Systematic Analysis And Prediction Of Stock Excess Returns Based On Improved Multi-factor Model

The research on stock returns has been a hot topic in recent years.Among them,the most widely used is the multi-factor model.The fama-french three-factor model and the five-factor model are the most typical.However,these models were generated in the context of western stock markets,and their applicability to the Chinese stock market is not strong.In the literature of fama-french,it is also pointed out that whether the three factors or the five factors cannot fully explain the stock returns rate.This shows that there are other unexplored factors that affect stock returns.In terms of calculation methods,the use of a simple linear model for time series regression is not effective,and the relationship between various factors at each time cannot be considered.Based on these shortcomings,this article will improve the multi-factor model in the following two aspects.First,emotional factors are added on the basis of the three-factor model.The existing well-known investor sentiment factors include the bw index and the CICSI index.The proxy variables used are generally the number of new investor accounts,consumer confidence index and other variables that can indirectly reflect investor sentiment.The construction methods are generally principal component analysis and partial least squares.This article hopes to use the Internet public opinion data as the main proxy variable to combine the two methods,and use the results obtained by the principal component analysis method as prior information to construct the weights of each proxy variable,and then use the partial least squares method to extract effective sentiment factors from online public opinion data,and from the goodness of fit and Two aspects of correlation verify its effectiveness.Second,var is used to solve the multi-factor model and the improved multi-factor model.In order to find the sparse solution of the model in high-dimensional cases,we also introduce the varlasso method.Finally,the model effect is compared by backward forecasting the sum of squared errors of ten periods,and its stability is analyzed.The conclusion is that the multi-factor model with improved emotional factors works best.And using lasso to solve the var model works better.