Business Context

NorthStar Mobile, a regional telecom provider with 1.2M subscribers, wants to predict which customers will churn in the next 30 days so the retention team can target save offers efficiently. The team needs a model that is fast, explainable, and easy to retrain monthly.

Dataset

You are given a customer-level training dataset built from the last 18 months of account, billing, usage, and support history.

• Rows: 240K customers, one row per active customer snapshot
• Target: churn_30d = 1 if the customer cancels within 30 days, else 0
• Class balance: 11.5% churn, 88.5% non-churn
• Missing data: ~8% missing in usage fields for new customers, ~12% missing in device-related fields, sparse missing categorical values in region and payment method

Success Criteria

A good solution should achieve strong ranking performance and actionable recall for the retention team: ROC-AUC >= 0.82, PR-AUC >= 0.42, and recall >= 0.70 at precision >= 0.35 on a held-out test set. The model should also provide interpretable coefficients or feature effects.

Constraints

• Prefer interpretable models over black-box ensembles
• Nightly batch scoring must finish in under 10 minutes for 1.2M customers
• Retraining should be simple enough for a small ML platform team to maintain
• Avoid leakage from post-churn or future information

Deliverables

1. Build a regression-based churn prediction pipeline using logistic regression.
2. Explain feature engineering, preprocessing, and how you handle missing data and class imbalance.
3. Evaluate the model with appropriate classification metrics and threshold selection.
4. Interpret the model outputs for business stakeholders.
5. Describe how you would deploy, monitor, and retrain the model in production.