Credit Risk Modeling (Basel III Framework)

A machine learning project that implements the three core components of the Basel III credit risk framework using historical Lending Club loan data (2007–2014).

---

Models

1. Probability of Default (PD)

Predicts the likelihood that a borrower will default on a loan.

• Algorithm: Logistic Regression with custom p-value computation
• Feature Selection: LassoCV (5-fold cross-validation)
• Evaluation: ROC-AUC, Confusion Matrix, Classification threshold = 0.75
• Output: pd_model.sav (generated on run, not tracked in repo)

2. Loss Given Default (LGD)

Estimates the percentage of the loan amount lost if a default occurs. Uses a two-stage approach:

• Stage 1 — Classification: Logistic Regression to predict whether any loss occurs (threshold = 0.45)
• Stage 2 — Regression: Linear Regression to estimate the magnitude of loss
• Evaluation: ROC-AUC (Stage 1), MSE + R² (Stage 2)
• Output: lgd_model_stage_1.sav, lgd_model_stage_2.sav (generated on run, not tracked in repo)

3. Exposure at Default (EAD)

Estimates the Credit Conversion Factor (CCF) — the proportion of the loan outstanding at the time of default.

• Algorithm: Linear Regression with custom p-value computation
• Target: CCF = (funded_amnt − total_rec_prncp) / funded_amnt
• Evaluation: R², MSE, Pearson Correlation
• Output: ead_model.sav (generated on run, not tracked in repo)

---

Dataset

Source: Lending Club Loan Data (2007–2014) — available on Kaggle.

Split	Samples
Train	373,028
Test	93,257

Class Distribution (PD target): ~89% good loans, ~11% bad loans

Required raw file: loan_data_2007_2014.csv (place in project root before running preprocessing)

---

Project Structure

Credit risk/
├── 01_Data_Preprocessing.ipynb          # Data cleaning, encoding, feature engineering
├── 02_Probability_of_Default_Model.ipynb # PD model training and evaluation
├── 03_LGD_and_EAD_Models.ipynb          # LGD (2-stage) and EAD model training
│
├── loan_data_2007_2014.csv      # Raw dataset (not included — download separately)
├── loan_data_inputs_train.csv   # Preprocessed training features (generated)
├── loan_data_inputs_test.csv    # Preprocessed test features (generated)
├── loan_data_targets_train.csv  # Training labels
├── loan_data_targets_test.csv   # Test labels
│
└── requirements.txt

---

Setup

pip install -r requirements.txt

Python version: 3.9+

---

Usage

Run the notebooks in order:

1. 01_Data_Preprocessing.ipynb — Cleans and preprocesses loan_data_2007_2014.csv, outputs train/test CSV files.
2. 02_Probability_of_Default_Model.ipynb — Trains and saves the PD model.
3. 03_LGD_and_EAD_Models.ipynb — Trains and saves the LGD and EAD models.

---

Key Technical Details

Custom Statistical Inference

Both logistic and linear regression models are extended to compute p-values for every feature coefficient, enabling statistical significance testing beyond standard sklearn outputs.

• Logistic Regression: Uses the Fisher Information Matrix to derive coefficient variances, then computes two-tailed z-test p-values.
• Linear Regression: Computes t-statistics and p-values using the t-distribution.

Feature Selection

LassoCV automatically selects significant features for the PD model by shrinking irrelevant coefficients to zero, reducing noise and improving generalization.

---

Results

Probability of Default (PD)

Metric	Value
ROC-AUC	0.690
Accuracy (threshold = 0.75)	87.3%

Confusion Matrix (test set, 93,257 samples):

	Predicted Good	Predicted Default
Actual Good	1,066	9,128
Actual Default	2,680	80,383

---

Loss Given Default — Stage 1 (Classification)

Metric	Value
ROC-AUC	0.620
Accuracy (threshold = 0.45)	57.7%

Confusion Matrix:

	Predicted 0	Predicted 1
Actual 0	523	3,250
Actual 1	410	4,465

---

Loss Given Default — Stage 2 (Regression)

Metric	Value
Pearson Correlation (actual vs predicted)	0.325
Mean predicted LGD	8.9%
Std	5.0%
Range	0% – 23.5%

---

Exposure at Default (EAD)

Metric	Value
R²	0.278
MSE	0.0290
Pearson Correlation	0.527
Mean predicted CCF	73.6%

---

Expected Loss Formula

After training all three models, the expected loss for any loan can be estimated as:

Expected Loss (EL) = PD × LGD × EAD

This is the standard Basel III formula used by financial institutions to quantify credit risk.