Feature selection for financial trading rules

Several sources have suggested the use of neural networks for discovering trading strategies for buying and selling financial commodities on a stock exchange. However, a problem with such networks is understanding their behaviour. The trading strategy is embedded in the network weights and hence not readily interpretable. It is not sufficient to only measure the performance of the neural trader in terms of the return that it is able to achieve. We also want to understand why the agent decided to buy or sell. This paper proposes a methodology by which neural traders with an interpretable structure can be discovered. We present results from applying the methodology to the Dow Jones Industrial Average index over a four year period. The results show that the rules represented by such networks have a format similar to commonly proposed trading rules based on a combination of moving average rules, and that application of the procedure over an extended period results in substantially better return than a buy-and-hold strategy.