Credit Risk Prediction — ML System

Project Overview

This project implements an end-to-end machine learning system for predicting loan default risk using real-world financial application data.

The system is designed with production ML engineering principles, including:

• Modular feature engineering pipelines
• Reusable preprocessing and model training modules
• Multi-model benchmarking
• Cross-validation based performance validation
• Business-aware decision optimization using profit simulation
• Model interpretability via feature importance analysis
• Inference-ready model serialization

This project emphasizes real-world ML workflow, not just model accuracy.

---

Business Problem

Financial institutions must estimate the probability that a customer will default on a loan.

Accurate credit risk prediction enables:

• Risk-adjusted loan approvals
• Portfolio loss reduction
• Interest rate optimization
• Regulatory-compliant risk modeling

Objective

Predict probability of default (PD) for each applicant and optimize approval threshold for maximum portfolio profit.

---

Dataset

Dataset: Home Credit Default Risk Primary Table Used: application_train.csv

Future extensions could include:

• Credit bureau history
• Previous loan performance
• Installment payment behavior

---

System Architecture

Raw Data
↓
Feature Engineering (src/features)
↓
Preprocessing Pipeline (ColumnTransformer)
↓
Model Training Modules (src/models)
↓
Cross Validation Evaluation
↓
Model Comparison & Selection
↓
Business Decision Optimization (Profit Simulation)
↓
Model Serialization (Deployment Ready)

---

Repository Structure

src/
 ├ features/
 │   ├ build_features.py
 │   ├ pipeline.py
 │
 ├ models/
 │   ├ train_model.py
 │   ├ train_tree_model.py
 │   ├ train_histgb_model.py
 │   ├ train_xgb_model.py
 │   ├ evaluate_model.py
 │   ├ save_model.py
 │
 ├ decision/
 │   ├ profit_simulation.py
 │
notebooks/
 ├ 01_eda.ipynb
 ├ 03_modeling.ipynb

artifacts/
 ├ xgb_credit_model.joblib

---

Feature Engineering Strategy

Financial Stress Features

• Credit-to-Income Ratio
• Annuity-to-Income Ratio

Customer Stability Signals

• Employment anomaly detection
• Registration duration
• Phone activity recency

Demographic Features

• Age conversion from raw birth date encoding

Data Quality Handling

• Sentinel missing value handling
• Identifier column removal
• Redundant feature removal

---

Preprocessing Pipeline

Implemented using sklearn ColumnTransformer.

Numeric Pipeline

• Median Imputation
• Standard Scaling

Categorical Pipeline

• Most Frequent Imputation
• One-Hot Encoding with unknown category safety

---

Models Evaluated

Model	Purpose
Logistic Regression	Linear baseline
Random Forest	Nonlinear bagging baseline
Gradient Boosting	Sequential boosting baseline
HistGradientBoosting	Modern histogram boosting
XGBoost	Final production candidate

---

Model Performance

Validation ROC AUC

Model	ROC AUC
Logistic Regression	~0.749
Random Forest	~0.726
Gradient Boosting	~0.753
HistGradientBoosting	~0.759
XGBoost	~0.762

---

ROC Curve Comparison

---

Cross Validation Stability

Logistic Baseline Cross Validation:

• Mean ROC AUC: ~0.746
• Std Dev: ~0.0026

Indicates stable model generalization.

---

Final Model Selection

Selected Model: XGBoost

Selected because:

• Highest validation ROC AUC
• Strong tabular feature interaction modeling
• Industry standard for structured financial ML
• Stable training behavior

---

Model Interpretability

Feature importance analysis confirms dominant signals from:

• External credit risk score features (EXT_SOURCE variables)
• Financial stress ratio features
• Customer stability indicators
• Age / lifecycle features

---

Feature Importance Visualization

---

Business Decision Optimization (Profit Simulation)

Instead of using default probability threshold (0.5), a profit simulation layer was implemented to optimize loan approval decisions.

Simulation Includes:

• Interest revenue modeling
• Loss given default modeling
• Operational cost modeling

---

Key Finding

Optimal Approval Threshold ≈ 0.20

This reflects real-world credit risk asymmetry: Default losses are much larger than interest gains.

---

Profit vs Threshold Visualization

---

Deployment Readiness

Final model is saved as serialized pipeline artifact:

artifacts/xgb_credit_model.joblib

This includes: Feature Engineering Preprocessing Model Inference

---

▶ How To Run

Install Dependencies

pip install -r requirements.txt

---

Train Models

Run:

notebooks/03_modeling.ipynb

---

Load Model For Inference

import joblib

model = joblib.load("artifacts/xgb_credit_model.joblib")
preds = model.predict_proba(X_new)

---

Key Technical Learnings

• Feature engineering dominates tabular ML performance
• Boosting models outperform bagging on structured financial data
• Histogram boosting improves training efficiency significantly
• Cross-validation is critical for stable evaluation
• Business-aligned metrics outperform pure accuracy metrics

---

Future Improvements

Potential next enhancements:

• Multi-table feature aggregation
• Probability calibration for financial risk pricing
• Model monitoring and drift detection
• Real-time inference pipeline

---

Author

Built as a production-style machine learning system demonstrating:

• End-to-end ML pipeline engineering
• Financial tabular modeling best practices
• Business-aligned ML decision making