Business Context
LendFlow, a digital consumer lending platform processing roughly 200K loan applications per month, wants to improve default prediction at underwriting time. The risk team needs a neural-network-based classifier that can outperform a logistic regression baseline while remaining fast enough for batch retraining and low-latency scoring.
Dataset
The training data contains historical funded loans with borrower profile, credit bureau, application, and early repayment behavior features.
| Feature Group | Count | Examples |
|---|
| Borrower demographics | 6 | age, employment_length_years, housing_status |
| Credit history | 12 | fico_score, revolving_utilization, delinquencies_2y |
| Application details | 9 | loan_amount, term_months, interest_rate, purpose |
| Banking and cash flow | 8 | monthly_income, debt_to_income, avg_balance_90d |
| Early payment behavior | 7 | first_payment_missed, autopay_enabled, days_past_due_30d |
- Size: 320K loans, 42 features
- Target: Binary — default within 12 months of origination
- Class balance: 11.4% default, 88.6% non-default
- Missing data: ~9% missing in banking features, ~4% missing in employment fields, and sparse missingness in bureau attributes
Success Criteria
A strong solution should improve minority-class detection over a linear baseline and achieve:
- ROC-AUC >= 0.84
- PR-AUC >= 0.42
- Recall >= 0.70 at precision >= 0.35
Constraints
- Inference must stay under 20 ms per application in online scoring.
- The model should support monthly retraining on newly booked loans.
- The risk team needs enough interpretability to review top drivers at portfolio level.
- The solution must avoid leakage from post-decision features.
Deliverables
- Build a neural network classifier for 12-month default prediction.
- Explain why a neural network is appropriate versus simpler tabular baselines.
- Design preprocessing for mixed numerical and categorical features with missing values.
- Define a validation strategy and threshold-selection approach aligned to credit risk.
- Report model performance, calibration, and operational tradeoffs for deployment.
