Robust Cross-Validation of Predictive Models Used in Credit Default Risk

• DOI: 10.3390/app15105495
• Publication Date: 2025-05-14T14:27:41Z
• Authors: Jose Vicente Alonso; Lorenzo Escot

Model validation is a challenging Machine Learning task, usually more difficult for consumer credit default models because of the availability small datasets, modeling low-frequency events (imbalanced data), and bias in explanatory variables induced by train/test sets split techniques (covariate shift). While many methodologies have been developed, cross-validation perhaps most widely accepted, often being part model development process optimizing hyperparameters predictive algorithms. This experimental research focuses on evaluating existing robust variants to address issues validating models. In addition, some improvements those methods are proposed compared with wide range techniques, including fuzzy methods. To reach solid practical conclusions, this work limits its scope logistic regression, as it best-practice technique real-world applications context. It shown that algorithms lead stable estimates, expected due homogeneous partitions, which positive impact selection enhancements improved results when there data restrictions.