Achieving Nonlinearity and Interpretability in Credit Scoring: The Penalised Tree Ensemble for Region Merging (PERM) Model

Balancing predictive performance with interpretability is a critical issue in the application of machine learning (ML) to credit scoring. In this research, we propose a novel intrinsically interpretable classification model called Penalised Ensemble for Region Merging (PERM) that effectively captures potential nonlinearities in data. While existing methods in literature are either explainable without consideration of nonlinear structures (e.g., Logistic Regression) or nonlinear without explanations (e.g., Random Forest and Gradient Boosting), PERM is both intrinsically interpretable and capable of capturing potential nonlinearity.

The PERM model adopts an ensemble structure that converts a series of weak learners into a strong learner, enhancing overall performance. A recently developed regularization term called SCOPE, which originally addressed nominal data, is utilized to overcome the challenge of interpreting a high number of weak learners. This regularization term promotes similar coefficients to have the same value, which merges tree-based rules or discretization regions in the weak learners, thus improving interpretability. The PERM modeling framework is also flexible for extension to meet legislative requirements, further increasing its resilience.

Empirical analyses were conducted on credit default datasets and compared to popular credit scoring models. Results demonstrate that PERM predicts credit risk competitively and consistently while providing valuable insights for decision-makers.