Skip to content

README.md

Latest commit

 

History

History
142 lines (90 loc) · 3.29 KB

README.md

File metadata and controls

142 lines (90 loc) · 3.29 KB

📊 Credit Risk – Probability of Default (PD) Modelling

🔎 Project Overview

This project builds a Probability of Default (PD) model to estimate borrower credit risk using the Home Credit Default Risk dataset.

The goal is to compare interpretable and performance-driven models and translate model outputs into business-friendly risk decisions.

📁 Dataset

  • Source: Home Credit Default Risk (Kaggle)

  • Total records: ~307K loan applications

  • Working sample: 70,000 observations

  • Target variable:

    • TARGET = 1 → Default
    • TARGET = 0 → Non-default

⚙️ Modelling Workflow

1️⃣ Data Preparation

  • Loaded loan-level retail credit data
  • Selected numeric features for modelling
  • Missing value handling using median imputation
  • Addressed class imbalance using class-weighted models

2️⃣ Models Built

Logistic Regression (Interpretability Model)

  • Baseline credit risk model
  • Used for transparent and explainable risk assessment

Random Forest (Performance Model)

  • Captures non-linear relationships
  • Used for stronger predictive ranking power

3️⃣ Model Evaluation

  • ROC-AUC used as primary metric due to imbalanced classes
  • Confusion matrix used for threshold analysis
  • Model comparison focused on ranking quality rather than accuracy

4️⃣ Threshold Optimization

  • Probability threshold adjusted to 0.30
  • Increased default detection while controlling false approvals
  • Demonstrates risk appetite trade-off used in credit strategy

📈 Results

Model ROC-AUC
Logistic Regression ~0.61
Random Forest ~0.69

Key Insight: Random Forest achieved stronger ranking power and better separation of high-risk borrowers.

🔑 Top Predictive Features

  • EXT_SOURCE_2
  • EXT_SOURCE_3
  • DAYS_BIRTH
  • DAYS_EMPLOYED
  • DAYS_LAST_PHONE_CHANGE
  • DAYS_REGISTRATION
  • DAYS_ID_PUBLISH

These variables were identified as strong drivers of borrower default risk.

🏦 Risk Bucket Segmentation

Predicted probabilities were converted into business-friendly risk bands:

  • Low Risk (PD < 10%)
  • Medium Risk (10% – 30%)
  • High Risk (PD > 30%)

Observed default rates increased consistently across buckets, demonstrating strong ranking capability of the model and enabling portfolio-level monitoring.

📊 Visualizations

ROC Curve

ROC Curve

Feature Importance

Feature Importance

Risk Bucket Analysis

Risk Buckets

💼 Business Interpretation

  • Logistic Regression supports explainability for regulatory and business stakeholders.
  • Random Forest improves predictive performance for borrower ranking.
  • Risk buckets translate model outputs into underwriting decisions.
  • Lower thresholds improve default capture but increase approval risk, highlighting the balance between growth and risk control.

🛠 Tools & Libraries

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

🚀 Key Learning Outcomes

  • End-to-end credit risk modelling workflow
  • Handling imbalanced datasets
  • Model comparison using ROC-AUC
  • Threshold tuning aligned with risk appetite
  • Feature importance interpretation
  • Risk segmentation for portfolio monitoring