deseasonalized. The mixed results from past studies may be due to in consistent handling of seasonality. Our findings give guidance to both prac titioners and researchers developing neural networks. Copyright (C) 1999 Jo hn Wiley & Sons, Ltd.