Neural Network Models for Time Series Forecasts
Tim Hill, Marcus O’Connor, William Remus
Neural networks have been advocated as an alternative to traditional statistical forecasting methods. In the present experiment, time series forecasts produced by neural networks are compared with forecasts from six statistical time series methods generated in a major forecasting competition (Makridakis et al. [Makridakis, S., A. Anderson, R. Carbone, R. Fildes, M. Hibon, R. Lewandowski, J. Newton, E. Parzen, R. Winkler. 1982. The accuracy of extrapolation (time series) methods: Results of a forecasting competition. J. Forecasting 1 111–153.]); the traditional method forecasts were estimated by experts in the particular technique. The neural networks were estimated using the same ground rules as the competition. Across monthly and quarterly time series, the neural networks did significantly better than traditional methods. As suggested by theory, the neural networks were particularly…