An Empirical Research On Multi-factor Stock Selection Model Based On SVM Algorithm

The paper is to study how to combine the traditional multi-factor model with machine learning algorithm to build a multi-factor model based on SVM algorithm.The model can construct stock portfolios through selecting valuable stocks in the CSI 300 index stocks in order to obtain steady excess return.This paper enriched methods of multi-factor stock selection,as well as provide good ideas and reference for other ways of stock selection.The paper selects stocks in the CSI 300 index from 2010-01 to 2016-12 and uses the cross-section data of last trading day of each month as the data sample,in which,the data from 2010-01 to 2012-12 is training sample of model parameters,and the data from 2013-01 to 2016-12 is used for back-test outside the sample.The process of the model construction is mainly divided into five parts:data preprocessing,factor validation test,model parameter optimization,the model construction and result analysis,model development and optimization.The cumulated yield and annualized yield of the model were respectively 85.37 percent and 20.54 percent in four years from 2013-2016,which were much higher than the yields of the benchmark of CSI 300 Index.Through ranking comparative analysis,we can found that the yields decreases significantly with rankings of descending order,hence,the model shows its significant effect of stock classification,which can effectively distinguish strong and weak shares.The paper compares the online learning model and offline learning model and found the performance of the excess accumulative total yields of two models respectively were different,which proved that it's better to classify stocks via online learning and constantly adapt to the change of market environment.In the part of development and optimization of the model,in order to guarantee the timeliness,the paper constructs a new dynamic model based on the original model,meanwhile,it also sets weights for the stock portfolio according to predicting probability.Both two ways improve the model performance.