How to Train a Scoring Model in the Age of Artificial Intelligence

A structured methodology is presented for training, comparing, and selecting a scoring model, particularly within a professional credit risk context. Beyond simple predictive performance, a robust model must be statistically sound, stable over time, interpretable, and aligned with business logic. Logistic regression is positioned as a key reference model due to its advantages in interpretability and stability, even when compared to more complex AI models. The process involves splitting data into training, test, and out-of-time samples and evaluating candidate models against statistical, performance, and stability criteria.