| Reinforcement learning is a kind of machine learning method for solving sequential decision-making problems.Through the continuous "interaction-trial and error" mechanism,it realizes the continuous interaction between the agent(Agent)and the environment,so as to learn the optimal strategy for completing the task.Deep reinforcement learning combines the perception ability of deep learning with the decision-making ability of reinforcement learning,and can make end-to-end perceptual decision-making in a complex and high-dimensional state space.Pairing trading strategy is a classic statistical arbitrage strategy.The analysis framework of the pairing trading strategy includes:matching of the same industry,matching of fundamentals,matching of upstream and downstream industry chains,matching of underlying assets of the same company on different exchanges,etc.This paper combines the DQN(Deep Q Network),Double DQN,Prioritized Experience Replay(DQN),Dueling DQN algorithms in reinforcement learning with paired trading strategies,and makes full use of the algorithm advantages of reinforcement learning’s self-learning and self-renewal to construct a class of Cointegration pair trading strategy with artificial intelligence attributes,and empirical research is carried out with the US stock market and Chinese futures market as samples.In research based on the U.S.stock market,the trading effect of the new strategy is overall better than the classic pair trading investment strategy(GGR)and other benchmark strategies.The Sharpe ratio,annualized rate of return and other indicators have been greatly improved,and better investment returns have been achieved.At the same time,it has a lower drawdown and reduces investment risks;the new strategy can automatically select the timing of opening and closing positions,and has better Capture potential trading opportunities and generate relatively stable positive returns in the short and medium-to-long term;paired trading investment strategies based on Double DQN,Prioritized Experience Replay(DQN),Dueling DQN algorithms have faster training speed,better convergence,The generalization ability is stronger.In the research based on China’s futures market,this paper further introduces deep neural networks such as GRU and CNN into strategy design,and constructs a deep reinforcement learning paired trading investment strategy based on GRU-DQN and CNN-GRU-DQN.In the paired trading of stock index futures-stock index futures ETF,the yield and risk resistance of the new strategy are also better than other strategies.The study also found that using a more complex deep neural network structure can more effectively mine potential features,which has positive significance for improving policy performance.This paper combines artificial intelligence algorithms such as reinforcement learning and deep neural network with pairing trading strategies to construct an artificial intelligence pairing trading strategy with adaptive ability,which effectively solves the defects of classic pairing trading and improves the profitability of pairing trading strategies.The ability and transaction efficiency also expand the application scenarios of artificial intelligence algorithms in quantitative trading problems.With the continuous development of my country’s margin financing and securities lending market,the new strategy can also provide investors with an effective arbitrage method and risk control tool. |
