Deep Reinforcement Learning Pairs Trading with a Double Deep Q-Network

Abstract

This research applies a deep reinforcement learning technique, Deep Q-network (DQN), to a stock market pairs trading strategy for profit. There is a need for this work, not only to further the use of reinforcement learning in stock market trading, but in many other areas of financial markets. The work utilizes a specific type of DQN, a Double Deep Q-Network to learn a pairs trading strategy. The DDQN is able to learn a cointegrated stock pair's mean reversion pattern, and successfully make predictions based on this pattern. Attesting that a reinforcement learning system, can effectively learn and execute a pairs trading strategy in the stock market. It also introduces a parameter, Negative Rewards Multiplier, during training that adjusts the system's ability to take more conservative actions. Based on the results, the next steps would be to employ this method in other financial markets, or perhaps use a DDQN to learn additional trading strategies.