Research On Stock Price Prediction Combined With Financial Text Sentiment Analysis

In recent years,with the development of deep learning,the using of neural network models to predict the future price trend of stocks has become a research hotspot in professional investment institutions and financial and computer science circles.These studies usually predict the future ups and downs of a stock based on its historical price data.In addition to stock price data,major financial websites are active with stock-related reviews,financial news and other text data.Textual information usually contains market investor sentiment,and investor sentiment affects stock price movements.Therefore,in recent years,many studies have integrated text features and price features to build stock prediction models.However,the current stock prediction research fused with text is partial to personalization and privatization,that is,to build private datasets for specific research objects.Moreover most of the researches focus only on the US stock market.Therefore,in order to start research in this field and realize flexible study for any stock,we must implement every link involved in the whole process.We aim to study stock price prediction incorporating text sentiment,to explore how to mine stock-related text,how to extract sentiment representation from text,and finally to apply it to stock price prediction.The whole work of this paper mainly exists in the following two aspects.1.Constituent Parsing Compatible with Entities:The key point of text sentiment analysis based on deep learning model is how to extract effective text representation.We choose tree-based model RvNN among LSTM,attention mechanism weighted summation,convolutional neural network,and tree-structured sentence representation RvNN.RvNN relies on the constituent parsing tree of a sentence,starting from the leaf node,and recursively getting the representation of the root node as the representation of the sentence.In the practice process,we found that current existing constituent parsing algorithms all suffer from entity-violating issue.Due to the rigor requirements of data,we branched out a research point in the main line of stock prediction research:how to reduce the entity violation problem in constituent parsing?In this study,we proposed a constituent parsing algorithm compatible with entity structures based on biaffine mechanism.We clarify the scientificity of our proposed model in theoretical point of view.Later extensive experiments on three constituent parsing datasets,ONTONOTES,PTB,and CTB5.1,prove that the model proposed in this study can effectively reduce the entity violation rate.In addition,in order to demonstrate the practical significance of this study,we conduct our parsing model on downstream tasks,and the experiment results prove that entity compatible parsing model can obtain better performance.2.Stock Price Prediction Combined with Text Sentiment:The paper chooses LONGi shares and CSI300 index as research object.We implement the whole process of stock-related web text crawling,text sentiment analysis,and finally integrating text sentiment into stock price trend forecasting.1.We write a crawler system based on the Scrapy framework to crawl specific text on Oriental Fortune website,Xueqiu website,China Securities News website,and foreign website StockTwits;2.Based on the constituent parsing model realized by research 1,we build a tree-structured text sentiment analysis model relying on the RvNN architecture.The built model is trained on two professional financial sentiment analysis datasets Financial PhraseBank,FIGA-2018 and self-built dataset StockTwits to obtain a financial sentiment model.3.We use the sentiment analysis model of the second step to extract sentiment representation from the text data crawled in the first step,and fuse it with stock price features to predict the future trend of stock price.The final experiments prove that the text sentiment representation extracted by our study can achieve higher stock prediction accuracy than only using stock price data or directly representing the text and then merging with price features.At the same time,it also proves that stock-related text does contain sentiment and help to predict the trend of stock price.