Machine learning can improve segmentation, default prediction and early-warning signals, but a credit-capital model has to remain controlled, explainable and valid through the cycle. The EBA's discussion of machine learning in IRB models is a useful reminder that predictive performance does not replace prudential evidence.
Separate use cases
Distinguish regulatory capital estimation, underwriting, collections, monitoring and management information. A feature or model suitable for early warning may not be appropriate for a regulatory risk parameter. The approved purpose must be explicit.
Validate the economics and the statistics
Test discriminatory power, calibration, stability, representativeness, default definition, downturn behaviour, missing data and treatment of overrides. Challenge whether a high-performing feature captures a temporary relationship or a structural credit driver.
Document conservatism and change
Where uncertainty cannot be eliminated, quantify the effect and apply a governed conservatism approach. Record the model-development sample, version, monitoring thresholds, validation findings and approval. The EBA IRB machine-learning paper is a useful starting point for a capital-model discussion.