Advanced Risk Scoring: Gradient Boosting, Fairness, and Production
Machine learning methods applied to credit scoring have shifted from theoretical research to standard practice. The models in this program reflect that shift — built from data patterns, validated on real portfolios, and interpreted with tools that regulators and risk teams actually use.
about this program
Building a model that scores well on a holdout set is one problem. Keeping it reliable in production, compliant with fair lending rules, and maintainable by a team of three is a different problem entirely.
The production gap
Many data scientists can train a LightGBM model. Far fewer have dealt with score instability after a macroeconomic shift, or had to explain a model rejection to a regulator. Both scenarios are covered here with documented case walkthroughs.
Fairness and disparate impact
Canadian and US fair lending frameworks require that automated decisions do not produce disparate impact on protected groups — even unintentionally. This course covers statistical fairness metrics, disparate impact testing, and how to adjust feature selection and thresholds without gutting predictive power.
Fairness constraints and model performance are not always in opposition. The course demonstrates cases where fairer models also generalize better out of sample.
MLOps for scoring pipelines
You will set up a basic MLflow tracking server, log model versions, and build a PSI-based monitoring script that flags score distribution shifts. The tooling is kept minimal and reproducible — no cloud vendor lock-in.
program curriculum
structured by topic period and learning outcome
-
Gradient Boosting Deep Dive
Week 1LightGBM and XGBoost internals, monotonicity constraints for regulatory compliance, hyperparameter search with Optuna.
-
Fairness Auditing
Week 2Disparate impact ratio, equal opportunity difference, testing feature proxies, fairness-aware feature selection.
-
Explainability at Scale
Week 3SHAP waterfall plots, global feature importance, building automated reason-code generators.
-
Monitoring and Drift
Week 4PSI and CSI metrics, setting alert thresholds, scheduling retraining pipelines with Prefect.
-
MLflow and Model Registry
Week 5Experiment tracking, model versioning, staging and production promotion workflow.
ready to work with REAL scoring data?
The program runs on actual credit portfolio datasets under controlled conditions. You leave with code, documentation, and a validated model — not slides. Enrolment closes when seats fill.
8 seats remaining