James D. Thomas and Katia Sycara
A genetic programming system for data mining trading rules out of past foreign exchange data is described. The system is tested on real data from the dollar/yen and dollar/DM markets, and shown to produce considerable excess returns in the dollar/yen market. Design issues relating to potential rule complexity and validation regimes are explored empirically. Keeping potential rules as simple as possible is shown to be the most important component of success. Validation issues are more complicated. Inspection of fitness on a validation set is used to cut-off search in hopes of avoiding overfitting. Additional attempts to use the validation set to improve performance are shown to be ineffective in the standard framework. An examination of correlations between performance on the validation set and on the test set leads to an understanding of how such measures can be marginally benificial; unfortunately, this suggests that further attemps to improve performance through validation will be difficult.