Discovering effective technical trading rules with genetic programming: towards robustly outperforming buy-and-hold

Genetic programming is now a common research tool in financial applications. One classic line of exploration is their use to find effective trading rules for individual stocks or for groups of stocks (such as an index). The classic work in this area (Allen & Karjaleinen, 99) found profitable rules, but which did not outperform a straightforward ¿buy and hold¿ strategy. Several later works report similar outcomes, while a small number of works achieve out-performance of buy and hold, but prove difficult to replicate. We focus here on indicating clearly how the performance in one such study (Becker & Seshadri, 03) was replicated, and we carry out additional investigations which point towards guidelines for generating results that robustly outperform buy-and-hold. These guidelines relate to strategies for organizing the training dataset, and aspects of the fitness function.