Forecasting of 15 time series of DGOR with neural networks

In 1982, the working group ''Forecasting Methods'' of the Deutsche Gesellschaft fur Operations Research (DGOR) carried out a forecasting comparison between 12 various models which were applied to 15 time series. The results of this study can be considered as a good benchmark for further prediction techniques. This paper reports upon the prediction of these 15 time series by using a Neural Network which was developed by the Backpropagation algorithm. The four highest autocorrelated lag-variables were used as the input variables of the Neural Network. The results show that the Neural Network delivered worse predictions than the other methods including the naive pre diction by forecasting non-stationary time series. Stationary time series could be predicted better than the naive pre diction, but in comparison to the other techniques the results were only average. After regarding the problem of non-stationarity by using the Dickey-Fuller-Test, first differences were chosen as the input-variables of the Neural Network. In this case, there was a considerable improvement, but the best method (Box-Jenkins' ARIMA technique) could not be surpassed.