Using Genetic Algorithms for Robust Optimization in Financial Applications

In this study, optimal indicators and strategies for foreign exchange trading models are investigated in the framework of genetic algorithms. We first explain how the relevant quantities of our application can be encoded in quot;genesquot; so as to fit the requirements of the genetic evolutionary optimization technique. In financial problems, sharp peaks of high fitness are usually not representative of a general solution but, rather, indicative of some accidental fluctuations. Such fluctuations may arise out of inherent noise in the time series or due to threshold effects in the trading model performance. Peaks in such a discontinuous, noisy and multimodal fitness space generally correspond to trading models which will not perform well in out-of-sample tests. In this paper we show that standard genetic algorithms will be quickly attracted to one of the accidental peaks of the fitness space whereas genetic algorithms for multimodal functions employing clustering and a specially designed fitness sharing scheme will find optimal parameters which correspond to broad regions where the fitness function is higher on average. The optimization and the quality tests have been performed over eight years of high frequency data of the main foreign exchange rates. The authors acknowledge a careful review of the manuscript by Rakhal D. Dave and useful discussions with Ulrich M. Muller. The Swiss National Science Foundation is gratefully acknowledged for its financial support