A modified adaptive retraining procedure for data forecasting

The paper presents a further improvement of the adaptive retraining procedure of Artificial Neural Networks (ANNs) used for time series predictions. An important advantage of this approach is that the model is periodically adapted to the changes of the non-stationary environment. The retraining starts from proportionally reduced values of the parameters used in the previous version of the ANN model. As usual, variously delayed versions of the time series to be predicted and of the previous outputs are applied at the input of the ANN. In addition, the newly developed model also uses as inputs the averaged seasonal values from the previous years, obtained for the desired target variables in some specified time windows.