SHARC: SHAP-Based Interpretability in Machine Learning Risk Models for Regulatory Capital under ICAAP and CCAR
创建于 更新于
摘要
SHARC (SHAP for Regulatory Capital) applies SHAP to SVaR outputs from a Hybrid Gaussian Process Regression - Historical Simulation framework to obtain a regulator-friendly decomposition into baseline, mean-driven, and volatility-driven components. This enables transparent linking of scenario inputs (West Asia War, Climate Risk, AI Bubble/Regulatory Burden) to capital outcomes, supporting auditable traceability of capital drivers and guiding regulatory reporting. A key finding is that under stress, the mean return component dominates the SVaR (rather than the variance), implying that directional losses are the primary driver of capital levels and influencing hedging and position-management strategies. The SHARC force plots provide zero-remainder explanations suitable for inclusion in ICAAP/CCAR model documentation, aligning with FRTB and operational resilience expectations. [page::0][page::9]
