Discussion Paper on Machine Learning for IRB Models (OpenAlex (recherche mondiale))

Summary

• The EBA published its discussion paper on machine learning for IRB models under reference EBA/DP/2021/04 on 11 November 2021.p.1
• Consultation responses to the proposals in the discussion paper are requested by 11.02.2022.p.4
• Personal data processing by the EBA adheres to Regulation (EU) 1725/2018 of 23 October 2018.p.4
• The EBA previously issued a report on big data and advanced analytics trends in banking on 13 January 2020.p.7
• Under the CRR, historical observation periods used for PD and LGD parameter estimations must cover a minimum of 5 years.p.15

Key insights

1. Internal ratings-based models used to estimate credit risk capital requirements have remained largely unchanged from approaches introduced 15 to 20 years ago.p.7
2. Machine learning adoption in credit risk is primarily focused on credit decisions and pricing rather than regulatory capital requirements.p.10
3. Institutions frequently apply machine learning techniques as challenger models to benchmark standard regulatory models during validation.p.11
4. High structural complexity in machine learning models makes demonstrating an intuitive economic rationale connecting input variables to outputs challenging.p.14
5. Data retention constraints under GDPR may conflict with CRR rules requiring a historical observation period of at least 5 years.p.15
6. Estimating hyperparameters from data requires splitting datasets into training, validation, and test samples to prevent overfitting.p.17
7. High-frequency model updates present governance and validation risks if minor revisions accumulate into material unapproved model changes.p.18
8. The EBA advises institutions to avoid unnecessary model complexity unless it yields a significant gain in predictive accuracy.p.24