Intraday FX Trading: An Evolutionary Reinforcement Learning Approach

Abstract

We have previously described trading systems based on un-supervised learning approaches such as reinforcement learning and genetic algorithms which take as input a collection of commonly used technical indicators and generate profitable trading decisions from them. This article demonstrates the advantages of applying evolutionary algorithms to the reinforcement learning problem using a hybrid credit assignment approach. In earlier work, the temporal difference reinforcement learning approach suffered from problems with overfitting the in-sample data. This motivated the present approach.

Technical analysis has been shown previously to have predictive value regarding future movements of foreign exchange prices and this article presents methods for automated high-frequency FX trading based on evolutionary reinforcement learning about signals from a variety of technical indicators. These methods are applied to GBPUSD, USDCHF and USDJPY exchange rates at various frequencies. Statistically significant profits are made consistently at transaction costs of up to 4bp for the hybrid system while the standard RL is only able to trade profitably up to about 1bp slippage per trade.