Research On Multi-factor Quantitative Stocks Selection Strategy Based On Stacking Ensemble Learning Framework

Through the establishment of accurate and complex mathematical models,quantitative investment technology can enable investors to dig and analyze the rules in the historical data of stocks,so as to develop a systematic,procedural and automated trading strategy.The traditional multi-factor stock selection method analyzes the linear relationship between the financial factors of listed companies and the stock price,and regularly buys and holds the asset portfolio that is expected to rise,so as to obtain profits.However,as the data volume and data dimension of the factors in the stock market become higher and higher,the traditional linear regression method is difficult to effectively capture the nonlinear relationship between the increasing high-dimensional factor data.Therefore,it has certain research significance to combine artificial intelligence algorithm with traditional multi-factor investment method and construct quantitative investment strategy with excess return ability.In this paper,the effective factors are firstly mined,and the daily factor data of each component stock of Shanghai and Shenzhen 300 and China Securities 500 from 2018 to 2021 are selected with the help of the width-quantization data interface.Through data pretreatment and factor validity test,Twelve effective candidate factors were screened from 60 candidate factors including valuation,capital structure,earnings,growth and technology,and a multi-factor regression model was established for market capitalization(logarithm).In order to break through the prediction effect of a single model,when constructing a multi-factor model,the LGL deep learning integrated model is obtained by fusing LSTM and GRU deep learning algorithms as base learners in the first layer and linear regression LR algorithm as meta-learners in the second layer,based on the low-level logic of Stacking integrated learning framework.In the training model,select fixed window rolling training prediction method to predict the stock market value.In the stage of designing quantitative stock selection investment strategy,the candidate stocks in the stock pool are sorted from large to small according to the expected stock price growth range in the future transfer period,and the top fixed number of stocks are selected for holding in each transfer period.In addition,taking 1million as the initial investment capital and considering the transfer cost of the transaction,we set up multiple groups of horizontal comparison experiments for different number of shares,different transfer periods,different stock pools and risk control strategies.Experimental results show that the presented LGL deep learning integration model on the csi 500 stock pool with 30 day fixed tone warehouse cycle each with 10 holdings and add tracking stop-loss risk control strategy for investment effect is best,and the model under different combination strategy parameters were significantly better than the market gains and the control model of earnings,With higher returns and lower risks;The annual return and anti-risk ability of the investment strategy with the addition of tracking stop-loss risk control are significantly better than those without risk control.