The invention provides a stock prediction method which combines news corpus and stock market transaction data. The method makes full use of a large amount of corpus information of a network and breaksthe traditional boundary of a single analysis data source. Through a deep learning model, the stock market news corpus can be analyzed in batches, and the importance of corpus for prediction can be judged automatically, thus the automation and precision of network information analysis are realized. Modeling is performed on news corpus and transaction data through deep learning, and the relationship between different data is comprehensively analyzed according to different information from many aspects. The influence of stock market information on the stock price, the persistence of the stock market information and the investor's psychological factor are grasped, so that the stock market forecast accuracy is further improved; and the word vector, a GRU neural network, an attention mechanismand other in-depth learning cutting-edge technology are used, so that the implementation of science into the industry is realized, and scientific and technological innovation is achieved.