How Not to Learn Trading Rules

James Thomas

Abstract

There is no need to explain why applying AI to to financial markets is an intriguing problem. But recently there has been a growing interest in AI on the part of economists, as a methodology for testing the idea of efficient markets. The core approach is to use excess returns produced by AI trading rule learners as operational evidence of market inefficiency. Much of this work has used a straightforward genetic programming approach to learn trading rules based on simple technical analysis-inspired components. Unfortunately, financial data is far noisier than typical machine learning problems, and I show that the standard genetic programming approach is poorly suited to learning trading rules in financial markets due to overfitting issues. I then present a new approach, based on an extremely simplified genetic learner combined with ensemble methods. In addition, I present results from recent work exploring the use of news stories as an additional source of data for building trading rules.

Back to the Main Page