Enhancing Credit Scoring Models with Explainable AI Techniques

Abstract

The financial services industry has experienced a paradigm shift from traditional statistical credit scoring methods toward sophisticated machine learning algorithms, offering superior predictive accuracy but raising critical concerns regarding model interpretability and regulatory compliance. This research investigates the integration of Explainable Artificial Intelligence (XAI) techniques with ensemble learning methods to address the transparency challenges inherent in advanced credit risk assessment systems. We systematically evaluate the performance of gradient boosting models enhanced with SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) frameworks, comparing them against traditional logistic regression baselines. Our empirical analysis demonstrates that XGBoost models achieve Area Under the Receiver Operating Characteristic curve values of 0.89, substantially exceeding logistic regression performance of 0.78, while SHAP-based feature importance analysis consistently identifies loan amount, checking account status, and borrower age as primary default predictors. The feature attribution analysis reveals that these top three factors collectively account for approximately thirty-five percent of model discriminative power, with loan amount demonstrating the highest individual importance at twelve percent. This research contributes empirical evidence that explainable machine learning frameworks successfully reconcile the competing objectives of predictive accuracy and model transparency, enabling financial institutions to deploy sophisticated algorithms while maintaining regulatory compliance and stakeholder trust.