Robustness test of genetic algorithm on generating rules for currency trading

In trading in currency markets, reducing the mean of absolute or squared errors of predicted values is not valuable unless it results in profits. A trading rule is a set of conditions that describe when to buy or sell a currency or to close a position, which can be used for automated trading. To optimize the rule to obtain a profit in the future, a probabilistic method such as a genetic algorithm (GA) or genetic programming (GP) is utilized, since the profit is a discrete and multimodal function with many parameters. Although the rules optimized by GA/GP reportedly obtain a profit in out-of-sample testing periods, it is hard to believe that they yield a profit in distant out-of-sample periods. In this paper, we first consider a framework where we optimize the parameters of the trading rule in an in-sample training period, and then execute trades according to the rule in its succeeding out-of-sample period. We experimentally show that the framework very often results in a profit. We then consider a framework in which we conduct optimization as above and then execute trades in distant out-of-sample periods. We empirically show that the results depend on the similarity of the trends in the training and testing periods.