Empirical Research On Paired Trading Strategy Based On Reinforcement Learning

Since the birth of the Chinese A-share market in the 1990 s,after nearly 30 years of development,it is still in an immature stage.After entering the twenty-first century,China's stock market has experienced two major crises in 2008 and 2015.The entire market has fluctuated greatly,making most investors miserable,and many new entrants have become one crop after another.In this environment,China launched a margin trading system in 2010,and pairing trading strategies began to take the mainland by storm.Paired trading is the most important trading strategy in statistical arbitrage.During periods of low market effectiveness,this strategy has the opportunity to obtain more returns.However,with the continuous improvement of market efficiency,traditional fixed-parameter trading models have been unable to guarantee that pairing transactions have always achieved greater returns.The parameters of the trading model need to be optimized,but also dynamically and automatically adjusted.To a certain extent,it is necessary to research and develop a trading model with dynamic optimization parameters of machine learning attributes,which is of great significance for improving the profitability and execution efficiency of the paired trading model.In the overall demonstration process of this article,the main constituent stocks of the CSI 300 Index are used as samples.Compared with the results of the traditional pairing trading strategy and the improved pairing trading strategy of the reinforcement learning algorithm,it is tested whether the latter can obtain more High yield.First select the CSI300 index stocks as the sample stocks.The sample time range is from January 1,2006 to December 31,2016.133 stocks are retained,and the daily closing price data of each stock is analyzed.Z-Score is standardized,and according to the CSI Tier 1 industry classification,133 stocks are classified into the top ten industries.Among them,27 are in the real estate industry,5 in the telecommunications business,17 in the optional industry,21 in the industrial industry,9 in the information technology industry,16 in the medical and health industry,9 in the major consumer industry,20 in the raw materials industry,and 4 in the energy industry.Only five are in the utility industry.Then,during the formation period of the trading strategy,the stocks in each industry were modeled using the Copula model,and stock combinations with correlation coefficients of 0.7 and above were retained,followed by unit root test and co-integration test,and finally those who passed the co-integration testwere left.Finally,during the trading period of the trading strategy,a traditional paired trading model and a paired trading model based on reinforcement learning are separately designed,and the cumulative returns and information ratios of the two models are compared.The former uses fixed parameters,that is,the value of fixed trading signals,including opening and closing positions and take profit and stop loss,while the latter uses internal algorithms to optimize various parameters in real time in order to achieve greater returns.From the results of empirical analysis,it can be seen that paired trading strategies can still obtain considerable returns in China's A-share market.After the introduction of reinforcement learning algorithms,the yield and information ratio have been greatly improved.Combining reinforcement learning algorithms and paired trading strategies,a new paired trading strategy based on intelligent algorithms is further designed to achieve effective adjustment and optimization of trading models.Although the traditional model can still obtain good benefits,it is declining.Based on case studies,it can be found that such new models are beneficial to increase the overall level of income and performance.In the case where the short-selling system is becoming more and more perfect,the new model can be used as a more valuable arbitrage tool.