Machine Learning for IRBA Rating Systems
Analysis of the EBA follow-up report on the application of machine learning in IRBA credit risk modelling – opportunities, challenges, and regulatory requirements.
Authors: Manfred Puckhaber, Prof. Dr. Dirk Schieborn
Publication: banking.vision – Blog by msg for banking
Year of publication: 2023
Abstract
The article analyses the follow-up report of the European Banking Authority (EBA) on the application of machine learning methods in IRBA credit risk modelling. The central question is how banks can deploy ML methods – in particular for PD estimation – without violating regulatory requirements.
Key Topics
Methodological strengths and limitations: ML methods can achieve better discriminatory power than classical rating models, but require more elaborate validation and documentation processes.
Interpretability: An appropriate balance between model performance and the explainability of results is a mandatory prerequisite for regulatory approval. The SHAP approach is highlighted as a key tool for improving model interpretability.
Regulatory framework: The article examines the interactions with CRR/CRD and the EU AI Act, which may classify IRBA systems as high-risk AI systems.
EBA recommendations: Comprehensive model documentation, requirements for team expertise, rigorous out-of-sample and out-of-time testing, and SHAP-supported explainability.