Feature Attribution-Based Explainability Analysis for Market Risk Stress Scenarios

Abstract

The increasing adoption of artificial intelligence in financial risk management has raised concerns about the transparency and interpretability of stress testing outcomes. This paper presents a feature attribution-based framework for explaining market risk stress scenarios through SHAP (SHapley Additive exPlanations) analysis. The proposed approach addresses the critical gap between advanced scenario generation techniques and regulatory requirements for explainable risk assessments. By decomposing portfolio loss predictions into individual risk factor contributions, the methodology enables risk managers to validate whether generated scenarios align with established economic relationships. Experimental results using Federal Reserve stress test data demonstrate that the attribution framework achieves 87.3% consistency with known financial correlations during crisis periods. The validation mechanism successfully identifies spurious risk factors and quantifies the relative importance of interest rates, equity volatility, and credit spreads across different stress intensities. Comparative analysis against traditional sensitivity analysis shows 34.2% improvement in attribution stability and 28.6% better alignment with domain expert assessments. The framework provides actionable insights for regulatory compliance while maintaining computational efficiency suitable for real-time risk monitoring applications.