One running question: A payments company hands you two days of card transactions — 284,807 of them, 492 fraudulent. A model that flags nothing is 99.83% accurate. Build the one they should actually switch on at the point of sale, and tell them honestly what it will and won't catch, what it will cost them, and what would make it stop working. This is the course's capstone — no new algorithm family, just Modules 01-04's tools (logistic regression, decision trees, an MLP, confidence intervals, hypothesis tests) applied end-to-end to a messier, higher-stakes problem. Arc 1 makes you commit to a target and a metric before any model exists. Arc 2 hits two leaks Module 03 didn't cover: a feature that already knows the answer (the Epic sepsis story, cashed out on the diabetes data), and a temporal leak you build yourself. Arc 3 picks a model under real constraints, not by "newest." Arc 4 turns Module 03's threshold sweep into a dollar cost matrix, tuned properly. Arc 5 tells the truth: a real population-shift collapse (NSL-KDD), an optional drift-detector tangent, and a bootstrap check on whether the scoreboard has a real winner.
Builds forward from Modules 01-04 — cleaning, statistics, the model toolbox, and the MLP are assumed and reused, not re-taught. New anchor data: the ULB/Worldline Credit Card Fraud dataset (Arcs 1-5, the project) and NSL-KDD (Arc 5's s5-1, and the optional mini-project) — chosen for genuine class imbalance and a real, dollar-denominated cost asymmetry the diabetes/MNIST anchors couldn't supply. The diabetes data returns for one session (s2-1/s3-2) to host a hands-on target-leakage example on data you already own. Telco churn appears only as a never-loaded text brief.
This module vendors two new datasets (the ULB credit-card fraud transactions, ~150 MB, and the NSL-KDD
intrusion files). Re-run env/setup.sh once — it downloads them with checksum checks and
re-runs the smoke test. No new Python packages.
This closes out the data science course's currently-planned module list. See
progress.md — a future Module 06 would be a new /lecture
request, not assumed here.