Skip to content

README.md

Latest commit

 

History

History
236 lines (189 loc) · 8.41 KB

README.md

File metadata and controls

236 lines (189 loc) · 8.41 KB

Credit Risk Modeling: End-to-End Quantitative Framework

A production-grade credit risk modeling pipeline implementing Probability of Default (PD), Loss Given Default (LGD), Exposure at Default (EAD), credit scorecard development, macroeconomic stress testing, and Basel III regulatory capital calculations.

Project Overview

This project builds a complete credit risk analytics framework from the ground up:

Component Description
Data Generation Synthetic portfolio of 50,000 consumer loans with realistic correlations
EDA 6+ diagnostic visualizations with statistical summaries
Feature Engineering 35+ features including ratios, interactions, and log transforms
PD Models Logistic Regression, Random Forest, Gradient Boosting, XGBoost
LGD Model Gradient Boosting with logit-transformed beta-distributed target
EAD Model Credit Conversion Factor (CCF) approach per Basel framework
Credit Scorecard WoE/IV-based scorecard with PDO scaling (300-850 range)
Model Validation ROC/AUC, KS, Gini, PSI, calibration, lift, gains charts
Stress Testing 4 macroeconomic scenarios (baseline through deep depression)
Regulatory Capital Basel III IRB capital formula with Vasicek ASRF model
Thesis PDF Auto-generated 20-page thesis with embedded charts

Project Structure

credit-risk-model/
├── main.py                      # Run the complete pipeline
├── generate_thesis_pdf.py       # Generate thesis/documentation PDF
├── config.py                    # Configuration & hyperparameters
├── requirements.txt             # Python dependencies
├── README.md
├── src/
│   ├── __init__.py
│   ├── data_generator.py        # Synthetic data with latent variable model
│   ├── eda.py                   # Exploratory Data Analysis & plots
│   ├── feature_engineering.py   # Feature pipeline (scaling, encoding, transforms)
│   ├── pd_model.py              # PD model suite (4 algorithms)
│   ├── lgd_model.py             # Loss Given Default model
│   ├── ead_model.py             # Exposure at Default / CCF model
│   ├── scorecard.py             # WoE credit scorecard
│   ├── validation.py            # 9-panel validation dashboard
│   ├── stress_testing.py        # Macroeconomic stress scenarios
│   └── capital.py               # Basel III IRB capital calculator
├── output/
│   ├── figures/                 # All generated charts (15+ PNGs)
│   ├── models/                  # Model comparison results
│   └── data/                    # Dataset & summary statistics
└── docs/
    └── Credit_Risk_Modeling_Thesis.pdf

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Run the Full Pipeline

python main.py

This single command executes all 10 stages and generates:

  • 15+ publication-quality figures in output/figures/
  • Model results in output/models/
  • Complete dataset in output/data/
  • 20-page thesis PDF in docs/

3. Generate Only the Thesis PDF

python generate_thesis_pdf.py

Pipeline Stages

Stage 1: Data Generation

Generates 50,000 synthetic loan records using a latent variable model where default probability is driven by:

  • Borrower features: FICO score, income, employment length, home ownership
  • Credit history: utilization, delinquencies, inquiries, public records
  • Loan terms: amount, rate, term, purpose, DTI ratio
  • Macro variables: GDP growth, unemployment, fed funds rate, house price index

Stage 2: Exploratory Data Analysis

Produces diagnostic visualizations including:

  • Default rate distribution (count & percentage)
  • Feature distributions segmented by default status
  • Correlation heatmap of 20+ features
  • FICO score analysis with default rate by score band
  • Macroeconomic sensitivity analysis
  • Loan purpose default rate comparison

Stage 3: Feature Engineering

Creates 15+ derived features:

  • Ratios: loan-to-income, balance-to-income, payment-to-income
  • Interactions: FICO x utilization, FICO x DTI
  • Transforms: log(income), log(loan), log(balance)
  • Indicators: high utilization flag, delinquency flag
  • Economic: real interest rate (nominal - GDP growth)

Stage 4: PD Model Training

Trains and compares four models with 5-fold stratified cross-validation:

Model Key Hyperparameters
Logistic Regression C=0.1, L2, balanced weights
Random Forest 300 trees, depth=8, leaf=50
Gradient Boosting 200 trees, lr=0.05, depth=4
XGBoost 300 trees, lr=0.05, L1+L2 reg

Stage 5: Model Validation

Nine-panel validation dashboard covering:

  • ROC curve (train vs test)
  • Precision-Recall curve
  • Kolmogorov-Smirnov chart
  • Calibration plot
  • Score distribution by class
  • Confusion matrix (at optimal threshold)
  • Cumulative gains chart
  • Lift chart (by decile)
  • Population Stability Index (PSI)

Stage 6: Credit Scorecard

WoE-based scorecard with:

  • Information Value (IV) ranking of all features
  • Points-to-Double-Odds (PDO) scaling: base 600, PDO=20
  • Rating grades: AAA through D
  • Score range: 300-850

Stage 7-8: LGD & EAD Models

  • LGD: Gradient Boosting with logit-transformed target on defaulted loans
  • EAD: Credit Conversion Factor model: EAD = Drawn + CCF x (Limit - Drawn)

Stage 9: Stress Testing

Four macroeconomic scenarios with second-order effects:

Scenario GDP Shock Unemployment Rate Shock
Baseline 0% 0% 0%
Mild Recession -2% +3% +1%
Severe Recession -5% +7% +2.5%
Deep Depression -10% +12% +4%

Stage 10: Basel III Capital

IRB capital calculation using:

  • Vasicek asymptotic single-risk-factor (ASRF) model
  • 99.9% confidence level
  • Asset correlation function: R(PD)
  • Maturity adjustment: 2.5 years
  • Capital components: CET1 (4.5%), Tier 1 (6%), Total (8%), Conservation Buffer (2.5%)

Key Equations

Expected Loss:

EL = PD x LGD x EAD

Basel IRB Capital:

Conditional PD = Phi[(Phi^-1(PD) + sqrt(R) * Phi^-1(0.999)) / sqrt(1-R)]
K = [LGD * Conditional_PD - PD * LGD] * Maturity_Adjustment
RWA = K * 12.5 * EAD

Credit Score:

Score = Offset + Factor * ln((1-PD)/PD)
Factor = PDO / ln(2)
Offset = BaseScore - Factor * ln(TargetOdds)

Output Files

File Description
output/figures/01_default_distribution.png Target variable analysis
output/figures/02_feature_distributions.png Feature distributions by default
output/figures/03_correlation_matrix.png Feature correlation heatmap
output/figures/04_fico_analysis.png FICO score deep-dive
output/figures/05_macro_impact.png Macroeconomic sensitivity
output/figures/06_loan_purpose.png Purpose-based analysis
output/figures/07_model_comparison.png PD model CV comparison
output/figures/08_feature_importance.png RF & XGBoost importances
output/figures/09_validation_dashboard.png 9-panel validation suite
output/figures/10_decile_analysis.png Risk decile calibration
output/figures/11_scorecard.png Credit scorecard results
output/figures/12_lgd_analysis.png LGD model diagnostics
output/figures/13_ead_analysis.png EAD/CCF model results
output/figures/14_stress_testing.png Stress test scenarios
output/figures/15_capital_analysis.png Basel III capital breakdown
output/all_results.json All numerical results
docs/Credit_Risk_Modeling_Thesis.pdf Full thesis document

Dependencies

  • Python 3.9+
  • numpy, pandas, scipy, statsmodels
  • scikit-learn, xgboost
  • matplotlib, seaborn
  • reportlab, Pillow

References

  1. Merton, R.C. (1974). On the pricing of corporate debt. Journal of Finance.
  2. Basel Committee (2006). International Convergence of Capital Measurement.
  3. Basel Committee (2010). Basel III: Global regulatory framework.
  4. Vasicek, O.A. (2002). The distribution of loan portfolio value. Risk.
  5. Chen, T. & Guestrin, C. (2016). XGBoost. KDD.
  6. Altman, E.I. (1968). Financial ratios and corporate bankruptcy. Journal of Finance.
  7. Thomas, L.C. (2009). Consumer Credit Models. Oxford University Press.

License

This project is for educational and research purposes.