Artificial Intelligence, Machine Learning and Business Analytics in Banking
The exam tests model lifecycle governance, validation statistics, explainability and RBI's expectations for responsible AI - not how neural networks compute gradients.
Core rules
- Model risk management: maintain a model inventory, tier models by materiality, keep developer and validator separate, validate independently before deployment, revalidate at least annually or on material change, and monitor with PSI, KS and Gini.
- Every model has three performances that fail independently: discrimination (AUC/Gini), calibration (predicted PD versus observed default rate) and stability (score and population drift).
- RBI's FREE-AI framework (2025) sets out principles - trust, fairness, accountability, transparency, human oversight - and expects a Board-approved AI policy with an AI governance committee and third-party AI risk assessment.
- Explainability is a compliance requirement: rejections need adverse action reason codes, and proxy variables such as pin code or device type can create disparate impact even when protected attributes are excluded.
- Analytics spans underwriting, early warning signals, AML transaction monitoring, collections propensity, cross-sell and Ind AS 109 expected credit loss.
Worked example
A bank develops a probability-of-default model with in-sample AUC 0.78. Out-of-time validation on the following 12 months gives AUC 0.69 and an observed default rate 1.4 times the predicted rate. Diagnosis: overfitting plus macro drift, not a data error. Action: recalibrate the intercept to the out-of-time default rate, retain the rank ordering, keep the model live only if Gini stays above the approved threshold, add monthly PSI monitoring and a documented macro-overlay policy. Discrimination did not collapse, so the fix is recalibration rather than rebuilding.
Traps that cost marks
Quoting 99% accuracy on a 1% default portfolio where the model predicts nobody defaults. Using AUC as proof of calibration. Assuming vendor-supplied AI escapes model risk or outsourcing rules. Treating chatbot product advice as exempt from misselling liability. Ignoring reject inference and sample bias in scorecards.