Intelligent Credit Risk Scoring System

Google GenAI Capstone — Milestone 1 (Mid-Semester Submission)

---

Problem Statement

Predict borrower creditworthiness and default probability from historical loan data using supervised machine learning, without LLMs or Agentic AI.

---

Project Structure

Credit_Risk-_System/
├── Data/
│   └── loan_credit_data.csv       # 45,000 borrower records
├── src/
│   ├── preprocessing.py           # Data cleaning, encoding, scaling, splitting
│   ├── train_logistic_regression.py  # Standalone LR model training
│   ├── train_decision_tree.py     # Standalone DT model training
│   ├── training.py                # Orchestrator to run both models
│   └── app.py                     # Streamlit web application
├── models/                        # Auto-generated: saved models & plots
├── .gitignore
├── requirements.txt
└── README.md

---

Dataset Overview

Property	Value
Rows	45,000
Features	14 columns
Target	loan_status (0 = no default, 1 = default)
Missing values	None
Class distribution	78% Non-Default / 22% Default

---

Setup Instructions

1. Clone the repository

git clone <your-repo-url>
cd Credit_Risk-_System

2. Install dependencies

pip install -r requirements.txt

3. Run Preprocessing

python src/preprocessing.py

Outputs: scaled train/test arrays + scaler.pkl in models/

4. Train Models

python src/training.py

Outputs: logistic_regression.pkl, decision_tree.pkl, metrics.json, confusion matrix & feature plots in models/

5. Launch Streamlit App

streamlit run src/app.py

Opens at http://localhost:8501

---

Models Used

Model	Role	Key Parameters	Accuracy	ROC-AUC
Logistic Regression	Primary	max_iter=1000	89.7%	0.953
Decision Tree	Secondary	max_depth=12, min_samples_leaf=15	92.3%	0.966

Both models are evaluated on Accuracy, ROC-AUC, Classification Report, and Confusion Matrix.

---

UI Features (Streamlit)

Tab	Description
Single Borrower	Input borrower details via sliders/dropdowns → get risk score, probability & badge
Batch CSV Upload	Upload CSV → predict all rows → download results
Model Performance	View accuracy, ROC-AUC, confusion matrices, feature plots

---

Evaluation Metrics

• Accuracy — Overall correct predictions
• ROC-AUC — Ability to discriminate defaults from non-defaults
• Confusion Matrix — TP, FP, TN, FN breakdown
• Classification Report — Precision, Recall, F1-score per class

---

Key Design Decisions

• Outlier removal: age capped at 80, employment experience at 50 yrs, income at 99th percentile
• No data leakage: scaler fitted only on training set, applied to test set
• Stratified split: maintains class ratio across train and test
• Prediction Consistency: Kept class weights default so Logistic Regression maintains calibrated probability outputs compared to Decision Tree.

---

Team

Google GenAI Capstone Project — Milestone 1 (Mid-Semester)

---

• No LLMs, Agentic AI, or external APIs used in this milestone.*