Commodity Futures Pair Trading Strategy Design Based On Reinforcement Learning Algorithm

As a trading strategy,statistical arbitrage can avoid certain market risks while generating risk-free profits.Among many statistical arbitrage strategies,the paired trading strategy has received extensive attention and application,and it is also a major trading strategy.Generally speaking,pair trading is to construct long and short positions of paired assets first,so as to obtain some income from asset spread convergence.One of the obvious advantages of this trading strategy is that it can avoid some of the systemic risks of investing in the form of a hedging mechanism,which means that even when the financial market as a whole is in a downward state,paired trading can be profitable..One of the characteristics of the operation of the paired trading strategy is that the securities market is required to have a short-selling mechanism,but the stock financial market in our country only allows longs and does not allow shorts,which is a unilateral operation mode.Although my country began to gradually implement the margin trading and securities lending business in 2010,the securities lending business requires a higher cost,which makes it difficult for the matching trading strategy in my country's financial market to make a difference.The futures market has a good short-selling mechanism,and in recent years,my country's financial futures market has gradually enriched products,making the matching trading strategy more likely to be applied to the futures market.Based on the above,this article takes the domestic futures market as the research object and conducts an empirical study on the paired trading strategy.This article combines the reinforcement learning algorithm with the traditional pair trading model.The main improvement is to use the dynamic parameter optimization method to replace the fixed parameter method in the traditional pair trading model,and apply it to the Chinese futures market,making the trading strategy of this article increase Some profit opportunities.Using the daily closing price data from January 2010 to January 2020 in the Shanghai Futures Exchange,first conduct a liquidity analysis on the futures data during the training period to screen out products with better liquidity;then correlate the pairwise combinations Cointegration analysis is performed on matching pairs whose correlation coefficient is greater than the threshold(the correlation coefficient threshold is set to 0.85 in this article),and finally 6 matching pairs with better cointegration are screened out,and these 6 final pairing combinations That is the research object of this article.Then,after calculating the matching contract ratio for the final pairing combination,the traditional static pairing trading model's trading strategy signals are formulated,and the initial parameters of the dynamic pairing trading strategy based on reinforcement learning are initialized.The two models are used to conduct research on the research objects.Empirical research,and calculate the indicators of profitability,such as the cumulative return rate,the Sotino rate,and the information rate,based on the data of the training set and the test set.Through the research results,we can find that the paired trading model based on the reinforcement learning algorithm has good performance in the results of the six paired combinations,and compared with the traditional model,the new model significantly promotes the increase in the rate of return,and there are also investment risks.The reduction has the ability to continue learning.This article combines the reinforcement learning algorithm and the pairing trading strategy to design a new pairing trading strategy based on the Sarsa algorithm,so as to achieve the purpose of effective adjustment and optimization of the trading model,and help to improve the profitability of the traditional pairing trading strategy.The shortcomings,and provide investors with more effective arbitrage tools.