The paper presents HIST (a graph-based framework for stock trend forecasting via mining concept-oriented shared information), which addresses limitations of prior work that assumes stationary stock-concept relations and ignores hidden concepts. HIST constructs a stock-concept bipartite graph to learn dynamic representations of predefined concepts by aggregating stock information with market-cap weights and correcting representations via cosine similarity to all stocks. It then extracts hidden concepts using a residual-based algorithm: after removing predefined concept effects via a doubly residual architecture, it initializes hidden concepts from stock embeddings, connects each stock to its most similar hidden concept, and aggregates information to form hidden concept representations. Finally, individual stock information is extracted by further filtering shared information. The three types of information are combined via a feed-forward network for prediction. Experiments on real-world stock data show that HIST outperforms baselines (e.g., LSTM, GCN, HATS) in forecasting accuracy and investment returns. Ablation studies validate the importance of each component, and visualization of hidden concepts demonstrates their practical relevance, such as capturing pandemic-related stocks. The framework effectively leverages dynamic shared information from both predefined and hidden concepts to improve stock trend forecasting.