AI Act compliance in credit scoring: reconciling interpretability and accuracy via quadratic terms and Powershap

The increasing regulatory demands, notably the EU’s GDPR and upcoming AI Act, underscore the critical need for credit‑scoring models to balance high predictive accuracy with essential interpretability. While traditional logistic regression has been widely used in credit risk assessment due to its interpretability, its predictive performance often lags behind machine learning methods like XGBoost. We propose a systematic approach that integrates quadratic features and powershap feature selection to enhance both model transparency and predictive performance. Empirical analysis demonstrates that logistic regression optimized through our procedure can achieve predictive performance comparable to XGBoost while maintaining strong interpretability. This approach meets regulatory compliance and practical financial requirements, offering a trustworthy and effective credit‑scoring framework.