What does EBA discussion paper on machine learning for IRB models say about AI in banking?

What are the key points of EBA discussion paper on machine learning for IRB models?

• Published November 11, 2021; consultation closed February 11, 2022.
• Aim: set supervisory expectations on how ML models can adhere to CRR requirements for IRB models (Articles 142–191 CRR) — not to change the rules.
• Identifies challenges: interpretability and explainability, model complexity and overfitting, data quality and representativeness, staff skills, and the CRR's requirements on human judgement and documentation.
• Identifies opportunities: better risk differentiation, use of unstructured data, ML as challenger models in validation, collateral valuation.
• Principle-based recommendations: ensure adequate understanding of the model by management and validation, document the model and data, guard against overfitting, keep a consistent model over time, and be able to explain individual outputs where needed.
• Applies equally to PD, LGD and EAD models and to ML used in supporting steps such as data preparation and risk differentiation.

What did EBA discussion paper on machine learning for IRB models change for banks?

Before this paper, EU banks had no supervisory statement on whether ML could sit inside capital models; most assumed the ECB would not approve it. The paper confirmed ML is permissible within the IRB framework if it meets existing CRR requirements, and it defined the tests — explainability, understanding, documentation, stability — that the ECB now applies in model approvals.