Business Context
You’re on the Risk Modeling team at MetroBank, a digital lender operating in the US and EU with 12M consumer customers and $9.5B in outstanding unsecured personal loans. The bank is launching an instant-decision loan product where approvals must be explained to regulators and to customers under adverse action requirements. A recent model review found that your current logistic regression model is unstable across retrains: coefficients flip signs month-to-month and small data refreshes cause large swings in predicted risk. The review suspects multicollinearity among engineered financial features (e.g., multiple debt-to-income variants, overlapping utilization metrics, and correlated bureau aggregates).
Your task is to propose and implement a robust approach to detect, diagnose, and mitigate multicollinearity while maintaining strong predictive performance and meeting interpretability constraints.
Dataset
You have a supervised dataset built from loan applications and 12 months of bureau + transaction aggregates.
| Feature Group | Count | Examples | Notes |
|---|
| Applicant demographics | 6 | age_bucket, region, employment_status | Categorical + ordinal |
| Income & cashflow | 12 | monthly_income, income_volatility_90d, paycheck_count_90d | Some missing for gig workers |
| Debt & utilization | 14 | dti, dti_alt, revolving_util, total_balance, balance_to_income | Highly correlated ratios |
| Credit bureau aggregates | 10 | inquiries_6m, tradelines_open, delinquencies_24m | Lagged aggregates |
| Product & channel | 5 | requested_amount, term_months, acquisition_channel | Channel drift over time |
Additional notes:
- Data is time-indexed by application date.
- You’ve observed that dti, balance_to_income, and revolving_util are strongly correlated, and several features are near-linear combinations due to shared denominators.
Success Criteria
- Model stability: coefficient sign flips for top 20 features should be rare across monthly retrains (e.g., <10% flip rate).
- Predictive performance: maintain strong discrimination on a forward-looking test set.
- Target: AUC-ROC ≥ 0.78 and AUC-PR ≥ 0.32 (default rate is low).
- Interpretability: provide a defensible explanation of how correlated features were handled and how the final feature set/model supports adverse action reasons.
- Operational usability: training must run in <30 minutes on a single 16-core machine; batch scoring for 200K apps/day must be <100 ms per application (CPU).
Constraints
- Regulatory: Must be able to justify feature inclusion/removal; avoid opaque transformations that prevent reason codes.
- Temporal leakage: Must use time-based validation (no random K-fold).
- Data drift: acquisition_channel mix changes quarterly; correlated features may shift.
Deliverables
- A step-by-step plan to detect multicollinearity (what diagnostics you run, thresholds, and why).
- A modeling approach that mitigates multicollinearity (e.g., regularization, feature selection, dimensionality reduction) and a clear rationale.
- A training/validation strategy that avoids leakage and measures stability.
- A short proposal for how you would communicate the approach to Risk/Compliance (what artifacts you’d produce).
- (Optional) How you’d monitor multicollinearity and coefficient stability in production over time.
