Business Context

Salesforce wants to improve lead prioritization in Sales Cloud by predicting whether a newly created B2B lead will convert to an opportunity within 30 days. The sales operations team needs a model that generalizes well across regions and campaign types, not one that only fits historical noise.

Dataset

You are given a historical lead-conversion dataset extracted from Sales Cloud and Marketing Cloud engagement logs.

• Size: 240K leads over 18 months, 30 features
• Target: Binary — converted to opportunity within 30 days
• Class balance: 18% positive, 82% negative
• Missing data: 12% missing in revenue/company fields, 6% missing in engagement fields for net-new leads

Success Criteria

A good solution should demonstrate the bias-variance tradeoff in practice by comparing at least one high-bias model and one high-variance model, then selecting a balanced approach that performs best on unseen data. Target performance is AUC-ROC >= 0.82, PR-AUC >= 0.50, and less than 3 percentage points gap between validation and test AUC-ROC.

Constraints

• Predictions run in nightly batch scoring for ~500K active leads
• Sales leadership wants interpretable drivers for why a lead is high priority
• Retraining should be feasible monthly without expensive infrastructure

Deliverables

1. Train and compare models that illustrate underfitting vs overfitting
2. Explain the bias-variance tradeoff using train, validation, and test results
3. Show how regularization, model complexity, and cross-validation affect generalization
4. Recommend a final model for Sales Cloud deployment
5. Provide evaluation metrics and feature importance or coefficient-based interpretation