Abstract: Traditional stock price prediction methods are mostly based on the time series of a single stock， ignoring the complex interrelationships between stocks. In response to this issue， the article proposes a stock price prediction method based on business content to construct stock correlation relationships from the perspective of building a more effective stock portfolio. The model consists of three components: the association relationship construction component， the temporal feature extraction component and the association capture component. The association relationship construction component uses improved TF-IDF to extract the similarity of business content keywords in the annual reports of listed companies to construct stock correlation relationships. The temporal feature extraction component uses LSTM to extract temporal features of stock trading data. The association capture component utilizes GCN to capture high-dimensional features of stock interactions， and finally outputs the predicted stock price through the fully connected layer. The experimental results in the Chinese A-share market indicate that this model has the smallest error， the better fit， and can more effectively predict stock prices compared to single stocks and industry relationship based prediction methods. It is a stock price prediction model that captures the mutual influence between stocks more fully.

Key words: stock price forecast, business content, stock related relationships, term frequency-inverse document frequency, long and short term memory neural network, graph convolution neural network