Curriculum / Module 06
core2 hours3 lessonsTimed assessment

Model Risk and Explainability

Interrogating the black box

The discipline of not trusting a model you cannot interrogate. It introduces model-risk management in the lineage of supervisory guidance, the main explainability methods and their genuine limits, and the difference between an explanation that satisfies a regulator and one that satisfies an engineer — framed for a South African institution and its Prudential Authority expectations. It closes the Core certificate by giving every prior module its governance spine: the verification routine, the monitoring stack and the fairness audit all assume a professional who can read, stress and challenge a model's documentation. This module builds that professional.

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
  • The human-oversight posture the EU AI Act requires

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 genuine 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 own the framework: the model inventory, the validation cycle, effective challenge, and the documentation standard a supervisor samples. This module gives you the vocabulary and the checklist.

Credit & Underwriting

You apply the framework to the scoring models of Module 5 — and gain the explainability literacy to turn feature contributions into the lawful adverse-action reasons that module demanded.

Fraud & AML

You apply it to the detection models of Module 4: drift monitoring is why your typology feedback matters, and explanation outputs are what turn a score into an investigable alert.

Wealth & Advisory

You learn what a defensible model explanation sounds like to a client — and what questions 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 load-bearing 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