| The stock market,as an essential part of the capital market,reflects the economic performance of the country and the growth of the industry.The prediction of stock market share price fluctuations helps establish a risk prevention mechanism,provides a basis for the investment behavior of a large number of retail investors in China’s A-share market,and assists relevant institutions in making important decisions to keep healthy economic and financial market conditions.In the financial field,a large number of institutions publish real-time domestic and international news information on their platforms,and these news texts can fully influence investors’ investment decisions and thus the stock market trend.This study uses text sentiment analysis methods in the field of natural language processing to mine and analyze the sentiment embedded in news texts to determine stock trends by quantifying the current market investors’ investment sentiment.Since publicly available financial text datasets are difficult to obtain,this paper obtains the financial news sentiment classification dataset(called WIND)from Github and the CCF2019 dataset from the "Financial Information Negative and Subject Determination" competition.The study uses a BERT pre-training model to fine-tune the two labeled financial texts to obtain a language model suitable for the sentiment classification task in the financial domain.In addition,the BERT model is distilled to the lightweight networks Bi LSTM and Text CNN using knowledge distillation for comparison experiments.The experiments show that the effect of both distillation and compression is higher than that of support vector machines and can explain more than 90% of the information in the original BERT model,which is superior.Also,the experiments find that the feature extraction capability of the BERT model in this task is sufficient and the effect exceeds that of the more complex ERINE and XLNet pre-trained models.By comparing the time consumption of model training and validation,it is concluded that the distilled model is faster in inference and takes up less computational resources.Based on the model comparison experiments,we use the BERT model,which performs well on both datasets,as a prediction model to obtain real-time news texts of Sina Finance through crawling techniques to obtain the sentiment tendency of the news texts and use them to construct a financial opinion factor.The closing price of SSE index is analyzed empirically to investigate whether there is a correlation between the opinion factor and the SSE index and the strength of the relationship.Using the SSE index as the main variable,the DTW distance between it and the public opinion factor,Baidu search index and Baidu information index is calculated by the dynamic time regularization algorithm(DTW),and the experiment finds that the DTW distance between the public opinion factor and the SSE index is smaller,which proves that the constructed financial public opinion factor has a stronger correlation with the SSE index under the conditions of time advance and lag,and illustrates the scientific and interpretable nature of this public opinion factor. |
