End-to-end Basel II compliant credit risk modeling pipeline built on the Lending Club dataset (800K+ consumer loans). Implements Probability of Default (PD), Loss Given Default (LGD), and Exposure at Default (EAD) models with SHAP explainability, PyTorch neural network comparison, and Population Stability Index (PSI) monitoring.
| Model | Train AUC | Test AUC | PSI |
|---|---|---|---|
| Logistic Regression (PD) | 0.7059 | 0.6892 | 0.0863 |
| Neural Network (PD) | 0.7141 | 0.6955 | 0.0550 |
| LGD Stage 1 Classification | - | 0.7137 | - |
| EAD Linear Regression | - | R2: 0.1885 | - |
credit-risk-analytics/
├── notebooks/
│ ├── 01_data_preprocessing.ipynb # Feature engineering, WoE encoding, Basel II temporal split
│ ├── 02_pd_model_scorecard.ipynb # PD model, credit scorecard, cutoff analysis
│ ├── 03_lgd_ead_modeling.ipynb # LGD/EAD modeling, Expected Loss, stress testing
│ ├── 04_shap_explainability.ipynb # SHAP global/local explanations, adverse action notices
│ └── 05_pytorch_nn_comparison.ipynb # PyTorch NN vs logistic regression, PSI monitoring
├── requirements.txt
└── README.md
- Loads 500K Lending Club loans (2007-2018)
- Defines default under Basel II criteria (90+ days past due or charged off)
- Handles class imbalance (5:1 ratio) through weighted loss in modeling stage
- Weight of Evidence (WoE) encoding for categorical features with Information Value analysis
- Temporal train/test split: Train 2007-2015, Test 2016-2018 (no data leakage)
- Logistic regression with custom weighted loss (class_weight='balanced')
- ROC-AUC: 0.7023 (train), 0.6892 (test)
- KS Statistic: 0.2928 (train), 0.2766 (test)
- Gini Coefficient: 0.4045 (train), 0.3784 (test)
- Credit scorecard converting PD to interpretable 223-831 score range
- Cutoff analysis: score of 600 yields 56.7% approval rate at 10.6% default rate
- Two-stage LGD model: Stage 1 classification (AUC 0.7137) + Stage 2 regression
- Grade-level historical average LGD (40.07% Grade A to 47.47% Grade G)
- Grade-level EAD rate (53.77% Grade A to 83.94% Grade G)
- Expected Loss = PD x LGD x EAD: Mean EL $2,047 per loan, Total Portfolio EL $862M
- Stress testing: 295% EL increase under extreme scenario (3x PD, 2x LGD)
- LinearExplainer SHAP values on 2000 loan sample
- Global importance: sub_grade_woe dominant (mean abs SHAP 0.5413)
- Local explanations: waterfall plots for individual loan decisions
- Adverse action notices for regulatory compliance (GDPR Article 22, ECOA)
- SHAP vs coefficient ranking: top 5 features identical confirming model consistency
- 3-layer neural network (64-32-16 neurons) with BatchNorm and Dropout
- Custom weighted BCE loss addressing 5:1 class imbalance (pos_weight=4.4x)
- Neural network test AUC 0.6955 vs logistic regression 0.6892
- PSI monitoring: both models stable (PSI < 0.1) between training and test periods
- Quarterly PSI trend simulation showing gradual drift detection over 2016-2018
In credit risk modeling, random train/test splits violate a fundamental principle — a model built today will be used to score future applicants, not randomly sampled past applicants. Basel II explicitly requires out-of-time validation where the model is trained on historical data and validated on a more recent holdout period. Using 2007-2015 for training and 2016-2018 for testing replicates real production conditions and provides an honest estimate of how the model will perform on new loan applications.
WoE encoding is the industry standard for credit scorecard development, used by banks since the 1950s. It has three key advantages over one-hot encoding for credit applications: it handles missing values naturally by treating them as a separate bin, it captures monotonic relationships between categorical variables and default probability, and it directly measures the predictive power of each category through Information Value (IV). The IV threshold (0.1 for medium, 0.3 for strong predictors) is a well-established industry benchmark documented in Basel II internal ratings-based guidance.
Loss Given Default has a bimodal distribution — a significant proportion of defaulted loans result in either total loss (LGD = 1) or near-full recovery (LGD close to 0), with a continuous distribution in between. A single regression model poorly captures this bimodal behavior. The two-stage approach (classification first, then regression on partial recovery cases) is described in the Basel Committee on Banking Supervision working paper on LGD estimation and is implemented by major banks including in their IFRS 9 models.
Basel II Foundation IRB approach permits the use of supervisory estimates for LGD and EAD for retail exposures when loan-level recovery data is limited. The Lending Club dataset records total payments received but does not capture post-default collection costs, time value of recoveries, or workout expenses — all of which are required for accurate loan-level LGD regression. Using grade-level historical averages is consistent with the Foundation IRB approach and is standard practice for homogeneous retail loan portfolios where individual recovery drivers are not fully observable in the data.
Population Stability Index is the standard metric for credit model monitoring in retail banking, documented in OCC guidance on model risk management (SR 11-7). Unlike accuracy metrics which require knowing actual outcomes (which take months or years to realize for loans), PSI measures score distribution shift immediately using only the input score distribution. Banks typically run PSI monthly or quarterly — PSI below 0.1 indicates stability, 0.1-0.2 warrants investigation, and above 0.2 triggers model recalibration.
Regulatory frameworks including GDPR Article 22 (right to explanation for automated decisions) and the Equal Credit Opportunity Act (ECOA) require that credit decisions be explainable at the individual loan level. SHAP (SHapley Additive exPlanations) is grounded in cooperative game theory and provides the only attribution method that satisfies all three desirable properties simultaneously: local accuracy, missingness, and consistency. It is model-agnostic and has become the industry standard for explainable AI in banking since its adoption in the Federal Reserve's guidance on fair lending model validation.
Basel II Compliance The temporal train/test split (2007-2015 train, 2016-2018 test) follows Basel II out-of-time validation requirements. LGD and EAD use grade-level historical averages consistent with the Foundation IRB approach for retail portfolios.
Custom Weighted Loss Both logistic regression and neural network use weighted loss functions to prevent majority class collapse. The positive class weight (4.4x) is derived from the training data imbalance ratio.
SHAP Explainability SHAP values provide loan-level adverse action notices satisfying GDPR Article 22 and SR 11-7 Model Risk Management guidelines for explainable automated credit decisions.
PSI Monitoring Population Stability Index tracks score distribution shift between training and production populations. PSI below 0.1 indicates model stability; above 0.2 triggers recalibration.
Source: Lending Club Loan Data
- 800K+ consumer loans issued 2007-2018
- 150+ features including credit history, employment, loan characteristics
- Download and place
accepted_2007_to_2018Q4.csv.gzindata/folder
# Clone repository
git clone https://github.com/karthik-1604/Credit-Risk-Analytics.git
cd Credit-Risk-Analytics
# Create virtual environment
python -m venv venv
venv\Scripts\activate # Windows
source venv/bin/activate # Mac/Linux
# Install dependencies
pip install -r requirements.txt
# Run notebooks in order
# Start with 01_data_preprocessing.ipynbPython, scikit-learn, PyTorch, SHAP, pandas, NumPy, matplotlib, seaborn