Business Context
You’re on the Risk Modeling team at MetroLine Bank, a digital lender offering instant point-of-sale financing for e-commerce checkouts. The product serves 18M monthly active shoppers across the US and EU and underwrites about 2.5M applications/month. A 20–40 bps change in default rate meaningfully impacts profitability, and regulators require that adverse action reasons be explainable.
The bank is rebuilding its probability of default (PD) model after a product expansion into new merchant categories caused a distribution shift. Your task is to propose, implement, and evaluate parametric vs non-parametric modeling approaches, and justify which you would ship given business, compliance, and operational constraints.
Dataset
You have 12 months of labeled outcomes for applicants. The label is whether the applicant became 90+ days past due within 6 months.
| Feature Group | Count | Examples | Notes |
|---|
| Applicant credit bureau | 18 | fico_score, utilization_pct, num_open_trades, delinquency_24m | Some missing for thin-file users |
| Transaction & merchant | 10 | cart_amount, merchant_category, merchant_risk_score, device_type | High-cardinality categorical |
| Behavioral / device | 9 | ip_risk_score, velocity_1h, account_age_days, email_domain | Potential leakage risks |
| Geography & compliance | 6 | country, region, sanctions_flag, pep_flag | Must be handled carefully |
| Derived aggregates | 8 | rolling_spend_30d, prior_defaults, repayment_ratio | Built from internal history |
- Size: ~30M applications, 51 features after cleaning
- Target:
default_6m (1 if 90+ DPD within 6 months, else 0)
- Class balance: 2.2% positive (defaults)
- Missingness: ~12% missing in bureau features (thin-file); ~3% missing in device signals
Success Criteria
- Risk performance: Improve AUC-PR over baseline by at least +0.05 while maintaining stable calibration.
- Operational: Batch scoring 2.5M apps/day with p95 < 25 ms/app on CPU in an online service.
- Compliance: Provide actionable explanations (top drivers) for adverse action; avoid using prohibited attributes.
- Stability: Robust to distribution shift across merchant categories; minimize performance cliffs.
Constraints
- Latency: p95 < 25 ms per application (including preprocessing).
- Interpretability: Must support reason codes; model governance prefers simpler forms unless lift is material.
- Monitoring: Must detect drift and calibration decay weekly.
- Data leakage: Any feature derived after application time is disallowed.
Deliverables
- Explain, in the context of this problem, the difference between parametric and non-parametric models (capacity, assumptions, sample complexity, interpretability).
- Propose one parametric and one non-parametric candidate model for PD (e.g., regularized logistic regression vs gradient-boosted trees), including when each is likely to win.
- Implement both models with a leakage-safe split, handle missingness and categorical features, and evaluate with metrics suited for imbalanced risk modeling.
- Recommend which model to ship and describe how you would monitor and retrain it.
