The Study On Multi-Factor Quantitative Stock Selection Based On Self-attention Neural Network

In recent years,with the significant improvement in the computing power of computer hardware and the rapid development of Artificial Intelligence,quantitative investment has begun to emerge in the Chinese financial market.With the development of big data,the advantage of machine learning has been gradually demonstrated in the field of quantitative investment.Among many quantitative investment strategies,multi-factor stock selection strategy has become one of the hot issues in the field of quantitative investment by virtue of its high stability and wide coverage.At the same time,the powerful nonlinear fitting ability of the neural network model improves the performance of various tasks in the machine learning field.The construction of multi-factor quantitative stock selection strategy based on the self-attention neural network model is studied in this paper.Based on the financial big data platform‘Tushare Pro'and the quantitative trading platform‘JoinQuant',the daily data related to the constituent stocks of CSI 300 index is selected from October 2009 to October 2018 as the research object in this paper.In order to fully consider all factors affecting stock price volatility,the market factors,financial factors,technology factors and investor sentiment factors are selected to form an initial factor set.In order to ensure the quality of data utilization,some preprocess related to hysteresis,missing values and standardization is performed.At the same time,based on the idea of model ensemble in machine learning domain,68 factors are selected for the construction of the stock selection model,comprehensively considering Pearson correlation coefficient,distance correlation coefficient,Elastic Net based on AIC criterion,Elastic Net based on BIC criterion,random forest and GBDT.Finally,the factor data of the past 60 trading days is used to predict the price trend of the CSI 300 constituents in the next month.The top 50stocks are selected to construct the portfolio with equal weight allocations each time,according to the predicted rising probability.Meanwhile,the portfolio is updated monthly.In order to ensure the generalization ability of the model,some regularization methods of₂ regularization,dropout and layer normalization are used in this paper.In the entire back-test period from February 2015 to October 2018,the multi-factor stock selection strategy based on the self-attention neural network model constructed in this paper is significantly better than the performance of CSI 300 index.Compared with 1.69%annualized loss of the CSI 300 Index,the strategy achieved an annualized return of 27.02%.The maximum retracement of this strategy is-22.10%,significantly lower than the maximum retracement of the CSI 300?-46.70%?.Considering the evaluation of Sharpe ratio,information ratio and other indicators comprehensively,both high yield and low risk have been obtained from the trading strategy constructed in this paper.At the same time,CSI 300stock index futures is proposed to introduce into this strategy to eliminate systemic risks and obtain more stable excess returns.When the whole back-test period is divided into different trend stages to evaluate the strategy constructed in this paper,the price trend of each constituent stock can be predicted by this strategy in the rising trend and the oscillating trend stage,with the₁ score of 0.89and 0.91 respectively.The higher cumulative return rates of 61.59%and 59.88%are achieved from this strategy than CSI 300 Index.And the trading risk of this strategy is lower than CSI 300 Index's,keeping the maximum retracement at-7.01%and-6.99%respectively.In the downward trend phase,this strategy controls the cumulative loss at 12.73%compared to the cumulative loss of 43.13%in the CSI 300 Index.Considering the evaluation of Sharpe ratio,information ratio and other indicators comprehensively,the balance of return and risk has been obtained from the trading strategy constructed in this paper in different trend stages.