Credit Default Prediction Pipeline

Overview

This project builds a production-ready binary classification pipeline for predicting the probability of credit default within the next 12 months.

Problem Statement

Given bureau data and application features for a loan applicant, predict the probability of default (PD) to inform underwriting decisions.

Approach

1. Data Ingestion — Pulling from multiple bureau data sources
2. Feature Engineering — WoE encoding, lag features, aggregation across accounts
3. Model Training — XGBoost with Optuna hyperparameter search
4. Validation — Walk-forward validation preserving temporal ordering
5. Deployment — Batch scoring via MLflow model registry

Results

Key Learnings

• Temporal leakage is the #1 enemy in credit risk models
• WoE encoding significantly improved model stability
• Monotonicity constraints improved trust with the business

Code

Full code available on GitHub.