Automated stock trading : a multi-agent, evolutionary approach

Abstract

Stock market trading has garnered much interest over the past few decades as it has been made easier for the general public to trade. It is certainly an avenue for wealth growth, but like all risky undertakings, it must be understood for one to be consistently successful. There are, however, too many factors that influence it for one to make completely confident predictions. Automated computer trading has therefore been championed as a potential solution to this problem and is used in major brokerage houses world-wide. In fact, a third of all EU and US stock trades in 2006 were driven by computer algorithms. In this thesis we look at the challenges posed by the automatic generation of stock trading rules and portfolio management. We explore the viability of evolutionary algorithms, including genetic algorithms and genetic programming, for this problem and introduce an agent-based learning framework for individual and social intelligence that is applicable to general stock markets. Statistical tests were applied to determine whether or not there was a significant difference between the evolutionary trading approach and an accepted benchmark. It was found that while the evolutionary trading agents comfortably realised higher portfolio values than the ALSI, there was insufficient evidence to suggest that the agents outperformed the ALSI in terms of portfolio performance. Additionally, it was observed that while the traders combined knowledge from the expert traders to form complex trading models, these models did not result in any statistically significant positive returns. It must be said, however, that there was overwhelming evidence to suggest that the traders learned rules that were highly successful in predicting stock movement.