Reinforcement learning applied to Forex trading

This paper describes a new system for short-term speculation in the foreign exchange market, based on recent reinforcement learning (RL) developments. Neural networks with three hidden layers of ReLU neurons are trained as RL agents under the Q-learning algorithm by a novel simulated market environment framework which consistently induces stable learning that generalizes to out-of-sample data. This framework includes new state and reward signals, and a method for more efficient use of available historical tick data that provides improved training quality and testing accuracy. In the EUR/USD market from 2010 to 2017 the system yielded, over 10 tests with varying initial conditions, an average total profit of 114.0 ± 19.6% for an yearly average of 16.3 ± 2.8%.