An Empirical Study On Muti-factor Stock Selection Model Based On XGBoost Algorithm

Quantitative investment is an investment technology which is used to identify the opportunity cost of financial market transactions by advanced computer models.It has the advantages of dispersion,efficiency and discipline.This technology has been applied in overseas investment markets for nearly 40 years.Although the domestic stock market was established relatively late,with the improvement of financial legal system and computer technology quantitative investment technology has been applied more widely and extensively in the domestic market.In this paper,the XGBoost stock selection strategy based on Random Forest is used to conduct an empirical study on the CSI 300 component stocks.Combing the characteristics of factors associated with investments in the stock market,including technical indicators,income risk,valuation and other nine categories of factors.Taking July 1,2017 solstice and December 31,2020 as backtest interval,the interval is divided into 7 subintervals by rolling backtest method,while the current month and the previous 42 months are taken as the training interval.Firstly,the Random Forest method is used for feature dimension reduction,and the importance of each factor is calculated then,so as to screen out those factors which have important influence on the stock return rate.Based on the important factors selected above,a quantitative stock selection model combining with machine learning and multi-factor stock selection model is constructed.In this paper,logistic regression model,support vector machine model and XGBoost model are constructed respectively.In the process of empirical analysis,this paper compares the classification ability of the three algorithms in single phase modeling and the stability of the three algorithms during the period of backtest.The results show that XGBoost's classification ability and stability are better than the other two algorithms.And according to the backtest results the dimensionality reduction with Random Forest can effectively improve the investment benefit of the strategy.The three algorithms mentioned above are obviously better than CSI300 in terms of cumulative return rate and annual return rate,among which XGBoost algorithm strategy has the best performance.By comparing the return performance of other investment strategies of XGBoost model in different position numbers and position adjustment cycles after the reduction of the dimension,we found that the best performance was achieved by holding 100 stocks per month,and the excess return exceeds the CSI 300 benchmark return of 20.35%,and the annualized return is 14.57%.Finally,this paper optimizes above model according to the weight adjustment of probability,and the results show that the optimized model is superior to the original model in terms of cumulative return rate and annual return rate.