You already know: classification and calibration (D3); experiments and causal thinking (D2, D10); feature stores and point-in-time correctness (D11); the fraud and spam cases in this bank. This case adds churn: predicting who will cancel a subscription, and, more importantly, deciding who to act on. The design questions: how to define a label that is not leaking the answer, why ranking by risk wastes money, how to spend a retention budget, and how to measure the effect honestly.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "build a model that predicts churn in the next 30 days, send the top decile a discount". A staff candidate asks what action follows the prediction and whether that action changes the outcome. Churn prediction is only useful through an intervention, and interventions cost money, help some people, do nothing for others and occasionally make things worse. The real system is: a clean label, a model of uplift (the effect of treating someone), a budget-aware allocation, and a randomised holdout that measures what actually changed.
Assumptions you state: