SHAP Stability in Credit Risk Management: A Case Study in Credit Card Default Model

SHAP explanations for credit risk models are not uniformly reliable: The study found that SHAP's consistency in explaining which factors drive credit card default predictions varies depending on the importance of the variable being analyzed, meaning financial institutions cannot equally trust SHAP explanations across all factors in their models.

High-importance variables are more stable than low-importance ones: The research demonstrates that SHAP provides more consistent explanations for the most influential factors in default prediction, but becomes less reliable when explaining less important variables—a critical distinction for regulators and risk managers relying on these tools.

Transparency tools require context-specific validation: Rather than assuming explainability methods work uniformly, this case study shows that financial institutions need to evaluate and validate XAI tools like SHAP within their specific credit models before fully relying on them for regulatory compliance or fairness assessments.

Practical guidance for responsible AI implementation: The authors provide actionable recommendations for credit risk management, suggesting that institutions should be more cautious when using SHAP explanations for lower-importance variables and should implement additional validation steps when these explanations inform high-stakes decisions about credit approval.