Discovering stock market trading rules using multi-layer perceptrons

Abstract

This paper presents an approach to extracting stock market trading rules from stock market data. Trading rules are based on two multi-layer perceptrons, one generating buy signals and one generating sell signals. Inputs of these perceptrons are fed with values of technical indicators computed on historical stock quotations. Results of a large number of experiments on real-life data from the Paris Stock Exchange confirm that the model of trading rules is reasonable and the trading rules are able to generate reasonable trading signals, not only over a training period, used in the training process, but also over a test period, unknown during constructing trading rules. Moreover, trading strategies defined by such trading rules are profitable and often outperform the simple Buy&Hold strategy. © Springer-Verlag Berlin Heidelberg 2007.