Stock Price Forecasting Method Based On Feature Selection And Income Decomposition

Stock is a common way of investment in daily life.It is very valuable for investors to accurately predict the trend of stock price.The stock market fluctuates greatly and its style changes constantly.The features selected by the conventional filter feature selection method are locally important for the whole stock price prediction period;Secondly,the current common stock price forecasting model is to directly fit the feature data with the stock price or return rate that needs to be predicted,which can not dig out the potential return of factors in a deeper level.Therefore,in order to solve the problems of the locality of feature selection and the lack of deep mining of factor potential returns,this thesis proposes the feature selection method of mean decrease rate of return and the stock price prediction model of income decomposition.(1)For the problem of locality of current feature selection,this thesis proposes a feature selection prediction model based on the mean decrease rate of return.By comparing the income changes in the global time period,the factors that have a great impact on the income changes are selected out.In this way,the selected features are suitable for the global situation,have a large overall contribution in a long time period,and can cover more types of market styles.In the stock simulation trading with CSI 1000 constituent stocks as the stock selection pool and a time span of 18 months,the return of the method of retraining after feature screening can be increased by about 30% compared with the original situation,indicating that such a global feature screening is helpful for stock price prediction.(2)In order to further mine the hidden information of factors and improve the prediction ability of factors,this thesis decomposes the predicted return rate based on the factors used according to the multi factor model to form two parts: linear return and nonlinear return.Based on the existing machine learning prediction methods,a decomposition prediction model of stock return is proposed.The model models and fits the two decomposed benefits respectively,and uses the composite model to get the final prediction results.This decomposition prediction model can well mine the nonlinear prediction ability of factors.Through experiments,we can also prove that the yield decomposition prediction model is helpful for stock prediction.Compared with the original prediction method,the yield can be increased by about 30%.Combined with the feature screening method of the average yield decline method,the yield can be increased by about 55%.(3)In order to verify the practical application effect of the stock price prediction model proposed in this thesis,this thesis designs and implements a lightweight financial data query and quantitative strategy back testing platform,which realizes the functions of data download and query,simulation of stock trading,trading result analysis and so on;Through the back test platform verification,the two proposed models can increase the original yield by 30% and 55%respectively.Based on this platform,the feature selection prediction model based on the mean decrease rate of return and the simulation board strategy of the income decomposition stock price prediction model are realized.