All work

Home Credit Default Risk

Credit risk · Supervised ML November 2025

A credit default pipeline built the way a lender actually has to defend it.

Project links

Results

0.963 ROC AUC Held out set.
500+ Engineered features Across application, bureau and behavioral sources.
~300K Applications

Skills

  • Python
  • Feature engineering
  • Model stacking
  • Hypothesis testing
  • scikit-learn
  • LightGBM

About this project

A credit default prediction pipeline across roughly 300 thousand loan applications. Most of the lift came from relational history rather than the application form: ratios instead of levels, recency weighted behavior, and deviation from peer group, giving 500+ engineered features feeding a stacked ensemble.

The deliverable was the validation as much as the model: decile analysis, rank ordering, probability calibration and stability over time, because a well ranked but badly calibrated score is quietly wrong for any decision that uses the probability as a number.