| Pairs trading is a classic market-neutral quantitative investment strategy that originated in the 1980s.Its profit model is to pair two stocks with similar historical price movements and continuously monitor the performance of the paired stocks.When their spread deviates from the historical average,short the relatively overvalued one and buy the relatively undervalued one,then close the position when the two return to their long-term equilibrium relationship.At this point,an arbitrage is completed.Since the profit of pairs trading depends on the relative value of the paired stocks rather than the absolute value,it is almost independent of the rise and fall of the market.Coupled with its intelligent and rational programmed trading model as a part of quantitative investment,the pairs trading strategy is highly favored by researchers and investors under the background of the current shocking downturn caused by the trade disputes between China and the United States.The research in the field of pairs trading always follows a classic two-step framework,the first step is the selection of pairs,the second step is the setting of trading signals.Both thought optimization and algorithm upgrade are based on this framework.The traditional pairs trading models and most of the related researches today tend to restrict the scope of pair screening to the same industry,which is convenient and efficient,but also leads to the loss of a large number of effective pairing opportunities across industries.In the trading signal setting stage,the traditional and current mainstream thinking is to determine fixed thresholds as trade timing based on historical experience.However,this may cause unnecessary losses due to mechanical guidance of trade.In this paper,we use machine learning techniques to make improvements in the above two stages of pairs trading.In the pair selection stage,breaking the industry restriction,we use Kmeans++cluster algorithm to find stock groups with similar attributes and target the search range of potentially effective pairs.We use cointegration analysis and Hurst index to test whether there is a long-term equilibrium relationship between stock prices,and pick the optimal pairs according to the test results.In the trading signal setting stage,we make improvements based on the traditional fixed threshold signal settings.When the spread touches the threshold,the LSTM neural network model is used to predict the spread trend instantly and determine whether this is the best point to open or close a position.Specifically,if the predicted next move is consistent with the current direction,no action will be taken to avoid losses,and if the direction reverses,the corresponding trading operation will be executed immediately.The optimized trading signal is more intelligent than the traditional threshold,which can effectively improve the return of a single transaction and thus enhance the overall return of the strategy.This paper selects CSI 300 constituent stocks as the empirical object and tests the effectiveness of the pairs trading strategy constructed according to the above process.The results show that using the CSI 300 index as the market benchmark,the cumulative return of the proposed strategy significantly outperformed during the backtest period,with an annualized return of 38.29%,a sharpe ratio of 1.668763,and a max drawdown of 11.91%.All three indicators exceed the benchmark,indicating that it is able to achieve higher returns while providing greater protection against risk.The success of the strategy confirms that the machine learning approach selected in this paper is appropriate and effective for the integration of applications at all levels of the pairs trading framework.Through this excellent matching of classical problems and advanced techniques,a fairly complete and flexible framework of pairs trading strategies is established,which is paving and inspiring for subsequent investment research.. |
