Research On Cointegration Arbitrage Strategy Of High Frequency Trading Based On Machine Learning

Paired trading is a market-neutral investment strategy with low risk and stable returns,which was born in the United States in the mid-1980 s.It is mainly through matching the assets with long-term equilibrium relationship between the price series.When the price difference of the asset pair deviates from the equilibrium,the position is built,and when it returns,the position is closed for profit.At the end of March 2010,China began the pilot work of margin trading,and the A-share market ushered in the era of short selling,making the application of domestic matching transactions possible.Therefore,this paper studies the matching trading strategy based on the 300 components of Shanghai and Shenzhen A-share market in China.In this paper,the paired trading technology is applied to the A-share market.In the stock pool,there are the one-minute closing prices of 167 securities traded on the Shanghai Stock Exchange among the 300 constituent stocks of Shanghai and Shenzhen.Take January 3,2020 to December 31,2020 as the inner range of the sample,and January 4,2021 to July 4,2021 as the outer range of the sample.It is mainly analyzed from three aspects: asset transaction pair selection,transaction model design,and transaction back-test.(1)In terms of selection of asset trading pairs: In order to expand the industry scope of the stock pool while avoiding the surge of calculation,the stock pool is pre-divided by K-Means clustering and OPTICS clustering respectively according to the stock return sequence after dimensionality reduction,and the trading pairs with long-term cointegration relationship are selected from the cluster to form the trading portfolio by combining the cointegration analysis.The advantages and disadvantages of two stock pool pre-division methods are analyzed.(2)The design of the trading model: in view of the shortcomings of the traditional strategic model,that is,the standard deviation of the historical price difference series is used as the trading signal of the whole trading period,and the signal reliability is low.This paper constructs a trading strategy that uses the rolling volatility predicted by the LSTM model as the trading signal to explore whether the profitability and risk aversion ability of the new strategy have been improved.In order to further improve the prediction ability of the volatility of the price difference series,the GARCH model with good explanation performance for the volatility aggregation of the financial time series is introduced,and the LSTM-GARCH fusion model is constructed to explore whether the fusion of machine learning models with different performance advantages and time series models can improve the prediction accuracy of the standard deviation of the series,thus improving the performance of the trading strategy.(3)In terms of transaction back-testing: from the two aspects of income and risk,the corresponding indicators are designed as the evaluation criteria for the performance of the trading strategy,and the back-testing results of the trading strategy based on the new model are compared with the traditional trading strategy.At the same time,take the asset appreciation and indexation strategy as the benchmark investment strategy to compare and evaluate whether the transaction strategy based on the new model is superior to other asset allocation methods.The empirical results show that:(1)The trading pairs selected by the stock matching method combined with clustering and cointegration method are highly effective,and the trading pairs selected by this method in the same industry and non-same industry can generate arbitrage income.Among them,the average annual yield of the trading pairs selected based on K-Means clustering algorithm and OPTICS clustering algorithm under the traditional trading strategy reached 34.86%and 31.21% respectively.(2)There is no obvious difference in the profitability of the trading pairs selected by the two stock matching methods,but compared with K-Means clustering,the stock pool pre-division method based on OPTICS clustering algorithm has stronger stability.(3)Considering the transaction cost,compared with the traditional trading strategy,the paired trading strategy under the new model has a higher annualized yield and Sharp ratio,and a lower maximum withdrawal rate.It is believed that the latter has better profitability and lower risk.(4)Compared with the single LSTM model,the LSTM-GARCH hybrid model has better prediction ability for the standard deviation series.On the whole,the trading strategy based on the hybrid model has better profit performance.The possible innovations of this study are mainly reflected in:(1)In terms of the selection of trading pairs,in view of the limitations of traditional methods that are not suitable for processing large amounts of data and can only find tradable stock pairs in the same industry,this paper proposes a stock selection method that uses machine learning clustering method to pre-screen the entire industry stock pool to form a relevant cluster,and then uses the co-integration method to pair the stocks within the cluster,The expansion of the stock industry to the whole industry has improved the efficiency of computing.(2)In terms of trading model,based on the traditional model,the long and short memory neural network model is used to predict the volatility of the price difference,and a strategy model is proposed to predict the fixed multiple of the volatility as the trading signal.(3)In terms of volatility prediction,considering the limitations of machine learning model lacking financial theory support,it is proposed to integrate financial time series model and machine learning model to build GARCH-LSTM volatility prediction model,and further explore whether the fusion model has higher prediction accuracy and can improve the effect of strategic returns.The model proposed in this paper provides new ideas and methods for the practice of paired trading strategies.