Research On Stock Market Prediction Based On News Events And Knowledge Grap

The movements of the stock price are related to the investor’s profit return.Therefore,the prediction analysis of the stock market has attracted widespread attention from both industry and academia.Since the fluctuations of stock price are subject to various economic and political factors along with the investment psychology and trading technology of investors,and features of stock price are highly non-linear and non-parametric,the analysis and prediction of the stock price movements is a challenging task.Traditional statistical methods or machine learning methods rely on handcrafted features.However,the design of features is subjective,and it is difficult to fully capture some implicit features in the data that can affect the stock market,which often results in poor prediction results.In recent years,deep learning methods have leveraged their unique sequential modeling and fitting capabilities to predict stock movements instead of traditional time series models,which have achieved improved results.Nevertheless,there are still many shortcomings in the existing researches,and the use of data such as news events and relationships between different companies is not sufficient.News texts often contain economic,political and other relevant factors that may affect stock prices,and the relationships between different companies also have a certain impact on the price fluctuations of stocks.Therefore,how to utilize multi-source data such as news text data and the relationships between different companies to improve stock price prediction has become a research hotspot attracting widespread attention at home and abroad.With the latest achievements of natural language processing and graph neural networks,the article conducts systematic researches on stock price and trend forecasting issues based on multi-source data such as stock trading data,news text,and company relationships.The advantages and disadvantages of different methods for stock price forecasting are discussed and analyzed in depth.The stock price forecasting method based on news events and knowledge graph of company relationships are also studied.The main work and innovations of this article are as follows:(1)Although there are several researches on applying deep learning methods in stock index prediction currently,there is still lack of in-depth comparative researches covering different deep learning methods.To tackle this issue,this article selects mainstream neural network models(LSTM,GRU,CNN,MLP)to make stock index predictions on 29 stock markets including the Chinese mainland market,the US market,the European market,and the Asia-Pacific market.The price prediction are divided into short-term predictions in 20 trading days,medium-term predictions in 60 trading days and long-term predictions in 250 trading days.The percentage absolute error mean(MAPE)and absolute error standard deviation(SDAPE)are used to measure the prediction accuracy and stability of the model,respectively.Experiments demonstrate that the GRU model outperforms others in the stock prediction,which has the best overall performance and the highest generalization ability.To further improve the performance and overcome the prediction drawbacks of single model,the article proposes optimization algorithms for multiple models to make joint decisions.Experiments demonstrate that the joint-decision models outperform the GRU model,and improve the prediction performance for stock index prediction.(2)Aiming at the problems that current stock prediction models for news event exact semantic extraction ineffectively and lose key information probably,this article proposes a hierarchical attention stock trend prediction model based on local constraint Transformer.Compared to traditional Transformer,this model pays more attention to local important information of stock news.Leveraging a series of modules such as position coding module,self-attention module with local constraints,multi-head attention module and position-sensitive feedforward network,the model can effectively extract the semantic information of stock-related news texts.Experimental results show that the model performs better in stock trend prediction tasks than the existing methods.In addition,this article introduces a hierarchical attention mechanism with local constraints,where local-constraint attention focuses on the word-level features of news events and news-level self-attention focuses on news-level features.Finally,features are sent to the stock forecasting layer for trend prediction.The importance of different news information can be learned from different attention weights,thereby increasing the interpretability of the model.(3)In news text data,most simple samples have discriminative features,which are relatively easy to be captured,identified,and processed.However,there are still a few difficult samples,which are close to the classification boundary and whose features are difficult to identify.Ifthe simple samples and the difficult samples are not distinguished in the model training process,the model will be dominated by numerous simple samples,and the difficult samples will be ignored.To address this problem,this article proposes a difficult–case-balanced loss function in stock prediction based on news events,which is used to balance the loss contribution values of difficult training samples so that the difficult samples can play a role in model training The experimental results show that the performance of model prediction can be improved by adding the difficult-case-balanced loss function.(4)In current stock price prediction studies,each stock is considered as an independent individual and the relationships between companies are ignored.However,relationships often have a certain impact on the stock price.Therefore,this article proposes a stock prediction method based on the temporal similarity-constraint graph network.First,the knowledge graph is introduced into the stock prediction,as a result the relationships between stocks are represented and modeled,and a knowledge graph based on the association relationships of companines is established.Then,feature extraction of stock relationships is performed through a graph convolutional network whose weights change along with time-series feature to fully capture the linkage between stock prices.Finally,the relationship features are used for stock prediction analysis.Experimental results show that the model can obtain a higher cumulative return on investment compared with the previous methods.This article studies the stock price forecasting method based on news events and knowledge graphs between companies.The methods proposed in this article improve the accuracy of stock price and trend predictions,which can be used to improve existing stock trading systems and help investors make better investment decisions.