Investigation of the effect of training and prediction window sizes on neural financial prediction models

Several sources have reported on the use of neural networks in financial trading. One approach is to have the network represent a trading strategy. That is, rather than basing a trading decision on a predicted value or movement in the price of some asset, the network directly outputs a decision buy or sell. Thus, the network does not represent a continuous-valued function of its inputs, but rather, a discrete classifier. While the performance of such a network will depend on factors such as the structure of the network and the selection of input features, other factors also play an important role. In this paper we investigate the effects that training window size and prediction window size have in determining the ability of the network to successfully time the placement of trades. Results of experiments performed on the Dow Jones industrial average suggest that although some configurations of training and prediction windows sizes do yield satisfactory performance over an extended prediction period, improved performance might be achieved by dynamically adjusting window sizes on the basis of recent trends displayed by the time series.