A pair trading using reinforcement learning and wavelet decomposition

Abstract

In this study, we propose a trading optimization methodology for the pair trading strategy. The Johansen cointegration test and the correlation measure are used for pair selection. We apply Deep-Q-network (DQN) technique in which the trainable reinforcement learning agent is designed to directly control the trading positions. The maximum overlap discrete wavelet transformation (MODWT) algorithm is used for generating the trading signal from the spread time series. Wavelet signal preprocessing is used to extract the original time series into cyclic time series components and long-term behavior components. Based on the in-sample performance this trading model successfully solves a profit maximizing in the pair trading problem using wavelet components predictors. However, poor out-of-sample results observed in many sampled pairs indicate that the proposed procedure has an overfitting problem.