Quantitative Stock Selection Model Via Machine Learning

With the continuous development of information technology,various data processing technologies and machine learning algorithms are widely used in the financial field,which is profoundly affecting all aspects of the industry.The machine learning model has the mechanism of processing information in parallel,and has the advantages of strong self-learning and self-adaptive ability,so that the "uncertainty" in reality is no longer an obstacle but a "resource" that can be used,and has a broad theory development space and huge practical value.With the opening of the domestic financial market and product innovation,the quantitative long-term strategy represented by multi-factor stock selection will develop more and more.Therefore,this paper introduces machine learning into the multi-factor stock selection model to study its performance in the factor investment field.On the basis of referring to a large number of authoritative literatures in the field of factor investment and empirical asset pricing,this paper constructs 57 anomalous factors based on the daily frequency individual stock data of the A-share market from January 2010 to August 2021,and back-tests the performance of the long-short return portfolio in A shares.Afterwards,this paper uses these factors as feature inputs to systematically compare the stock selection performance of different machine learning models.In this paper,12 models including OLS and Lasso,Ridge,Elastic Net,Gradient Boosting Tree,Random Forest,Multilayer Perceptron,RNN,LSTM and GRU are selected,training with a 12-month rolling window.Then build a long-short portfolio based on the trained model for backtesting.The empirical results show that,compared with traditional linear regression,the nonlinear learning ability of machine learning has a significant advantage in dealing with a large number of factors,resulting a better performance than that the linear regression.The investment portfolio constructed based on the multi-layer perceptron with a single hidden layer(MLP1)can obtain a annualized return of 17.72% for long portfolio and 9.576%for annualized long-short portfolio.This paper also examines the importance of factors in predicting stock returns in the A-share market,and finds that the liquidity factor is the most important.This paper attempts to combine machine learning with multi-factor stock selection,which can effectively improve the predictive and explanatory power of the model,and help to mine effective information beyond traditional linear models.At the same time,according to the market characteristics reflected by the importance of factors,this paper also puts forward suggestions for the improvement of the A-share market system.However,limited by the computing power,the machine learning model in this paper still remain to be optimized.Secondly,introducing more different factors into models is also one of the directions for future research.