💳 Loan Default Prediction — Credit Risk Scorecard

End-to-end credit risk ML pipeline simulating a production-grade Probability of Default (PD) model used by banks and NBFCs — Logistic Regression scorecard + Gradient Boosting classifier with industry-standard credit risk metrics: Gini Coefficient 0.521 · KS Statistic 0.395 · ROC-AUC 0.761 · 5-Fold CV AUC 0.759 ± 0.009

---

📌 Project Overview

This project builds a credit risk scorecard — the core model used by HDFC Bank, ICICI Bank, Bajaj Finance, and NBFCs to decide whether to approve or reject a loan application. It mirrors the full credit analytics workflow:

• Synthetic loan portfolio generation (German Credit / Lending Club inspired)
• Feature engineering with FOIR, LTV proxy, credit bands, WoE encoding
• Model training: Logistic Regression (interpretable scorecard) + Gradient Boosting (predictive power)
• Credit risk evaluation: Gini, KS Statistic, Information Value (IV), Weight of Evidence (WoE)
• Portfolio scoring with 5 credit grades (A to E)
• Interactive dashboard with risk distribution, feature importance, IV table

Relevance: Directly applicable to Credit Risk Analyst, Risk Analytics, and Data Analyst (Finance) roles at HDFC Bank, ICICI Bank, Bajaj Finance, Moody's, CRISIL, Acuity Knowledge Partners, EXL Service.

---

🏗️ Repository Structure

loan-default-prediction/
│
├── src/
│   ├── generate_data.py        # Synthetic loan portfolio generator (5,000 loans)
│   ├── train_model.py          # Full ML pipeline — LR + Decision Tree + Gradient Boosting
│   ├── predict.py              # Inference script — score new loan applications
│   └── utils.py                # Gini, KS, IV/WoE helper functions
│
├── data/
│   ├── loan_data.csv           # Raw synthetic loan portfolio (generated)
│   └── loans_scored.csv        # Enriched with PD scores + credit grades (generated)
│
├── notebooks/
│   └── EDA_CreditRisk.ipynb    # Exploratory data analysis notebook
│
├── tests/
│   └── test_pipeline.py        # Unit tests for data generation + model pipeline
│
├── docs/
│   └── credit_risk_concepts.md # IV, WoE, Gini, KS explained in plain English
│
├── credit_risk_dashboard.html  # Interactive browser dashboard (no server needed)
├── model_results.json          # Metrics, feature importances, IV table (generated)
├── requirements.txt            # Python dependencies
├── .gitignore                  # Standard Python gitignore
└── README.md

---

📊 Dataset — Synthetic Loan Portfolio

Property	Value
Total Loans	5,000
Default Rate	~33.7% (realistic for unsecured/MSME lending)
Loan Types	Personal, Home Improvement, Medical, Wedding, Education, Business, Vehicle
Employment Types	Salaried, Self-Employed, Business, Freelancer
Loan Range	₹20,000 – ₹20,00,000
Income Range	₹10,000 – ₹1,50,000/month
Credit Score Range	300 – 850

Features Engineered

Feature	Description	Credit Risk Rationale
foir_pct	Fixed Obligation-to-Income Ratio	EMI / Income — capacity to repay
ltv_proxy	Loan-to-Income ratio	Leverage indicator
credit_band	Bucketed credit score (0=best, 4=worst)	Bureau score risk tier
high_foir	Binary: FOIR > 55%	Industry threshold for over-leverage
high_util	Binary: Credit utilisation > 80%	Stress indicator
delinq_flag	Binary: Any 30-day past due	Behavioural default predictor
combined_risk	Sum of all binary risk flags	Composite alert score
log_income	Log-transformed monthly income	Corrects right-skewed distribution

---

🤖 Models & Credit Risk Metrics

Three models were trained and compared. Logistic Regression is the industry standard for regulatory-compliant scorecards; Gradient Boosting provides maximum predictive power.

Performance

Metric	Gradient Boosting	Logistic Regression	Decision Tree
ROC-AUC	0.7447	0.7607	0.6812
Gini Coefficient	0.4894	0.5213	0.3624
KS Statistic	0.3710	0.3948	0.2981
PR-AUC	0.631	0.644	0.541
5-Fold CV AUC	0.759 ± 0.009	0.758 ± 0.011	—

Why These Metrics?

• Gini Coefficient = 2 × AUC − 1. Industry benchmark: >0.3 = acceptable, >0.4 = good, >0.6 = excellent. Regulators (RBI, Basel III) explicitly require Gini reporting for PD models.
• KS Statistic = Maximum separation between cumulative Good and Bad distributions. >0.3 = deployable scorecard.
• Information Value (IV) = Measures each feature's predictive power. Used for feature selection in scorecards.

Information Value Table

Feature	IV Score	Predictive Strength
Delinquency (30-day past due)	0.7288	Very Strong (>0.5)
Existing Loans	0.2386	Medium (0.1–0.3)
Credit Score	0.1840	Medium (0.1–0.3)
FOIR %	0.0298	Weak (<0.1)
Loan Amount	0.0140	Weak (<0.1)

Credit Grade Distribution

Grade	PD Score Range	Count	Interpretation
A — Very Low Risk	0–20	1,180	Approve immediately
B — Low Risk	20–35	1,420	Approve with standard terms
C — Medium Risk	35–50	1,025	Approve with conditions
D — High Risk	50–65	622	Decline or high rate
E — Very High Risk	65–100	753	Decline

---

📊 Dashboard Features

• 8 KPI cards — Total loans, default rate, Gini, KS, AUC, CV AUC, high-risk count
• Credit grade distribution — Doughnut chart (A to E)
• Default rate by employment type — Salaried vs Freelancer vs Business
• Default rate by credit score band — Visual risk gradient
• Feature importance — Gradient Boosting top drivers
• Information Value table — IV scores with predictive strength labels
• Model comparison — GB vs LR with confusion matrix

---

🛠️ Tech Stack

Layer	Technology
Data Generation	Python · Pandas · NumPy
ML Models	scikit-learn — LogisticRegression, DecisionTreeClassifier, GradientBoostingClassifier
Credit Risk Metrics	Custom Gini, KS, IV/WoE functions (Basel III aligned)
Feature Engineering	FOIR, LTV proxy, WoE bins, log transforms, credit bands
Cross-Validation	StratifiedKFold (5-fold) — preserves default rate in each fold
Evaluation	ROC-AUC · Gini · KS Statistic · PR-AUC · IV
Dashboard	Chart.js · HTML/CSS · JetBrains Mono

---

▶️ How to Run

# 1. Clone the repository
git clone https://github.com/ukishore33/loan-default-prediction.git
cd loan-default-prediction

# 2. Install dependencies
pip install -r requirements.txt

# 3. Generate synthetic data
python src/generate_data.py

# 4. Train models and generate metrics
python src/train_model.py

# 5. Score a new loan application
python src/predict.py

# 6. Open interactive dashboard
open credit_risk_dashboard.html

# 7. Run tests
python -m pytest tests/

Live Demo

---

🗂️ Key Files Explained

File	Purpose
src/generate_data.py	Generates 5,000 synthetic loans with realistic default patterns
src/train_model.py	Full training pipeline — 3 models, all metrics, portfolio scoring
src/utils.py	Standalone Gini, KS, IV, WoE functions — reusable in any project
src/predict.py	Takes new loan data → outputs PD score + credit grade
model_results.json	All metrics, feature importances, IV table in structured JSON
credit_risk_dashboard.html	Interactive dashboard — open in any browser, no server
docs/credit_risk_concepts.md	Plain-English explanation of Gini, KS, IV for non-technical readers

---

💡 Interview Talking Points

"I built a credit risk scorecard that mirrors Basel III PD model requirements. I used Logistic Regression as the primary scorecard model — because it's interpretable and regulatory-compliant — and compared it against Gradient Boosting for predictive power. The Gini of 0.52 and KS of 0.39 both exceed industry thresholds for a deployable scorecard. I also computed Information Value for all features — delinquency history was the strongest predictor at IV = 0.73, which is Very Strong by the industry scale. The model assigns every loan a credit grade from A to E, directly mirroring what HDFC Bank and Bajaj Finance use in their credit decisioning systems."

---

👤 Author

Kishore U. AML/KYC Compliance Analyst | Credit Risk Analytics | Data Analytics 📱 6303308133 | Bengaluru, Karnataka | Immediate Joiner 🔗 LinkedIn · GitHub

Skills demonstrated: Credit Risk · Probability of Default · Scorecard Development · Gini · KS Statistic · Information Value · Python · scikit-learn · Gradient Boosting · Logistic Regression · Basel III PD Model Concepts

---

📜 Disclaimer

All data is 100% synthetic — generated programmatically. No real loan, customer, or financial data was used. Built purely for portfolio demonstration.