Risk-Based Loan Pricing ML

Author: Telman Maghrebi
Role: Data Scientist

End-to-end credit-risk modelling and risk-based pricing project using real-world Freddie Mac loan-level performance data.

This project predicts whether a loan will become seriously delinquent, calibrates the predicted probability of default, assigns borrowers into interpretable risk bands, and converts credit risk into a suggested risk-based APR.

---

Project Objective

The goal of this project is to build a practical machine learning workflow for credit-risk decisioning and risk-based pricing.

The project demonstrates:

• Credit-risk modelling
• Probability of default estimation
• Class imbalance handling
• Feature engineering
• Exploratory data analysis
• Logistic Regression with L1/L2 regularisation
• XGBoost challenger modelling
• Simple deep learning baseline
• Probability calibration
• Automatic threshold optimisation
• Risk-band segmentation
• Expected-loss based pricing

---

Dataset

The project uses Freddie Mac Single-Family Loan-Level Dataset files for the 2023 vintage.

The processed dataset contains:

Item	Value
Loan origination records	931,749
Monthly performance records	22,771,770
Bad loans	17,290
Bad-loan rate	1.86%

A bad loan is defined as a loan that ever reaches 90+ days delinquent.

---

Project Workflow

Raw Freddie Mac files
        ↓
Origination feature table
        ↓
Monthly performance target table
        ↓
Loan-level modelling dataset
        ↓
EDA and feature engineering
        ↓
Baseline and challenger models
        ↓
Probability calibration
        ↓
Automatic threshold selection
        ↓
Risk-band segmentation
        ↓
Expected-loss pricing engine

---

Exploratory Data Analysis

Target Distribution

The dataset is highly imbalanced, with bad loans representing only 1.86% of the total portfolio.

---

Numerical Feature Correlation

A Pearson correlation heatmap was created for numerical features. Highly collinear variables were removed before modelling.

High-correlation numerical pairs were saved to:

reports/high_correlation_numeric_pairs.csv

Spearman correlation pairs were also saved to:

reports/nonlinear_spearman_numeric_pairs_all.csv

---

Cross Plots for Highly Correlated Features

Scatter plots were generated for numerical feature pairs with high Pearson correlation.

---

Feature Engineering

The feature-engineering process includes:

1. Removing leakage columns from monthly performance history
2. Removing highly collinear numerical variables
3. Removing categorical variables with only one class
4. Removing high-cardinality categorical variables
5. One-hot encoding categorical variables
6. Handling class imbalance through undersampling
7. Splitting data into train, calibration and test sets

Removed Collinear Numerical Features

Kept feature	Dropped feature	Absolute Pearson correlation
mortgage_insurance_percentage	original_cltv	0.70
mortgage_insurance_percentage	original_ltv	0.70
original_cltv	original_ltv	1.00

Removed Categorical Features

Feature	Reason	Unique values
prepayment_penalty_mortgage_flag	only one class	1
amortization_type	only one class	1
interest_only_indicator	only one class	1
msa	high cardinality	435
postal_code	high cardinality	888

---

Class Imbalance

The target is highly imbalanced.

Original training distribution:

Class	Count	Share
Good loan	731,567	98.14%
Bad loan	13,832	1.86%

After undersampling:

Class	Count	Share
Good loan	41,496	75.00%
Bad loan	13,832	25.00%

Undersampling was used only on the training data. Evaluation was performed on the original imbalanced test set.

---

Model Results

Three models were compared:

Model	Type	ROC-AUC	PR-AUC	Role
XGBoost	Tree-based challenger	0.8248	0.1313	Best-performing model
Simple MLP	Deep learning challenger	0.8213	0.1161	Tested neural-network baseline
Logistic Regression L1	Interpretable baseline	0.8165	0.1130	Scorecard-style benchmark

The dataset is highly imbalanced, so PR-AUC was treated as the main model-selection metric.

XGBoost improved PR-AUC by approximately 16% compared with Logistic Regression:

(0.1313 - 0.1130) / 0.1130 = 16.2%

---

Logistic Regression Baseline

Logistic Regression was used as an interpretable scorecard-style benchmark.

L1 and L2 regularisation were compared:

Model	Validation ROC-AUC	Validation PR-AUC
Logistic Regression L1	0.8227	0.1117
Logistic Regression L2	0.8227	0.1114

L1 was selected because it achieved slightly better PR-AUC and supports sparse, interpretable feature selection.

---

Logistic Regression Threshold Selection

The automatic threshold-selection rule was:

Maximise bad-loan recall while keeping approval rate at or above 80%.

Selected calibrated PD threshold:

Metric	Value
Threshold	3.00%
Approval rate	81.63%
Decline / refer rate	18.37%
Precision	6.31%
Bad-loan recall	62.52%
False positive rate	17.54%

---

Logistic Regression Risk Bands

Risk band	Loan share	Actual bad rate	Average calibrated PD
Low risk	63.96%	0.49%	0.59%
Medium risk	17.67%	2.17%	2.13%
High risk	18.37%	6.31%	6.08%

The high-risk band is around 3.4x riskier than the total portfolio average and around 12.9x riskier than the low-risk band.

---

Logistic Regression Risk-Based Pricing

The pricing engine uses:

Expected Loss = PD × LGD
Suggested APR = Funding Cost + Operating Cost + Profit Margin + Expected Loss

Assumptions:

Component	Value
Loss given default	40%
Funding cost	4%
Operating cost	2%
Profit margin	3%

Pricing output by risk band:

Risk band	Average PD	Average expected loss	Average suggested APR
Low risk	0.59%	0.24%	9.24%
Medium risk	2.13%	0.85%	9.85%
High risk	6.08%	2.43%	11.43%

---

XGBoost Challenger Model

XGBoost was the best-performing model by PR-AUC.

Metric	Value
ROC-AUC	0.8248
PR-AUC	0.1313

XGBoost improved ranking performance compared with Logistic Regression and captured stronger non-linear relationships in borrower and loan origination features.

---

XGBoost Threshold Selection

The same automatic threshold-selection rule was applied:

Maximise bad-loan recall while keeping approval rate at or above 80%.

Selected calibrated PD threshold:

Metric	Value
Threshold	2.40%
Approval rate	80.36%
Decline / refer rate	19.64%
Precision	6.22%
Bad-loan recall	65.88%
False positive rate	18.76%

Compared with Logistic Regression, XGBoost captured 116 additional bad loans under the same approval-rate constraint.

2,278 - 2,162 = 116 additional bad loans captured

---

XGBoost Risk Bands

Risk band	Loans	Loan share	Actual bad rate	Average calibrated PD
Low risk	126,044	67.64%	0.52%	0.55%
Medium risk	23,709	12.72%	2.24%	1.70%
High risk	36,597	19.64%	6.22%	6.56%

XGBoost creates clear separation between low, medium and high-risk groups.

---

XGBoost Risk-Based Pricing Output

Risk band	Loans	Actual bad rate	Average PD	Expected loss	Suggested APR
Low risk	126,044	0.52%	0.55%	0.22%	9.22%
Medium risk	23,709	2.24%	1.70%	0.68%	9.68%
High risk	36,597	6.22%	6.56%	2.62%	11.62%

The XGBoost model produces a wider pricing spread than Logistic Regression and is selected as the strongest challenger model for risk-based pricing.

---

Business Interpretation

The project shows how borrower and loan origination features can be transformed into a practical credit-risk decisioning workflow.

The final workflow supports:

• Identifying high-risk loans
• Ranking applicants by calibrated probability of default
• Assigning interpretable risk bands
• Applying threshold rules based on approval-rate constraints
• Converting PD into expected loss
• Producing risk-sensitive APR recommendations

This maps directly to real-world underwriting and risk-based pricing use cases.

---

Project Structure

risk-based-loan-pricing-ml/
│
├── data/
│   ├── raw/
│   └── processed/
│
├── reports/
│   ├── figures/
│   │   ├── eda/
│   │   └── model/
│   ├── threshold_analysis.csv
│   ├── selected_threshold.csv
│   ├── risk_band_analysis.csv
│   ├── pricing_decisions.csv
│   ├── challenger_model_comparison.csv
│   ├── final_model_comparison.csv
│   ├── xgboost_threshold_analysis.csv
│   ├── xgboost_selected_threshold.csv
│   ├── xgboost_risk_band_analysis.csv
│   ├── xgboost_pricing_decisions.csv
│   └── xgboost_pricing_summary_by_risk_band.csv
│
├── src/
│   ├── freddie_schema.py
│   ├── inspect_freddie_files.py
│   ├── count_freddie_rows.py
│   ├── inspect_freddie_target.py
│   ├── build_freddie_target.py
│   ├── build_freddie_features.py
│   ├── build_model_dataset.py
│   ├── eda_freddie.py
│   ├── train_logistic_regression.py
│   ├── threshold_analysis.py
│   ├── risk_band_analysis.py
│   ├── pricing_engine.py
│   ├── train_challenger_models.py
│   ├── train_xgboost_calibrated.py
│   ├── xgboost_threshold_analysis.py
│   ├── xgboost_risk_band_analysis.py
│   ├── xgboost_pricing_engine.py
│   ├── plot_pricing_summary.py
│   └── plot_xgboost_pricing_summary.py
│
├── requirements.txt
├── .gitignore
└── README.md

---

How to Run

Create and activate a virtual environment:

python -m venv .venv
.\.venv\Scripts\Activate.ps1

Install dependencies:

pip install -r requirements.txt

Build the modelling dataset:

python src\build_freddie_target.py
python src\build_freddie_features.py
python src\build_model_dataset.py

Run EDA:

python src\eda_freddie.py

Train Logistic Regression:

python src\train_logistic_regression.py
python src\threshold_analysis.py
python src\risk_band_analysis.py
python src\pricing_engine.py

Train challenger models:

python src\train_challenger_models.py

Train and price with XGBoost:

python src\train_xgboost_calibrated.py
python src\xgboost_threshold_analysis.py
python src\xgboost_risk_band_analysis.py
python src\xgboost_pricing_engine.py
python src\plot_xgboost_pricing_summary.py

---

Notes

The model is intentionally structured like a practical credit-risk workflow.

Logistic Regression is used as a transparent scorecard-style baseline.

XGBoost is used as a stronger challenger model for improved ranking performance.

The simple MLP model was tested as a deep learning baseline, but it did not outperform XGBoost on this structured tabular dataset.

---

Future Improvements

Planned improvements:

• Add SHAP-based reason codes
• Add model drift monitoring
• Add population stability index
• Add champion/challenger monitoring
• Add Streamlit pricing dashboard
• Add reject inference discussion
• Add macroeconomic features
• Add time-based validation
• Add fairness and bias analysis