ML Benchmarking for Rating Models

A systematic benchmark study compares machine learning methods — including random forests, gradient boosting and neural networks — with classical logistic regression models for the development of rating models in the banking sector.

The study analyses predictive performance, discriminatory power and calibration quality, taking into account regulatory requirements. The results show that ML methods are statistically superior in many scenarios — and thereby provide an early empirical basis for the discussion of AI use in credit risk management, which has grown strongly since then.