Case Studies

Case A: Hospital Readmission Score

The historical label reflects whether a patient returned to the same network, not whether they needed care. Discuss label validity, missing external visits, intervention effects, subgroup harm, and why predictive performance alone cannot authorize deployment.

Case B: Marketplace Fraud Queue

Review capacity is constrained and fraud value is heavy-tailed. Compare recall, expected prevented loss, customer friction, calibration, and queue latency. Decide what should be automated, reviewed, or declined.

Case C: Demand Forecast

A promotion changes demand and the model’s own forecast changes inventory. Identify temporal leakage, feedback, baseline seasonality, prediction intervals, and retraining triggers.

For each case submit a decision contract, credible baseline, split design, failure taxonomy, and go/limit/stop recommendation.