Business Context

FitLoop, a subscription-based fitness app with 2.4M monthly active users, wants to predict which new users will still be active on day 30. The growth team will use the model to trigger onboarding nudges and retention campaigns for users at risk of dropping off.

Dataset

You are given a user-level training table built from each user’s first 7 days after signup.

• Rows: 420K users from the last 12 months
• Target: retained_day30 = 1 if the user is active on day 30 after signup, else 0
• Class balance: 28% retained, 72% not retained
• Missing data: ~12% missing in campaign metadata, ~6% missing in notification fields, and sparse values for optional profile fields

Success Criteria

A solution is considered good enough if it achieves ROC-AUC >= 0.82, PR-AUC >= 0.58, and recall >= 0.70 at precision >= 0.50 on a held-out test set. The output should also support campaign targeting by ranking users by retention probability.

Constraints

• Predictions run daily in batch for ~80K new users
• Inference should complete within 15 minutes
• Marketing needs feature-level explanations at the segment level
• Retraining should be feasible weekly with standard Python ML tooling

Deliverables

1. Build a binary classification model to predict day-30 retention from first-7-day behavior.
2. Explain preprocessing, feature engineering, and how you handle missing/categorical data.
3. Justify model choice and validation strategy, especially for temporal leakage.
4. Report evaluation metrics and recommend an operating threshold for campaign targeting.
5. Describe how you would deploy, monitor, and retrain the model in production.