Improved time series prediction using evolutionary algorithms for the generation of feedback connections in neural networks

Some results from a method for generating recurrent neural networks (RNN) for prediction of financial and macroeconomic time series are presented. In the presented method, a feedforward neural network (FFNN) is first obtained using backpropagation. While backpropagation is usually able to find a fairly good predictor, all FFNN are limited by their lack of short-term dynamic memory. RNNs, by contrast, may exhibit short-term memory due to feedback connections in the network. In the method presented here, the RNNs are generated by an evolutionary algorithm (EA). The preliminary results indicate that the evolved RNN indeed outperform, by a few per cent, the FFNN obtained through backpropagation on several time series. However, it is noted that, regardless of the predictor used, the prediction error cannot be much improved over that obtained from a very simple predictor. Finally, another approach is tested as well, in which the evolved RNN generate not only a prediction but also a measure of confidence in the prediction.