Evolving Rule-Based Trading Systems

In this study, a market trading rulebase is optimised using genetic programming (GP). The rulebase is comprised of simple relationships between technical indicators, and generates signals to buy, sell short, and remain inactive. The methodology is applied to prediction of the Standard & Poor’s composite index (02-Jan-1990 to 18-Oct-2001). Two potential market systems are inferred: a simple system using few rules and nodes, and a more complex system. Results are compared with a benchmark buy-and-hold strategy. Neither trading system was found capable of consistently outperforming this benchmark. More complicated rulebases, in addition to being difficult to understand, are susceptible to overfitting. Simpler rulebases are more robust to changing market conditions, but cannot take advantage of high-profit-making opportunities. By increasing the richness of the available rulebase building-blocks and the variety of training data, it is anticipated that subsequent systems will surpass the benchmark strategy.