🏦 CreditRiskAI

End-to-end ML system to predict credit default probability — from raw data to a deployed REST API

Live Demo →  |  API Docs →  |  Notebook →

---

🎥 Demo Video

Demo.loan_prediction.mp4

Overview

Financial institutions lose billions annually to loan defaults. This project builds an end-to-end credit risk scoring system that estimates the probability a borrower will default, using gradient boosted trees trained on 32,000+ real loan records.

The system goes beyond a Jupyter notebook — it ships a FastAPI backend deployed on Railway and a Streamlit frontend on Streamlit Cloud, accepting live applicant data and returning a risk score in real time.

Dataset: Credit Risk Dataset – Kaggle  |  32,581 records · 12 features · 21.8% default rate

---

Results

Model	Precision	Recall	F1 Score	ROC-AUC
Logistic Regression	0.73	0.56	0.64	0.867
Random Forest	0.91	0.72	0.82	0.933
XGBoost	0.93	0.72	0.83	0.942

XGBoost selected as final model. Optimised for recall — in credit risk, a missed default (false negative) is costlier than a false alarm.

---

System Architecture

┌─────────────────────────────────────────────────────────┐
│                     Streamlit Cloud                      │
│   User fills applicant form → POST /predict              │
└──────────────────────┬──────────────────────────────────┘
                       │  JSON payload (12 features)
                       ▼
┌─────────────────────────────────────────────────────────┐
│                  FastAPI  (Railway)                       │
│                                                          │
│   Raw Input → ColumnTransformer → XGBoost.predict_proba  │
│                                                          │
│   Returns: { risk_category, default_probability }        │
└─────────────────────────────────────────────────────────┘

---

ML Pipeline

1. Data Cleaning

• Removed logical impossibilities: rows where person_emp_length > person_age
• Dropped rows with simultaneous nulls in loan_int_rate and person_emp_length
• Removed outlier: single row with loan_int_rate > 20 + missing employment (data error)

2. Feature Engineering

Feature	Type	Description
person_emp_length_missing	Binary flag	Missing employment length is itself predictive — borrowers with unknown employment default at higher rates
loan_percent_income	Ratio	loan_amnt / person_income — debt-to-income proxy

3. Imputation Strategy

• person_emp_length nulls → filled with median
• loan_int_rate nulls → filled with within-grade median (preserves grade-level interest rate signal)

4. Preprocessing

• OneHotEncoding on 4 categorical columns: person_home_ownership, loan_intent, loan_grade, cb_person_default_on_file
• StandardScaler applied only to Logistic Regression (tree models are scale-invariant)
• Preprocessing saved as preprocessor.pkl via ColumnTransformer for consistent inference

5. Key EDA Findings

• Loan grade is the strongest default predictor — Grade G default rate ~4× Grade A
• Loan-to-income ratio separates defaulters cleanly (defaulters avg 0.38 vs 0.16)
• RENT ownership defaults more than MORTGAGE or OWN
• Debt consolidation and medical loan intents have highest default rates
• Missing emp_length → higher default rate (informative missingness)

6. SHAP Explainability

Top features by SHAP importance:

1. loan_int_rate — higher rate = higher risk (also a proxy for perceived risk)
2. loan_percent_income — debt burden relative to income
3. loan_grade — lender's internal risk rating
4. person_home_ownership — financial stability signal
5. person_emp_length_missing — informative missingness flag

Project Workflow

flowchart TD
    A[Loan Application Data]
    --> B[Data Preprocessing]

    B --> C[Feature Engineering]

    C --> D[XGBoost Model]

    D --> E[Default Probability]

    E --> F[Risk Classification]

    F --> G[FastAPI Backend]

    G --> H[Streamlit Dashboard]

Loading

---

Project Structure

loan-default-prediction/
├── notebooks/
│   └── loan_default_prediction.ipynb   # EDA, feature engineering, training, SHAP
├── models/
│   ├── xgboost_model.pkl               # Trained XGBoost classifier
│   └── preprocessor.pkl                # Fitted ColumnTransformer
├── Fast_api_app.py                     # FastAPI prediction endpoint
├── streamlit_app.py                    # Streamlit frontend
└── requirements.txt

---

Project Features

• XGBoost Machine Learning Model
• Feature Engineering Pipeline
• SHAP Explainable AI
• FastAPI Backend
• Streamlit Frontend
• Probability-Based Risk Assessment
• Real-Time Predictions
• Railway Deployment
• Production-Ready Inference Pipeline

---

API Reference

POST /predict — Returns default probability for a loan applicant

Sample Request:

{
  "person_age": 28,
  "person_income": 60000,
  "person_home_ownership": "RENT",
  "person_emp_length": 3.0,
  "loan_intent": "PERSONAL",
  "loan_grade": "B",
  "loan_amnt": 10000,
  "loan_int_rate": 12.5,
  "loan_percent_income": 0.17,
  "cb_person_default_on_file": "N",
  "cb_person_cred_hist_length": 4.0,
  "person_emp_length_missing": 0
}

Sample Response:

{
  "risk_category": "Low Risk",
  "default_probability": 0.1342
}

GET / — Health check

---

Tech Stack

Layer	Tool	Purpose
ML Model	XGBoost 3.2	Gradient boosted classifier
Preprocessing	scikit-learn ColumnTransformer	OHE + scaling pipeline
Explainability	SHAP TreeExplainer	Feature importance + waterfall plots
API	FastAPI + Uvicorn	REST inference endpoint
Frontend	Streamlit	Interactive prediction UI
API Deployment	Railway	Cloud hosting for FastAPI
UI Deployment	Streamlit Cloud	Cloud hosting for frontend
Serialization	joblib	Model + preprocessor persistence

---

Author

Vipul Singh

LinkedIn