Michele Hibon

During 2007 we conducted an empirical evaluation of the accuracy of artificial Neural Networks (NN) and other methods of Computational Intelligence (CI) in time series prediction through a dedicated forecasting competition: the NN3 (www.neural-forecasting-competition.com). The competition aimed to resolve two research questions: (a) what is the current performance of CI methods in comparison to established statistical forecasting methods, and (b)

This paper extends the empirical evidence on the forecasting accuracy of extrapolative methods. The robustness of the major conclusions of the M-Competition data is examined in the context of the telecommunications data of Fildes (1992). The performance of Robust Trend, found to be a successful method for forecasting the telecommunications data by Fildes, is compared with that of other successful methods using the M-Competition data. Although it is established that the structure of the telecommunications data is more homogeneous than that of the M-Competition data, the major conclusions of the M-Competition continue to hold for this new data set. In addition, while the Robust Trend method is confirmed to be the best performing method for the telecommunications data, for the 1001 M-Competition series, this method is outperformed by methods such as Single or Damped Smoothing. The performance of smoothing methods depended on how the smoothing parameters are estimated. Optimisation at each time origin was more accurate than optimisation at the first time origin, which in turn is shown to be superior to arbitrary (literature based) fixed values. In contrast to the last point, a data based choice of fixed smoothing constants from a cross-sectional study of the time series was found to perform well.