Curriculum / Module 06
core2 hours3 lessonsTimed assessment

Model Risk and Explainability

Interrogating the black box

This module is about working responsibly with a model whose internal reasoning you cannot inspect directly. It introduces model-risk management as it developed from international supervisory guidance, the main explainability methods and their limits, and the difference between an explanation that satisfies a regulator and one that satisfies an engineer, all framed for a South African institution and the expectations of the Prudential Authority. It closes the Core certificate by setting out the governance framework the earlier modules assume. The verification routine, the monitoring stack and the fairness audit each depend on professionals who can read a model's documentation, test its claims and question its conclusions, and this module develops those skills.

Create an accountBack to curriculum

By the end of this module, you can
  • Explain what model risk is and set out a validation and governance lifecycle for a deployed model.
  • Use and critique post-hoc explanation methods, and know when an inherently interpretable model is the better choice.
  • Judge whether a model's explanation is adequate for the decision it supports and the duty attached to it.
Skills taught
  • Model-risk framing
  • Reading and stress-testing model documentation
  • Interpreting explainability outputs and their failure modes
  • Applying the human-oversight standard set out in the EU AI Act

Lessons in this module

L1
Model risk and the lifecycle: the SR 11-7 lineage
Define model risk, set out the validation and governance lifecycle of a deployed model, and explain 'effective challenge' as the organising principle of supervisory expectation.
30 min
L2
The explainability toolbox and its limits
Describe the main explanation methods (feature importance, LIME, SHAP, counterfactuals), interpret their outputs correctly, and state the failure modes that make Rudin's case for inherently interpretable models in high-stakes decisions.
30 min
L3
Explanation adequacy: the engineer, the regulator, and the client
Judge whether a model's explanation is adequate for the decision it supports and the duty attached to it, distinguishing the three audiences an explanation must serve.
30 min
Case 06.1
Case file: The explanation that explained nothing
An applied fact pattern worked against a model resolution, followed by the timed assessment (30 minutes, pass mark 70 percent).
25 min

How it lands across the four desks

Compliance & Risk

You are responsible for the framework itself: the model inventory, the validation cycle, effective challenge, and the documentation standard a supervisor will sample. This module supplies the vocabulary and a working checklist.

Credit & Underwriting

You apply the framework to the scoring models of Module 5, and you learn to read explainability outputs well enough to turn feature contributions into the lawful adverse-action reasons that module required.

Fraud & AML

You apply it to the detection models of Module 4. Drift monitoring is the reason your typology feedback matters, and explanation outputs are what turn a score into an alert an investigator can act on.

Wealth & Advisory

You learn what a defensible model explanation sounds like to a client, and what to ask before your practice relies on any scoring or recommendation engine.

Key literature · 6 sources

Every module rests on a verified scholarly and institutional evidence base. The full core and further reading lists open with the module.

  • Board of Governors of the Federal Reserve System / OCC (2011) 'SR 11-7: Supervisory Guidance on Model Risk Management': the founding framework: the two-limb definition, the lifecycle, effective challenge.
  • Ribeiro, M. T., Singh, S. & Guestrin, C. (2016) '"Why Should I Trust You?" Explaining the Predictions of Any Classifier (LIME).' KDD 2016: local model-agnostic explanation and its mechanism.
  • Lundberg, S. M. & Lee, S.-I. (2017) 'A Unified Approach to Interpreting Model Predictions (SHAP).' NeurIPS 2017: the Shapley-value attribution framework that dominates practice.
  • Rudin, C. (2019) 'Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.' Nature Machine Intelligence 1(5): the principal counter-position and this module's decision rule.
  • Barredo Arrieta, A. et al. (2020) 'Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI.' Information Fusion 58: the field survey framing the toolbox and its trade-offs.
  • SARB Prudential Authority & FSCA (2025) 'Artificial Intelligence in the South African Financial Sector': the explainability-methods survey data, the governance-gap findings, and the regulatory-constraint rankings cited in this module.
Create an account