LendingClub Loan Default Prediction

What is Loan Default Prediction?

Loan default prediction estimates the probability that a borrower will fail to repay their loan within the observation period. LendingClub, the largest P2P lender, has issued $70B+ in loans with a historical default rate of 12–15%. Accurate default prediction enables proper loan pricing — charging higher interest rates to riskier borrowers while remaining competitive for borrowers.

Traditional credit scoring (FICO) uses logistic regression with 5 factors, achieving ~70% AUC. Machine learning models incorporating 100+ features achieve 75–80% AUC by capturing non-linear interactions: high DTI + short credit history is riskier than the sum of its parts. The challenge is calibration — predicted probabilities must match observed default rates for accurate pricing.

Survival analysis provides richer insights than binary classification. Instead of predicting whether default occurs within 36 months, survival models estimate the time-to-default curve — the probability of default at each month. This enables: (1) more accurate expected loss calculation; (2) early warning systems that trigger at specific hazard rates; and (3) dynamic pricing that adjusts as the loan seasons.

Mathematical Foundation

Kaplan-Meier Survival Estimate:

Where:

• — number of defaults at time
• — number of loans at risk just before
• Intuition: Probability of surviving (not defaulting) beyond time

Cox Proportional Hazards:

Where:

• — baseline hazard function
• — linear predictor from features
• Intuition: Risk factors multiply the baseline hazard multiplicatively

Brier Score (calibration metric):

Where:

• — actual outcome (0/1)
• — predicted probability
• Intuition: Mean squared error of probability predictions

Model Architecture

Performance Results

Real-World Case Study

LendingClub's credit model evaluates 1,000+ features per loan application, including traditional credit data, employment verification, and bank account analysis. Their model achieves 79% AUC, enabling them to price loans from 6.0% (Grade A) to 29.0% (Grade G). Key insight: loans with high DTI but strong payment history perform better than their DTI suggests — ML models capture this interaction effect that logistic regression misses.

Deployment

Common Pitfalls

1. Observation bias: Loans with short histories haven't had time to default — use survival analysis
2. Competing risks: Prepayment removes loans from observation — account for censoring
3. Vintage effects: Loans originated in different economic conditions perform differently
4. Calibration drift: Default rates change with economic conditions — recalibrate quarterly
5. Sample bias: Only approved loans have outcomes — reject inference needed for full picture

Summary with Key Takeaways

This project built an XGBoost loan default predictor achieving 0.794 AUC with calibrated probabilities. Survival analysis provides time-to-default curves for dynamic pricing. Key insights: calibration is essential for accurate pricing — a 1% calibration error translates to millions in mispriced loans; survival models outperform binary classifiers for long-horizon predictions; and vintage analysis reveals economic regime effects invisible in pooled data.