Research On Multi-factor Stock Selection Based On Updatable Variable Selection Method

With the continuous expansion of the financial market and the continuous disclosure of effective information in the stock market,quantitative investment theories that can bring stable investment returns that are higher than market returns have come into people's sight and have received wide attention from investors.Among the quantitative investment models reused in the market,the multi-factor stock selection model has attracted the most attention from investors,brokers and scholars.The multi-factor model combines CAPM and APT two major financial investment theories,namely capital asset pricing and arbitrage pricing theory.This model seems to realize the construction of investment portfolio by screening out the effective factors used to explain the stock return rate.By actively selecting stocks and holding stocks for investment portfolios,the goal of obtaining excess returns can be achieved.Under the rapid development of statistical theory and machine learning,multi-factor stock selection has begun to decompose and use a variety of computer algorithms,thereby continuously updating the multi-factor stock selection strategy.In data research work,streaming data has a large amount of data,fast update speed,and historical data occupies a long memory,which brings huge difficulties to the research,and the updateable estimation can effectively solve the timeliness problem of streaming data modeling.At the same time,in the study of multi-factor stock selection,transformation scholars use the stepwise regression method or the method of adding the Lasso penalty item to select factors,and add the Lasso penalty.In practice,the method of item is likely to lead to a mixture of biased estimates and the use of stepwise regression methods,and the SCAD penalty function with asymptotically unbiased estimation properties can more effectively select variables and apply it to quantitative stock selection It is worth exploring.Fundamentally,although scholars and scholars have studied the optimization of multi-factor models,such as the addition of neural networks,logical forests,and the addition of lasso penalty items,a large amount of work has penetrated and the subsequent updated data information cannot be effectively used.Due to the classic OLS stock selection,the improved model based on the lasso stock selection model and the XGBoost model fixes the training set,while the multi-factor stock selection model based on the updatable variable selection method(denoted as "SCAD update")can store some summary statistics The training set can be updated in a rolling manner according to the characteristics of the model while retaining effective information and reducing memory.Therefore,this article considers the multi-factor stock selection research in combination with the renewable estimates of variable selection.The research object of this paper is the Shanghai and Shenzhen 300 shares.The OLS multi-factor stock selection model,the lasso-based stock selection model and the multi-factor stock selection model based on updatable variable selection(SCAD update)are used for empirical research on stock selection.In the former application,the Fama-Macbeth test and the selection factor stage of the correlation test will be carried out,and the redundant factors will be eliminated through the forward regression method to obtain the relevant effective factors,and then the OLS stock selection model will be established.In the empirical back-testing stage,this model is used in stock return forecasting.According to the forecasted return rate,15 stocks are dynamically screened out,and then the portfolio merger is completed.After 33 periods of dynamic portfolio holdings,the corresponding performance of the five models is relatively compared.The results found that:(1)In terms of the CSI 300 benchmark,the benefits of these five models are all real,and we see the feasibility and effectiveness of these five models in the domestic stock market;(2)From the comprehensive performance In terms of analysis,the income is based on the selection of stock selection models based on updatable variables,Lasso stock selection,multiple regression models,SCAD stock selection and XGBoost stock selection segmentation,and at the same time based on the risk portfolio substitution of the update variable selection stock selection model,The stock selection model is more robust and the winning rate is higher.