On the impact of streaming interface heuristics on GP trading agents: an FX benchmarking study

Most research into frameworks for evolving trading agents emphasize aspects associated with the evolution of technical indicators and decision trees / rules. One of the factors that drives the development of such frameworks is the non-stationary, streaming nature of the task. However, it is the heuristics used to interface the evolutionary framework to the streaming data which potentially have most impact on the quality of the resulting trading agents. We demonstrate that including a validation partition has a significant impact on determining the overall success of the trading agents. Moreover, rather than conduct evolution on a continuous basis, only retraining when changes in trading quality are detected also yields significant advantages. Neither of these heuristics are widely recognized by research in evolving trading agent frameworks, although both are relatively easy to add to current frameworks. Benchmarking over a 3 year period of the EURUSD foreign exchange supports these findings.