Research On Stock Price Trend Prediction Based On Text Topic Recognition

With the development of market economy,the stock market grows rapidly and occupies a significant place in the national economy.Therefore,stock investment becomes an important way for investors to benefit from financial management.Stock investment has the characteristics of high risk and high return.The situational awareness of stock price not only helps the government supervise the stock market,but also helps investors avoid risks and obtain returns at the same time.Thus,the prediction of stock price trend has become the focus of the government and relevant departments.The stock comments released by experts are the authoritative reference basis for predicting the future trend of stocks.However,a large number of expert stock comments will appear in the network at the same time,and reading stock comments manually is time-consuming and labor-consuming.Therefore,how to capture the subject information of stock comments from a large number of expert stock comments quickly and effectively has important theoretical value and practical significance.This thesis mainly carries out the following two works for the subject identification of expert stock evaluation.(1)Based on the LDA topic probability model(Latent Dirichlet Allocation),a topic recognition model of stock evaluation based on word frequency statistics is proposed.It is used for stock price trend prediction.The LDA based on the principle of statistics is a method that takes the statistical results of word frequency as the input and mines the potential semantic information of words.It extracts the distribution of hidden topics as the classification feature through clustering.At present,it is widely used in text topic recognition.In this thesis,the LDA is applied to extract the features of expert stock reviews,then Support Vector Machine(SVM)is used to identify stock reviews.And the results are used to predict the trend of stock price.This method has the following advantages: a)the hidden topic distribution is obtained by LDA through mining the internal semantic information of words.The hidden topic distribution can reduce the feature dimension and reduce the information loss at the same time.Then the effect of stock evaluation recognition is improved.b)The application scope of the LDA is expanded by the research on stock reviews.(2)Combined with the characteristics of expert stock evaluation,the improved BERT model(Bidirectional Encoder Representations from Transformers)is proposed.Then the improved BERT model is applied to the subject recognition of expert stock evaluation.The results are used to accurately predict stock price trend.BERT is a bidirectional coding model,which can capture context structure information comprehensively.It is one of the most advanced methods used in the field of text classification.In this thesis,the BERT model is improved,then the improved BERT model is applied to the stock market.In the input of BERT,the sliding window is used to segment stock comments.And in the output,the multilayer feature ablation method is applied to extract fusion features.The stock comments are recognized by the extracted features and used to predicted the stock price in this thesis.The following improvements are made in this method: a)the original stock comments are intercepted by the sliding window,which helps the BERT model obtain all the information of stock comments.This method can increase the sample size and reduce the over fitting problem.b)The multi-layer feature ablation method in BERT model can extract effective fusion features,then improve the accuracy of subject recognition in stock comments.c)The effect of the fusion features can be further improved by the multi-layer feature ablation method of BERT model on the basis of the context structure information learning in BERT model.The experimental results based on the experimental data show that the two models proposed in this thesis can accurately capture the theme of expert stock comments.The theme of numerous stock comments can guide the prediction of stock price trend.