Business Context

PulseBoard, a mid-market SaaS analytics company, wants a 30-day churn prediction model for 120K customer accounts. The retention team needs a model that is accurate enough to prioritize outreach but also explainable enough to justify interventions.

Dataset

You are given an account-level monthly training set built from product usage, billing, support, and account metadata.

• Size: 120K account-month rows, 42 features, 18 months of history
• Target: Binary label indicating whether the account churns in the next 30 days
• Class balance: 11.5% churn, 88.5% retained
• Missing data: 12% missing in CSAT, 7% missing in support features, <2% elsewhere

Success Criteria

A good solution should demonstrate whether a regularized logistic regression baseline is sufficient or whether a more complex model such as gradient boosting materially improves business outcomes. “Good enough” means achieving ROC-AUC >= 0.82, PR-AUC >= 0.42, and providing a clear recommendation based on both model quality and operational constraints.

Constraints

• Predictions run nightly in batch for 120K accounts
• Retention managers require interpretable drivers for each prediction
• Retraining should be feasible monthly with standard Python tooling
• Avoid unnecessary model complexity if gains are marginal

Deliverables

1. Build a strong logistic regression baseline with appropriate preprocessing.
2. Train at least one more complex model and compare it fairly.
3. Define the evaluation framework and thresholding strategy.
4. Explain when logistic regression should be preferred over the complex model.
5. Provide production-ready training and evaluation code.