Course 3Predictive Modeling & Modern Actuarial Machine Learning
Master Generalized Linear Models (GLMs), unsupervised risk profile discovery, tree-based ensemble methods, and regularization techniques using statsmodels and scikit-learn.
For: Students fluent in data preparation who want to build production-grade pricing, reserving, and risk classification models.
-
Lesson 1
Statistical Foundations of Predictive Modeling
OLS review, why it breaks on insurance data, log links, robust standard errors. Lab: health-claim severity by service category.
-
Lesson 2
GLMs for Claim Frequency
The exponential family, Poisson regression with exposure offsets, overdispersion and negative binomial. Lab: a motor frequency model.
-
Lesson 3
GLMs for Severity & Pure Premium
Gamma and inverse Gaussian severity, frequency × severity, Tweedie. Lab: a prospective rating algorithm.
-
Lesson 4
Overfitting, Validation & Regularization
Bias–variance, cross-validation, Lasso, Ridge, ElasticNet, scaling. Lab: a 60-feature health cost model.
-
Lesson 5
Unsupervised Learning & Risk Group Discovery
Member profiles, K-means, elbow and silhouette, z-score heatmaps and radar charts. Lab: three patient archetypes.
-
Lesson 6
Rule Extraction & Interpretability
Surrogate decision trees, depth and pruning, export_text. Lab: underwriting rules and risk tiers.
-
Lesson 7
Trees, Forests & Gradient Boosting
CART, random forests, XGBoost with a Poisson objective, SHAP. Lab: boosting vs the GLM, explained.
-
Lesson 8 · Part A
Production Pipelines & Model Comparison
scikit-learn Pipelines, penalized GLMs, joblib persistence, and an A/B test on holdout loss ratios.
-
Lesson 8 · Part B
Capstone: A Production Pricing Engine
One command from raw policy and claim records to fitted models, rate indications, and a management report.