Research On Stock Forecasting Method Based On Deep Learning

The ups and downs of stocks are closely related to the development of the national economy and have a great impact on every investor.Therefore,using deep learning technology to predict the financial market has important practical significance.However,there are still some problems in the research at this stage.First,it focuses on modeling the data of a single stock,ignoring the relationship between stocks.Second,traditional research focuses on predicting stock prices at a single time point,such as predicting the future.One-day stock prices cannot be effectively modeled for long-term price trends.This thesis conducts research on these two issues and obtains the following research results:For the existing models that focus on the study of a single stock price sequence to capture the time characteristics of stock historical prices,thus ignoring the correlation between different stocks,this thesis proposes a new deep learning method for stock price prediction.A stock correlation graph is constructed,where individual stocks act as nodes and stocks with highly correlated prices are connected to each other,and our model combines GCN and GRU.Among them,GCN is used to extract spatial features from the price of each stock and the prices of stocks connected to this stock in the graph.GRU is used to extract time features from the historical price of the stock,and finally integrate its time features and space features to obtain the stock price prediction result.Our extensive experiments on real stock price data show that our method outperforms baseline models.Traditional research focuses on the use of short-term stock price series to predict the stock price of a single time node,but cannot effectively predict the long-term price trend.This thesis proposes a stock price prediction method based on Transformer.The Transformer model can model long-term stock prices through its own attention mechanism,but an excessively long lookback window will increase the computational complexity of the model,so this thesis will input the stock price sequence Slicing,which not only reduces computing time complexity,but also better extracts information between different slices.Second,we also maintain the channel independence of the input data.For the information of multiple dimensions of the stock,independently predict the future value of each channel,and then combine to obtain the prediction result of the stock price.Experimental results show that our model outperforms baseline models.