Method Of Multiple-factor Stock Price Predictions With Deep Learning

With the development of economy,the prospering of stock market,the increase of the number of stock investors year by year,the demand for stock analysis has become increasingly prominent.Stock price predictions is one of the keys to the field of stock analysis.However,many of the models and methods for stock price predictions are not satisfactory,and they have not reached the standard of actual application.Thanks to the recent application of artificial intelligence technologies in the financial field,it has also promoted the development of stock price predictions.In order to further improve the accuracy of stock price predictions,this thesis introduces the multiple-factor model in quantitative into normal stock price predictions method and make some improvement.The summary of this thesis's main contents and researches are as follows:1.Focusing on the problem that the factor selection method in the general multiple-factor model is not effective and so the stock price prediction result is not ideal,this thesis proposes a stock factor selection method based on the GBDT and FFM fusion model.Using the model characteristics of GBDT and FFM to complete the stock factor selection,the effect of factor selection is improved.The method mainly includes two points:First,the original features are reconstructed by using the feature reconstruction characteristics of the GBDT model.The new feature set constructed by GBDT has a stronger ability to describe the stock than the original feature set.Second,using the feature cross characteristics of the FFM model,using the new feature set reconstructed by GBDT,the appropriate stock factor set are selected.Finally,the real stock dataset is used for experiments to verify the feasibility of the model.The set of factors unselected,the set of factors selected by traditional methods,the set of factors selected by single GBDT model,and the set of factors selected by the method proposed in this thesis are respectively used as input features to predict the stock price,the stock price prediction results by the factor set selected by the method proposed in this thesis is the highest,which proves that the method of this thesis can screen out the better stock factor set.2.In view of the fact that the general LSTM model cannot accept the input form of multiple-factor model,this thesis establishes a multi-factor LSTM stock price prediction model based on the general LSTM.By changing the mapping structure of the input layers,the recurrent layers and the output layers of the general LSTM and introducing dropout into the recurrent layer,the new model structure not only can adapt to the input form of the multiple-factor model,but also not lead to a significant increase in the time of model training.Finally,the experimental results on the real stock dataset show that,thanks to the multiple-factor model introduction into the LSTM,compared with the general LSTM model,it not only improves the accuracy of stock price predictions,but also brings better model robustness to some extent.3.Using the research results mentioned above,a mobile stock prediction and recommendation system is built.The model and method proposed in this thesis are actually running in this system,and its practicability is verified.