| Stock trend prediction has always been a common concern of both academia and industry.The fluctuation of stock prices is influenced by various factors,such as historical stock market data,news,and related stock fluctuations.At the same time,with the continuous development of machine learning and deep learning technologies,these technologies can be used to process large amounts of data,even multi-source and heterogeneous data.Therefore,this article conducts stock prediction research based on deep learning technology from the above three aspects,and validates the effectiveness of the model on multiple real historical data sets.The work and achievements of this paper are as follows:(1)Future trend prediction based on historical stock market data.Since the change in stock prices is usually somewhat continuous with their historical prices,many researchers use historical stock market data to predict future stock prices.Most existing studies use a single neural network model(such as convolutional neural network CNN,recurrent neural network RNN)to predict the trend of stock prices.However,this study proposes an improved DCT-CNN-LSTM-Attention stock prediction model,which first uses the discrete cosine transform to remove high-frequency signals in the original historical stock trading data,then introduces CNN to extract useful features from the stock data,and uses LSTM to learn the time dependence in the stock data.Finally,the attention mechanism is used to calculate the influence of feature states of data at different time points on stock prediction results,in order to effectively extract important feature information and filter out worthless feature information in the stock market.The results show that the proposed model improves the accuracy of stock trend prediction compared to a single neural network model.(2)A stock trend prediction model based on news text information.Some key news events(such as policy changes,natural disasters,company performance,etc.)may disrupt the coherence of stock trends.Therefore,many researchers view news text data as useful supplementary information for stock prediction.This study also adopts this approach and proposes a deep learning model based on FinBERT and attention mechanism to extract stock news features and improve the accuracy of stock prediction on the basis of the first model.Firstly,the feature vectors of each news headline are obtained through the FinBERT model.Then,an attention mechanism is introduced to assign different weights to different news items within a day,and the comprehensive vector feature of each day is obtained.Meanwhile,the CNN model-derived stock data features are integrated and input into the LSTM-Attention network for trend prediction.The experiment shows that compared with the stock prediction model using a single historical trading data,the inclusion of news text data can more comprehensively reflect market information,improving the accuracy and reliability of prediction.(3)Stock trend prediction based on the correlation between companies.Due to the spatial correlation of stock price trends caused by geographical location,industry,and technology,this study proposes a graph convolutional network-based stock prediction model based on the previous model and introduces the network relationship between different stock individuals.This model first constructs a graph of the relationships between companies,and then integrates the stock trading data features and news headline features obtained from the previous model.It then uses graph convolutional neural networks to learn node features as well as the relationship information between nodes,and finally inputs them into the LSTM-Attention network for trend prediction.The experiment shows that this model can more comprehensively consider market information and company interrelatedness,improving the accuracy of stock price prediction. |
