Business Context
BrightShield Insurance wants a simple, evidence-based explanation of when to use bagging vs boosting for a tabular risk model. The analytics team currently uses single decision trees to predict whether an auto insurance claim will become high-cost, but performance is unstable across retraining runs.
Dataset
You are given a historical claims dataset for binary classification.
| Feature Group | Count | Examples |
|---|
| Policyholder profile | 6 | age, tenure_months, region, vehicle_age |
| Claim details | 8 | claim_amount_initial, accident_type, repair_estimate, police_report_flag |
| Driving history | 5 | prior_claims_3y, violations_2y, annual_mileage |
| Policy attributes | 4 | coverage_type, deductible, premium_amount, channel |
| Derived operational fields | 3 | days_to_report, claim_to_premium_ratio, weekend_accident_flag |
- Rows: 120,000 claims, 26 features
- Target:
high_cost_claim = 1 if final payout exceeds $15,000, else 0
- Class balance: 18% positive, 82% negative
- Missing data: ~7% missing in repair estimates, ~4% missing in annual mileage, <1% elsewhere
Success Criteria
A strong solution should:
- clearly explain the practical difference between bagging and boosting in simple terms,
- implement and compare at least one bagging model and one boosting model,
- improve over a single decision tree on out-of-sample performance,
- provide a recommendation for production given interpretability and latency constraints.
Constraints
- Batch scoring runs hourly on up to 50,000 claims
- Underwriting stakeholders want a simple explanation of model behavior
- Training must be reproducible and finish within a standard laptop or small VM environment
Deliverables
- Explain bagging vs boosting in plain language for a non-technical stakeholder.
- Train a baseline decision tree, a bagging model, and a boosting model.
- Compare models using cross-validated and test-set metrics.
- Summarize bias-variance tradeoffs and failure modes for each approach.
- Recommend which approach to deploy for this claims dataset and why.
