Neural networks applied to intraday futures trading

Since september 1996, a neural network based trading system, fully automatically and independently trades futures on the DTB. Being part of an overall expert system dedicated to the task of predicting short term price movements in order to assist professional market markers when posting quotes, the system was put to this ultimate test of economic significance. Clearly, the prediction of price changes in a competitive market such as the index futures market is a non-trivial and ambitious goal. Certainly within the neural network paradigm, where the number of degrees of freedom is very large compared to the weakness of the dependencies that we attempt to model, extra effort is demanded in the area of feature selection: network construction and the validation of the predictions generated. In this paper we outline the way we tackled the problem of constructing optimal predictors by means of an heuristic ensemble construction scheme using a constructive sequential forward selection scheme to create the individual neural networks.