Technical trading rules as a prior knowledge to a neural networks prediction system for the S&P 500 index

Financial markets data is very noisy and nonstationary which makes modeling through machine learning from historical information a challenging problem. Our experience indicates that in markets modeling through neural network learning, significant data preprocessing is needed. We have proposed a promising multi-component prediction system for the S&P 500 index which yields a higher return with fewer trades as compared to a neural network predictor alone. The multicomponent system consists of a statistical feature selection, a simple data filtering, two specialized neural networks for extraction of nonlinear relationships from selected data, and a symbolic decision rule base for determining buy/sell recommendations. The objective of this study is to explore if a more sophisticated data filtering process in our multicomponent system leads to further improvements in return or to a reduced number of trades as compared to our current system. The new system uses some well-known technical trading rules/indicators as a prior symbolic knowledge to develop a directional filter that splits the financial data into up, down, and sideway data sets. We use the directional movement indicators to detect whether the market is trending, and to measure the strength of the trend if it exists. Various experimental results using this system to predict S&P 500 index returns are presented and the result compared to our previously developed multi-component system. The system performance is measured by computing the annual rate of return and the return per trade