Advance Credit Risk Modeling - Loan Default Prediction

This project presents a full-cycle Credit Risk Modeling solution to predict the likelihood of a borrower defaulting on a loan. It involves meticulous data cleaning, feature engineering, model training, business-aligned metric optimization, and deployment using Streamlit. Designed with real-world financial services impact in mind, the model prioritizes recall to minimize false negatives (i.e., not catching risky borrowers).

---

🚀 Project Overview

• Goal: Predict whether a borrower will default on a loan.
• Dataset: Provided by a financial institution with borrower-level and loan-level details.
• Target Variable: default (1 = default, 0 = not default)
• Business Objective: High recall for defaulters to minimize risk exposure.
• Deployment: Web app hosted using Streamlit Cloud.

---

📊 Exploratory Data Analysis (EDA) & Preprocessing

✅ Class Imbalance

The dataset was highly imbalanced:

• Techniques used: SMOTE-Tomek, oversampling, and threshold tuning.

🛑 Data Leakage

Handled properly by eliminating leak-prone features like disbursal_date, installment_start_dt, and derived leakage indicators.

📉 Processing Fee Anomaly

Boxplots revealed processing_fee > loan_amount, which is invalid. These anomalies were cleaned or capped appropriately.

🧼 Categorical Feature Cleaning

• loan_purpose cleaned and grouped into standard categories.
• One-hot encoding and WoE/IV analysis used for feature transformation and selection.

---

🔍 Feature Engineering

Key New Features:

• Loan-to-Income Ratio (LTI): loan_amount / income
• Delinquency Ratio
• Average DPD per Delinquency

Insights:

• High LTI, delinquency_ratio, and avg_dpd_per_delinquency were strong predictors of default.
• Defaulted customers had younger age, longer loan tenure, and higher credit utilization.

---

📐 Feature Selectio

Multicollinearity Check (VIF)

Dropped correlated features: sanction_amount, processing_fee, gst, net_disbursement, principal_outstanding.

WoE & IV-Based Categorical Feature Selection:

Top features:

• credit_utilization_ratio
• avg_dpd_per_delinquency
• loan_to_income
• loan_purpose
• residence_type
• loan_tenure_months
• loan_type
• age, etc.

---

🤖 Model Training & Optimization

Model Attempt 1 default:

Model	Accuracy	Recall (Defaulters)
Logistic Regression (Basic)	96%	0.70
Random Forest	96%	0.69
XGBoost	96%	0.75

Final Model:

• Logistic Regression
• SMOTE-Tomek
• Optuna for Hyperparameter Tuning
• Business chose Logistic Regression for explainability

Final Metrics:

• Accuracy: 93%
• Recall (Defaulters): 0.95
• AUC: 98.3%
• Gini Coefficient: 0.967

---

📈 Model Evaluation

Metrics per class

ROC Curve

KS Statistic

• KS Value: 85.98% at Decile 8

• Indicates strong rank-ordering capability.

Feature Importance

---

📦 Deployment

• App Framework: Streamlit
• Main Files: main.py, prediction_helper.py
• Hosting: Streamlit Cloud

---

🧠 Business Impact

• Enables better credit risk filtering.
• High recall helps reduce bad debt.
• Easy model interpretability aids compliance and auditing.

---

📁 Folder Structure

Advance_Credit_Risk_Model_Loan_prediction/
├── data/
├── notebooks/
├── main.py
├── prediction_helper.py
├── README.md
├── requirements.txt
├── images/
│   ├── ks_statistic.png
│   ├── roc_curve.png
│   ├── confusion_matrix.png
│   └── streamlit_app_screenshot.png
│   └── metrics.png
│   └── feature_importance.png
├── artifacts/
│   └── modeldata.joblib

---

✍️ Author

• Mehul Ligade
• GitHub: @mehulcode12
• CodeBasics
• GitHub: @mehulcode12

---

🙌 Acknowledgements

This project was completed as part of the Codebasics Data Science Bootcamp. Special thanks to mentors and the open-source community for libraries and frameworks.

📌 Note

You are welcome to use this project as a reference. Please give credit to CodeBasics if you find it helpful.