Research On Stock Price Prediction Based On Text Event Detection And Neural Network

With the rapid development of the economy,stocks have gradually become an important tool for corporate financing and personal finance.Changes in stock prices directly affect the healthy development of the national economy.Stock price fluctuations are affected by various factors such as news events,policy changes,and economic environment.Traditional investment analysis methods have poor non-linear mapping capabilities and cannot meet the needs of stock price prediction;the method based on time series prediction only uses stock volume-price data for prediction,and cannot fully explain the reasons for stock price fluctuations.The efficient market hypothesis shows that the fluctuations of stock price are affected by newly released news.Therefore,in order to improve the accuracy of stock price trend prediction,a stock price prediction model based on financial text events and volume-price data is proposed.The paper mainly makes two research contributions as follows:(1)In order to extract the event type information of the news text,this paper proposes a new Chinese event detection model—a multi-pooling gate convolution neural network(MGCNN)based on mixed character and word embedding.This paper investigates the existing event detection models,and points out that these models have problems such as incorrect word boundary division,wrong event classification problem of the same trigger word in the Chinese text data set.This paper improves the input,convolution,pooling,and loss function of the convolutional neural network to solve these problems.Specifically,the MGCNN model performs mixed character and word embedding on the input text,and uses dynamic multipooling residual gate convolutional neural network to model the text embedding to extract rich semantic and grammatical features.Under the loss function of adding complex sample information entropy regular term,the network is trained and optimized.Finally,the results of event trigger word recognition and classification are obtained.Through the experiments on the public ACE2005 data set and financial news text data set,the effectiveness of the proposed model is proved.Compared with the existing five models,the F1 value of the trigger word recognition and classification has increased by 2.2% and 1.5% respectively,which can meet the requirements of financial news text event detection.(2)In order to improve the accuracy of stock price trend prediction,after obtaining the financial text event type information,this paper proposes a stock price trend prediction model(TPM)driven by financial text and volume-price data.First,based on the basic volume-price data,we establish a regression model and an SDNE model to extract volume-price features such as unpredictable trading volume and stock sector information embedding,and use LSTM to model the extracted features to obtain the stock market information embedding;then based on the dilated gate convolutional neural network and the self-attention layer,the text event sentence and event type are modeled to obtain the event information embedding;finally,the market and event embedding are fused and input into the MLP neural network to get the prediction of the stock price.The experiment on the CSI 300 stock data set proves that the proposed TPM model can effectively improve the prediction performance by fusing financial text and volume-price data.Compared with other prediction models,our TPM model obtains higher prediction accuracy and profitability.