<p>Stock markets are characterised by intricate relational dynamics and continuously evolving market trends. However, existing methods have significant limitations. Firstly, current research struggles to address both the effects of momentum spillover between pairs of stocks and clustered localised systemic risks simultaneously, often treating these interdependent factors in isolation. Furthermore, these models fail to consider the way in which market trends impact individual stock characteristics based on their attributes. To overcome these limitations, we propose MARF (Market-Aware Relational Fusion), an innovative graph neural network framework that incorporates these dual relational dimensions within a market-informed context to improve stock trend prediction. Specifically, MARF incorporates a market-aware temporal embedding mechanism to capture the diverse influences of market conditions on stock-specific features. It utilises hypergraph neural networks and heterogeneous graph neural networks to model clustered and pairwise risks and employs a multi-head attention mechanism to synthesise embeddings dynamically across relationships involving industry, region, firm-executive and firm-shareholder interactions. When evaluated on two real-world datasets from China’s A-share market, MARF outperforms baseline models in terms of both DA and AUC, demonstrating superior predictive accuracy and profitability. This work offers a unified framework for understanding relational risks and market-driven influences, providing a foundation for more robust and advanced financial forecasting models.</p>