PSD2 Explainable AI Model for Credit Scoring

Model transparency enables regulatory compliance: The paper demonstrates that explainable AI techniques can be integrated into high-performing credit scoring models, addressing PSD2 requirements for banks to explain automated credit decisions to customers. By using SHAP (Shapley Additive exPlanations) values, the model provides both global and local interpretations of predictions, making individual lending decisions understandable to non-technical stakeholders.

Interpretability doesn't require sacrificing accuracy: The CatBoost model achieved a GINI coefficient of 0.68 (indicating strong predictive performance) while maintaining explainability through feature importance analysis. This challenges the common assumption that there's an inherent trade-off between model accuracy and interpretability in credit risk assessment.

Feature importance analysis reveals decision drivers: The identification of the 20 most important features using Shapley values allows banks and customers to understand which specific attributes (such as income, credit history, or employment status) most strongly influence credit decisions, enabling fairer assessment of lending outcomes.

Explainability supports responsible lending practices: By making the relationship between individual attributes and credit predictions transparent and "human-comprehensive," this approach helps detect potential biases in lending decisions and allows both institutions and applicants to identify whether decisions are based on legitimate risk factors or potentially discriminatory patterns.