Credit Risk Modeling & Policy Optimization

LendingClub Loan Default Prediction

Overview

This project builds an end-to-end credit risk modeling pipeline using historical LendingClub loan data.

The objective is not only to predict loan default risk, but to translate model outputs into actionable credit policies using cost-sensitive decisioning and Expected Loss (EL) analysis.

Unlike typical classification-focused projects, this work emphasizes:

• Leakage-aware modeling
• Business-aligned decision thresholds
• Probability calibration
• Expected Loss–driven approval policies

Problem Statement

Given borrower and loan characteristics available at loan origination, predict whether a loan will default and design a decision framework that balances:

• Loan approval volume
• Credit losses from defaulted loans

Dataset

• Source: LendingClub (Kaggle: https://www.kaggle.com/datasets/gabrieloliveirasan/lending-club-credit-risk)
• Observations: ~160k+ loans after cleaning
• Target Variable: loan_status
  • 0 → Non-default (Fully Paid)
  • 1 → Default (Charged Off / Default)

Note: Only features known before loan issuance were retained.

Key Methodological Choices

Target Construction

Loan status categories were mapped into a binary credit risk label to reflect real-world default outcomes.

Leakage Prevention

All post-origination variables were removed, including:

• Payment dates
• Recoveries
• Collections activity

This ensures:

• Honest model evaluation
• Realistic deployability

Data Preprocessing

• Median imputation (numerical features)
• Most-frequent imputation (categorical features)
• One-Hot Encoding for categorical variables
• Feature standardization for numerical variables
• Memory-safe sparse preprocessing

Models Trained

Model	Purpose
Logistic Regression	Interpretable baseline
XGBoost	Non-linear high-performance model

Validation Strategy

• Train/Test split
• Cross-validation using ROC-AUC
• Final test ROC-AUC (XGBoost): ~0.78

Cost-Sensitive Decisioning

Decision thresholds were optimized using a cost-based framework:

• False Negatives assigned higher cost than False Positives
• Threshold chosen to minimize expected misclassification cost

This aligns model behavior with real-world credit risk priorities.

Probability Calibration

XGBoost probabilities were calibrated using Isotonic Regression to produce reliable Probability of Default (PD) estimates.

Benefits:

• Interpretable PD values
• Meaningful downstream financial calculations

Expected Loss Framework

Expected Loss is computed as:

Expected Loss = PD × LGD × EAD

Assumptions:

• LGD approximated for unsecured personal loans
• EAD proxied by loan amount

This enables policy-level decision-making rather than pure classification.

Credit Policy Analysis

Approval policies were evaluated using Expected Loss to study trade-offs between:

• Approval rate
• Bad loan approval rate

A policy frontier enables selection of conservative or aggressive credit strategies.

Results Summary

• Cost-optimized approval rate: ~72%
• Significant reduction in high-cost default approvals
• Stable policy behavior across thresholds
• No evidence of leakage-driven performance inflation

Key Takeaways

• Credit risk modeling is not just about AUC
• Business costs must drive decision thresholds
• Probability calibration is critical
• Expected Loss bridges ML and real credit policy

Limitations & Future Work

• Simplified LGD assumptions
• No macroeconomic stress testing
• No time-based (vintage) validation

Potential extensions:

• Explicit LGD modeling
• Stress testing under adverse scenarios
• Time-aware validation

Tools & Libraries

• Python
• Pandas, NumPy
• Scikit-learn
• XGBoost
• Matplotlib

Final Note

This project emphasizes realistic, deployable credit risk modeling with strong business alignment rather than leaderboard optimization.