Pairs Trading Strategy Based On Deep Reinforcement Learning

Pairs trading is a very popular quantitative investment strategy,which consist of two steps: choice transaction pairs and discover trading opportunities.At present,the research on the selection of transaction pairs at home and abroad has been relatively thorough,mainly including distance approach,stochastic spread approach and cointegration approach.However,with the development of science and information technology,the frequency of transactions has become higher,in the face of massive historical transaction information,it is difficult for investors to find trading opportunities in the noisy dynamic financial market.This paper proposes a pairs trading strategy based on deep reinforcement learning,which combines the deep reinforcement learning model with the pairs trading and can adaptive market changes,First,the assets are pre-selected through correlation analysis,and then the EG two-step method is used to co-integrate the pre-selected assets.Finally,the deep reinforcement learning model automatically discovers the trading opportunities.The main contributions are as follows: one is to combine the deep reinforcement learning method with the pairs trading,and the Actor-Critic method was adopted in the deep reinforcement model,which can perform single-step update and the output strategy is relatively stable.The other is to use the LSTM method to automatically discover dynamic market features,avoiding the subjectivity of artificially extracted features and greatly improving the ability to find trading opportunities in pairs trading.In the end,this paper selects the representative digital currency with high transaction frequency and scattered market as the experimental object.The experimental results show that the strategy has good theoretical and practical value in the scene of high-frequency automatic quantitative trading.