A Study Of Investor Sentiment Measures And Their Predictive Effects On Stock Index

In the 1990 s,Shanghai Stock Exchange(Shanghai Stock Exchange)and Shenzhen Stock Exchange(Shenzhen Stock Exchange)were established one after another,and China’s stock market ushered in rapid development.As of December 31,2021,the number of listed companies on the SSE and SZSE reached 4,608,and the market value of stocks in circulation reached 75,155,612 billion yuan.All kinds of data show that China’s stock market has become a pivotal player in the world financial system.In recent years,as more and more people speak through Internet channels coupled with the growing maturity of artificial intelligence(AI)technology,the extraction of valid information from online textual big data for application in the economic and financial fields has become a popular area of research at present.In the field of finance,methods of obtaining stock market information have also increased with the maturity of technological fields such as machine learning.In this paper,based on behavioral finance as the theoretical basis,we crawl the text data of stock commentary on Eastern Fortune online stock bar through web crawlers,use BERT model to classify the sentiment of the text data,deeply explore the information of individual investors’ sentiment contained in the text of stock commentary,and construct the investor sentiment index according to this.Finally,the built BERT-LSTM deep learning model is used to fit the prediction analysis of SSE index return.In this paper,in the part of investor sentiment analysis,the text is crawled with the text of the SSE index stock bar on the Oriental Fortune website,and more than 6million stock bar comments are obtained through web crawling technology.After that,the BERT model was fine-tuned with the training data with sentiment annotation,so that it can handle the sentiment classification of stock bar text,and SVM and other classification models were used to compare the accuracy of the classification effect with the main model used in this paper.Finally,the text is transformed into a measurable sentiment score to determine investor sentiment by combining the sentiment category of each stock comment and the method of quantifying investor sentiment.In terms of stock index return prediction,this paper uses LSTM neural network to learn to fit and predict the future return of SSE index.In order to demonstrate the effectiveness of BERT-LSTM based investor sentiment for stock index prediction,this paper empirically compares the accuracy of BERT-LSTM with support vector machine,random forest,XGBoost and other models in various situations including investor sentiment indicators and different market stages.In summary,the main points in the empirical part of this paper are summarized as follows:(1)In the analysis of investor sentiment text mining as well as index construction results,it is found that the highest accuracy of 83.86% is obtained for the stock review sentiment classification results with the BERT pre-training model used in this paper.This indicates that the BERT model used in this paper performs well in the stock comment text sentiment classification task from the modeling perspective.(2)In the correlation test between the daily index of investor sentiment and the SSE Composite Index,a significant positive correlation was found between the two.And the correlations between them differed significantly in different periods of rise and fall of the SSE index.At the same time,the degree of stock index rise also affects the correlation between the two,indicating that the correlation effect of investor sentiment is different in different degrees of stock market rise.(3)In the results of the classification prediction effect of investor sentiment on the future rise and fall of the SSE Composite Index,it is found that the XGBoost model has better prediction effect than other models among different classification prediction models without adding investor sentiment index.After adding investor sentiment features,the BERT-LSTM model used in this paper has the best classification prediction effect.No matter in which classification prediction model,the investor sentiment indicator constructed in this paper can improve the prediction accuracy of the model,which has important reference significance for using the investor sentiment indicator proposed in this paper for trading strategy design.(4)In the prediction results of investor sentiment on SSE Composite Index return values,it is found that the investor sentiment with a time window of three trading days contains the best validity of information on investors’ irrational investor behavior,and the prediction error of the prediction model incorporating the investor sentiment index is smaller than that of the benchmark LSTM model,which can better capture the characteristics of oscillatory market fluctuations and fit significantly better than the benchmark model.