Research On The Construction Of Investor Sentiment Indicators Of Different Information Channels Based On Deep Learning And Its Prediction Ability To Stocks

In the 1980 s,behavioral finance theory rose quietly and began to shake the authority of capital asset pricing model(CAPM)and efficient market hypothesis(EMH).Behavioral finance theory holds that the market price is not only determined by its internal value,but also influenced by investors’ psychology and behavior to a great extent.Compared with foreign countries,the composition of investors in China’s stock market is more complex,and investors receive a wide range of sources of information.Information from different channels affects investors’ expectations and decisions on the stock market,resulting in fluctuations in the stock market.Therefore,the correct measurement of investor sentiment will not only help the regulatory authorities to monitor the market,but also help investors make investment decisions.So far,many scholars have begun to pay attention to how to mine investor sentiment from text data and construct investor sentiment indicators.In previous studies,in terms of data dimension,it is often only studied for a single information channel,but the channels for investors to receive information are diverse.With the development of the Internet,the massive text data,news,comments and Research Report texts will produce different emotional guidance for investors;In terms of text emotion analysis,at present,Bert model has achieved good performance in text emotion task,but there is still room for improvement.We can further optimize the construction of fin Bert model suitable for financial text from the aspects of training methods and model structure to improve the semantic learning ability of text;In terms of index construction,the existing research does not consider the impact degree and sustainability characteristics of information.In fact,the impact degree of information from different channels on investors is different,and may affect investors’ judgment and decision-making in the subsequent period of time.Based on the above background,this paper comprehensively considers the information from different channels,and uses news reports,stock bar comments and research reports to measure the macro information and the investor sentiment of retail investors and professional institutions on the stock market,which can comprehensively quantify the market sentiment.Firstly,according to the characteristics of news,comments and research reports,special data preprocessing methods are designed,and different emotion analysis models are established.For news and comments,a fin Bert emotion analysis model suitable for financial texts is constructed;For the research report,the semantic rule model is used to extract the text emotion.Then,considering the impact and sustainability of information on investors,this paper constructs daily investor sentiment indicators for news,comments and research reports by using half-life weighting method combined with text reading and institutional ranking.Finally,in order to explore the effectiveness of the investor sentiment index constructed in this paper,the Granger causality test is used to verify the causal relationship between it and stock return in a statistical sense;Taking the investor sentiment index as the input information,the LSTM model is used to explore its prediction ability of return.Combined with DM test,it is proved that the investor sentiment index based on multi information channels can significantly improve the prediction ability.