FinGraphRisk: A Graph Neural Network-Based Framework for Sentiment-Driven Financial Risk Assessment

Financial risk assessment depends on understanding market sentiment and its influence on financial decision-making. Traditional models struggle to capture the complex relationships between financial entities and sentiment-driven risk factors. This paper presents FinGraphRisk, a graph-based framework that integrates sentiment analysis with financial risk modeling. The approach constructs a heterogeneous financial knowledge graph linking companies, news articles, and market indicators. A graph neural network (GNN) extracts risk-sensitive representations, improving risk prediction accuracy. The model is evaluated on a financial sentiment dataset, achieving an F1-score of 90.3 for positive sentiment classification and outper-forming LSTM- and BERT-based baselines. For risk prediction, FinGraphRisk achieves the lowest mean squared error (MSE) of 0.021 and mean absolute error (MAE) of 0.107, outperforming transformer-based (MSE 0.027) and LSTM-based (MSE 0.045) models. An ablation study confirms that financial entity relationships, sentiment links, and market indicators significantly contribute to predictive performance, with the removal of financial entities increasing MSE to 0.040. The results highlight the importance of graph-based representations in financial modeling. By incorporating dynamic financial interactions, the proposed approach captures sentiment-driven risk patterns more effectively than traditional models.