Study Of Multi-factor Stock Selection Scheme Based On Boosting Algorithm

In recent years,the combination of financial trading strategy and computer science has become a research focus.Many scholars have conducted thorough research on quantitative stock selection,quantitative timing,intelligent risk management and so on.Recently,due to the fact that China's stock market has been impacted by the increasingly complex external environment and the pressure of both supply and demand of China's economic development,investors prefer to be stabilized and expect more robust returns.Therefore,the market urgently demands for the optimization of various existing investment models.Among many effective investment and trading models,the multi-factor model has been tested in multiple markets and periods,and has long been the preferred model for many actively managed funds.The multi-factor model could not only obtain excess returns by replacing the types and proportion of factors in time,but also show greater potential under the optimization of machine learning algorithm.The purpose of this thesis is to partially optimize the multi-factor model through cutting-edge technical means on the basis of predecessors' research,in order to obtain long-term and stable returns.Firstly,this thesis introduces the theoretical source and development process of multi-factor model,and expounds the principle of three boosting integrated machine learning algorithms: GBDT,XGBoost and Light GBM.Secondly,this thesis selects 9effective factors from the existing factor pool of the quantitative platform by using traditional methods such as IC value,line crossing rate,linear regression,hierarchical back test and economic significance judgment,and forms the initial factor combination with those selected factors.Then,this thesis uses the above three boosting machine learning classification algorithms to reduce the dimension of the initial factor combination,and selects three different factor combinations with the best performance.Finally,after backtesting the initial factor combination and three low dimensional factor combinations respectively,this thesis makes a comparative study on the return and risk indicators of each combination.The research result of this thesis is that the three machine learning algorithms can effectively reduce the dimension of the initial factor combination.After comparing with the multi-factor stock selection scheme constructed by the initial factor combination,it can be found that the return and anti-risk ability of the optimized factor combination stock selection scheme are higher in the back-test demonstration.At the same time,investors can choose different algorithms to optimize the model according to different preferences.In short,this thesis establishes a stock selection scheme from the perspective of optimizing factor combination,which provides a reference for market participants.