Business Context

You’re interviewing for an ML role at TelcoOne, a prepaid + postpaid telecom operator with 18M active subscribers across the US and LATAM. Churn is a major profit lever: the business estimates an average $140 gross margin per retained subscriber over the next 6 months, and the retention team can only contact ~250K customers/week with targeted offers. Leadership wants a model that flags customers likely to churn in the next 30 days, so the retention team can intervene with minimal wasted outreach.

The product constraint is important: sending an offer to a customer who would not churn has a real cost (discount leakage + support load), but missing a true churner is even more expensive. You must design a model that is accurate, operationally feasible, and robust to changing behavior (seasonality, promotions, network incidents).

Dataset

You have a historical training set built from daily event logs aggregated into a customer-week table. Each row represents a subscriber’s behavior over a trailing window ending on a given week.

• Size: ~9.5M customer-weeks spanning 18 months, 53 features
• Target: churn_30d = 1 if the subscriber disconnects or ports out within 30 days after the week end
• Class balance: Imbalanced — about 4.5% positive overall; spikes to 7–9% during competitor promos
• Missingness: 10–20% missing in network-quality features (coverage gaps/roaming); 35% missing in CSAT (no ticket)

Success Criteria (Business + ML)

1. Operational goal: In the top 10% highest-risk customers scored each week, achieve lift ≥ 3.5× versus random selection.
2. Model goal: Achieve AUC-PR ≥ 0.35 (baseline is the churn rate ~0.045) and recall ≥ 0.65 at precision ≥ 0.25.
3. Calibration: Risk scores should be usable as probabilities for budgeting; expected calibration error should be reasonable (you can propose how to measure/monitor it).

Constraints

• Latency: Weekly batch scoring of ~18M customers must finish within 2 hours on a modest cluster (e.g., Spark + Python model, or a single beefy machine if feasible).
• Interpretability: Retention ops requires top drivers per customer (reason codes) for scripting outreach.
• Leakage avoidance: Features must be computed strictly from data available before the prediction week ends.
• Model updates: Retrain at least monthly; be resilient to concept drift (promotions, network incidents).

Deliverables

1. Define the prediction framing (label window, observation window, and how you avoid leakage).
2. Propose a baseline and a production-ready model, including feature engineering for temporal behavior.
3. Describe your train/validation/test strategy (time-based) and how you tune hyperparameters.
4. Specify evaluation metrics (including ranking metrics like lift) and how you pick an operating threshold given outreach capacity.
5. Explain how you would generate per-customer explanations and how you would monitor drift and performance post-deploy.