Interpretable credit scoring based on an additive extreme gradient boosting

Credit scoring is a critical component of the financial ecosystem, driving lending decisions, risk assessment, and resource allocation. However, the inherent nonlinear and chaotic nature of creditworthiness poses a significant challenge for traditional credit scoring algorithms. Existing methods often struggle to balance predictive performance with interpretability, leading to a lack of transparency in the decision-making process and hindering their practical application. To address this issue, we propose Add-XGBoost, a novel additive ensemble model leveraging the strengths of Generalized Additive Models (GAM) and Extreme Gradient Boosting (XGBoost). Add-XGBoost enhances interpretability by employing an additive architecture wherein individual decision trees are trained for each feature, facilitating granular feature importance analysis and capturing nonlinear feature effects. Furthermore, Add-XGBoost incorporates second-order feature interactions within the additive framework, enabling a nuanced understanding of complex relationships within credit data. To further enhance robustness, a Lasso regression-based optimization method refines the weights assigned to individual feature functions (shape functions). Empirical evaluation across four diverse credit datasets demonstrates Add-XGBoost’s superior performance compared to existing statistical and machine learning-based credit scoring methods, including ensemble approaches. Crucially, Add-XGBoost excels in providing both global and local interpretability, offering valuable insights into feature interactions and their influence on credit risk. This makes Add-XGBoost a transformative tool, bridging the gap between high-performance credit scoring and transparent, comprehensible decision-making in the financial sector.