Reinforcement Learning Approach to Forex Trading using Price Changes Trend

An automated program that generates profits from the foreign exchange market is attractive to all forex investors. Reinforcement learning has been widely used in financial market trading to maximize profits. In this study, we interested to generate profit from high-frequency currency trading with Deep Q-Network. It has known that the state signals are extracted from forex price data, e.g. mean, maximum, minimum, standard deviation in the interval time. We introduced novel price change trend features that aim to capture directions and speed of price movement over different time and apply it to be the state signal for reinforcement learning. The state signals are consist of the following two parts: the market features that including novel slope of price changes trends features and the agent's current position features. There are three actions -open position, close position, and hold or do nothing position. The reward signal was formulated by tanh function that is used to maximize the profit. We tested our proposed features with Deep Q-Network in the EUR/USD market from 01-Aug-2021 to 30-Dec-2021. The experimental results shown that in comparisons of reinforcement learning with base line features, our proposed features outperforms better than for all performance metrics in both 3000 and 6000 ticks feature vector. These results confirm that the slope of price changes trends features are useful for forecasting under the reinforcement learning approach.