Research On The Mean Reversion Strategy Of A-shares Bank Stocks Based On Deep Learning

Fama put forward the efficient market hypothesis in 1970,which is one of the most controversial investment theories at present.A great deal of research has been done on the effectiveness of the stock market in China.Many articles point out that China’s stock market cannot meet the standard of weak efficient market.However,the actual effect of investment strategy based entirely on historical trading data is not ideal.This paper thinks that this is mainly due to the over-strict linear assumption of the traditional model,which affects the generalization ability of the model.Based on the idea of stock matching,this paper designs a mean regression strategy based on depth learning technology.With its good non-linear fitting ability and generalization ability,it can discover the information in historical price data to obtain excess profits and further demonstrate the ineffectiveness of the market.Deep learning is a multi-layer neural network based on artificial neural network technology,and we use long-short term memory model(LSTM),a special recurrent neural network in this paper.It has excellent nonlinear fitting ability and good generalization ability.Different from traditional approach of pairs trading,the strategy designed in this paper uses 16 bank stocks as the object of the transaction,and uses the deep learning model to fit the return of one stock with the return of the other 15 stocks.Then,we can calculate the long-term equilibrium level of each stock’s return rate with the well-trained deep neural network at each time point,and use the long-term average price of the stock rather than the historical average price as the real price of each stock.There are two strategies presented in this paper: the first one includes short selling,and the other one does not include short selling.The result shows that the profitability of the two strategies is equivalent,the average cumulative income during the trading period both exceeds 16%,and the profitability is significantly better than the general matching trading strategy,but the performance in different market environments is different.In addition,this paper compares influence of the three feature extraction methods of principal component analysis(PCA),independent component analysis(ICA),and autoencoder(AE).The results show that Single-DNN that does not extract any feature is performed.Single-DNN model achieved higher returns,but was not significantly higher than the other three models.This paper actually proposes a mean reversion strategy based on deep learning technology.Compared with the traditional pairs trading strategy,it has two major advantages: First,the deep learning model utilized in this paper is a powerful nonlinearity fitting ability,with good generalization ability and is conducive to capturing the long-term equilibrium yield of stocks;Second,different from the traditional paired trading strategy only make one-to-one pairs.The strategy presented in this paper make many-to-one pairs.The increase in the number of pairs of stocks brings about an increase in information and lead to a better profitability.At last,this paper finds that the analysis of the historical trading information of bank stocks can obtain excess returns through empirical results,which provides empirical basis for the fact that China’s stock market is not a weak efficient market,and analyzes the reasons from relevant theories.