Research On The Application Of Deep Learning In Multi-Factor Stock Selection And Trading

The price of stocks is affected by many factors.Finding and using these factors for stock selection and strategy formulation has gradually become a hot issue in the field of financial investment.However,since its establishment in 1990,China's A-share market has accumulated a large number of stock return data and impact factor data,which is difficult to take artificial selection.It needs to learn the massive factor data information by means of deep learning model,and deeply explore the linear and nonlinear relationships between stock return and impact factor,and based on the analysis results to build stock trading strategy to guide investment practices.Firstly,this paper defined 170 detailed indicators from three aspects:company value index,technical index and investor sentiment index,and then uses GBDT model to conduct single factor stock selection backtest,and found that only 23 factors select the portfolio to produce greater than zero.income.Based on the annualized rate of return,Sharpe ratio and other indicators,the impact factors were initially screened,the candidate factor combinations were constructed,and the IC value analysis method was used to screen the factors.Finally,19 factors including the market value of circulation were formed to form a multi-factor combination.Combined with the GBDT model,the semi-annual multi-factor stock picking back test finally obtained an annualized yield of 9.58%,which was higher than the gain of the Shanghai and Shenzhen 300 Index and a six-factor combination and an eight-factor combination for comparison during the same period,and the stock picking effect was better than the random forest model.However,the prediction accuracy of this model was relatively low,and the strategy had problems such as large income volatility and low Sharpe ratio.Secondly,in order to solve the problem of insufficient model prediction,the LSTM model was used to deeply mine the stock pool selected by the GBDT model to judge the possibility of stock rise.Based on the dual-model multi-factor stock selection strategy,in the semi-annual stock picking transaction,the annualized return rate of 3 1.4%was obtained,and the Sharpe ratio was 0.95.It showed that after adding the LSTM model,the stock yield prediction effect was better,and the income was greatly improved.However,there was also the problem that the maximum retracement value increases and the risk of the strategy rised.Finally,considering the positive correlation between the stock return rate and the market index,and the stock return rate is likely to fall when the market index falls.Therefore,a large-scale risk control strategy was formulated,which formed a complete stock selection transaction process based on the idea of "multi-factor system construction-stock pool screening-combination forecast-risk control".In the empirical study,the LSTM model got a good predictive effect on the Shanghai and Shenzhen 300 index yields.In the half-year backtest,the annualized rate of return of the strategy was 30.40%,and the Sharpe ratio had risen to 1.01,in terms of strategic returns and risk compensation.A big improvement had been achieved.In the overall effect of risk control,the market risk control strategy was better than the trailing stop loss strategy.