Neural networks for financial market prediction

The use of neural networks in financial market prediction presents a major challenge to the design of effective neural network predictors and classifiers. In this paper, the author examines several neural networks to evaluate their capability in prediction and in trend estimation which is treated as a classification problem. The networks considered are the backpropagation trained network (BPN), the general regression neural network (GRNN), the class-sensitive neural network (CSNN), and the conjugate gradient trained network (CGNN). It is concluded that CSNN is among the best performing networks in both prediction and trend estimation. All major indicators are evaluated by the neural networks. It is found that the use of good indicators like the rate of change, momentum, moving average, etc. can lead to about 5% improvement over the case that no indicator is used. Momentum computed from the previous 14 day data is the best single indicator. The complexity of the financial market probably explains why the large number of indicators cannot provide any significant improvement in network classification