Fusion of Neural Classifiers for Financial Market Prediction

Forecasting financial currency markets is an extremely challenging problem because of the complex and highly chaotic nature of such markets. Motivated by the substantial profits that could be gained by having a system that could accurately predict large trends in the market, financial institutions are looking on advances in machine learning, neural networks, and statistics to provide them with another analysis tool. Researchers are investigating the use of back-propagation neural networks for financial time series prediction, due to their success on other pattern recognition problems such as machine & handwritten character recognition. However, to date their performance has been considerably lower than that achieved on the character recognition problem domain. This is due in large part to the tremendous amount of noise inherent in the data, which hinders the learning of good mapping functions. We believe that redundant forecasting through the synergistic use of multiple neural network predictors in combination with an intelligent decision aggregation scheme, may be the key to increasing the success rate of computer-aided forecasting systems. In this paper, we conduct an empirical and comparative study on the use of alternative methods for data preprocessing, fitness evaluation, and decision fusion. We demonstrate the advantage of our multiple classifier approach in predicting changes in the foreign exchange rate of the U.S. Dollar versus the German Mark over 250 days of trading.