Machine learning models are making credit decisions, flagging fraud, and driving risk scores inside banks every day. Regulators and internal audit teams increasingly demand that those decisions be explained - not just accurate. This programme works through the methods, tools, and documentation practices that make that possible.
Delivered as live interactive webinars with recorded access, the curriculum is built around real bank data scenarios, not textbook abstractions.
Reserve Your Place
Eight sessions move from foundational interpretability concepts through to deployment-ready documentation. Each module builds on the previous one, so the order matters.
Covers the gap between model performance metrics and model explainability, with examples drawn from Basel III and SR 11-7 model risk guidance. Participants examine real cases where unexplainable outputs led to supervisory findings.
Hands-on session using SHAP on a gradient boosting credit model. Covers waterfall plots, beeswarm charts, and how to translate feature importance into language that satisfies a model validation committee.
Applies LIME to individual loan decisions. Discusses where LIME explanations are stable and where they break down, and how to detect instability before presenting results to a compliance team.
Explores how XGBoost and LightGBM monotonicity constraints align model behaviour with credit policy. Introduces fairness metrics relevant to South African financial services regulation.
Uses PDP and ICE plots to surface how individual variables drive predictions across the score range. Particular attention to detecting spurious interactions that inflate apparent model performance.
Generates counterfactual explanations for declined credit applications. Covers DiCE and manual counterfactual construction, and how to phrase adverse action notices that meet NCRA requirements.
Builds a model card and MRM documentation template using outputs from sessions 2 through 6. Participants leave with a reusable structure for documenting explainability evidence in their own institutions.
Addresses drift in feature importance distributions as a signal of model degradation. Introduces PSI and KS-based monitoring adapted for SHAP value distributions rather than raw predictions.
Credit risk analysts, model validators, data scientists embedded in risk functions, and compliance officers who need to understand what their models are actually doing - and be able to explain it to someone who does not write code.
Prior exposure to Python and basic ML concepts is assumed. Deep statistical background is not required.
Each session pairs concept explanation with a worked example using real-world bank data structures. By the end of the programme, participants have applied every method to a running credit model, not a toy dataset.
Sessions are led by practitioners with direct experience inside South African and international banking institutions - not academics presenting hypothetical scenarios.
Eight years building and validating scorecards at two of South Africa's four major retail banks. Has presented model documentation to the SARB twice during on-site reviews.
Spent six years in model risk management at a pan-African development finance institution. Specialises in translating technical model outputs into audit-ready documentation.
Former quantitative analyst in the risk division of a Johannesburg-based commercial bank. Focuses on the intersection of BCBS 239 data principles and ML model transparency requirements.
Data scientist who has deployed explainability pipelines in production fraud detection and credit origination systems. Leads the hands-on coding portions of sessions 2 through 6.
Answers to what participants typically ask before enrolling. For anything not covered here, reach out directly at support@maxbist.com or call +27 62 181 1245.
