Business Context

NimbusCRM, a mid-market SaaS company, wants a lead-scoring model to predict whether a trial account will convert to a paid subscription within 14 days. The sales team needs a model that generalizes well to new campaigns because recent experiments added many correlated behavioral and marketing features.

Dataset

You are given a historical training set of trial accounts collected over 18 months.

• Size: 42K trial accounts, 60 features
• Target: Binary — converted to paid within 14 days (1) vs not converted (0)
• Class balance: 28% positive, 72% negative
• Missing data: 12% missing in firmographic fields, 6% missing in sales interaction fields, and high correlation among campaign and usage variables

Success Criteria

A good solution should improve generalization versus an unregularized baseline, achieve test AUC-ROC >= 0.82, and keep the train-test performance gap under 0.03. The final model should also provide interpretable coefficients or feature importance for go-to-market stakeholders.

Constraints

• Inference must score 200K leads daily in batch
• The sales team prefers interpretable models over black-box ensembles for the first release
• Retraining should be simple enough to run weekly

Deliverables

1. Build a baseline model without regularization and compare it to regularized alternatives.
2. Explain when regularization should be used and what problem it solves in this dataset.
3. Implement a training pipeline with preprocessing, cross-validation, and hyperparameter tuning.
4. Evaluate performance using appropriate classification metrics and calibration checks.
5. Recommend a production-ready model and justify the regularization choice.