Financial Time-Series Prediction using Negatively Correlated Neural Network Ensembles

ABSTRACT

Previous research has shown that the use of neural network ensembles can increase the generalization of outputs leading to improved network performance. Additional studies have proven that forcing ensembles to give negatively correlated outputs can improve generalization even further. Negative correlation is introduced by adding a penalty to each network's error function. This penalty will increase a network's error if it gives output similar to other networks in the ensemble. Thus networks are given incentive to develop differently and therefore produce different outputs.

While many approaches have been used in attempts to predict financial time series, there is a lack of research in applying negative correlation in ensembles to the problem. Negatively correlated ensembles have already been proven to benefit time-series and classification problems. The research discussed here applies this approach to time series prediction of a financial index. Financial data is well suited as a data set. It contains a large amount of noise and requires a highly generalized output.

A program was implemented to verify the effects of negative correlation on predicting financial time-series. The results of the method were compared against other standard approaches. The experimental results showed that negative correlation performed significantly better than other methods.