Research On Stock Market Forecast Based On LSTM And News Sentiment

The problem of stock prediction has always been a hot research topic in the financial market.The scholar of behavioral finance believes that various economic behaviors of people will be affected by "irrationality",and emotions can affect individual behaviors and decisions profoundly.Since daily financial news reports will affect stockholders’ sentiments,their investment behaviors,and then stock prices,this paper proposes a stock short-term trend prediction model based on LSTM and news sentiment.The main research work of this paper is as follows:Firstly,we extract technical indicators from stock historical trading information,and construct an SVM model based on stock technical indicators as the benchmark model(denoted as model A).Then,we use the Vader method based on thesaurus and semantic rules to extract the news sentiment on corresponding trading days.The news sentiment and stock technical indicators are both used as input variables of SVM model(denoted as model B),and the accuracy of model A and model B in stock prediction is compared.Facts have proved that Model B is superior to Model A in all evaluation indicators,which also verifies the effectiveness of news sentiment in the stock prediction model.Secondly,based on stock technical indicators and news sentiment indicators,the traditional classifier SVM,ensemble learning method XGBoost and deep learning LSTM were used to build models(denoted as models B,C,and D respectively)to train and predict the rise and fall of stocks.Experiment shows that the LSTM model has the best effect.Its accuracy on the test set is improved by 13.1 percentage points compared with the SVM model and 4.5 percentage points higher than the XGBoost model.One of the reasons is that compared with the general neural network model,the LSTM model can learn the laws of long-term dependence more effectively,which also confirms the advantages of LSTM in stock prediction problems.Thirdly,we optimize news sentiment extraction methods and further improve the accuracy of the model.We take the improved news sentiment extraction method,which combines the Vader sentiment dictionary with the Loughran-McDonald Financial dictionary specific to the financial field,and stock technical indicators as input variables of a new model(denoted as model E)based on LSTM neural network.The model’s prediction accuracy of the test set reached 75.5%,which was an increase of 6 percentage points compared to the model D.This shows that the fused sentiment dictionary and the improved sentiment extraction method can extract the sentiment of the news more comprehensively and accurately,which will help improve the accuracy of stock prediction greatly.