emipredict-loan-default-xgboost

Loan default risk prediction on 404K records using XGBoost + SMOTE. Weighted F1 0.988, CV F1 0.88. SHAP explainability, MLflow experiment tracking, Streamlit deployment. Python · scikit-learn · XGBoost

💳 EMIPredict AI — Intelligent Financial Risk Assessment Platform

Predicts loan default risk using XGBoost + SMOTE on 404K records. F1 0.988 on test set. Fully explained via SHAP. Tracked with MLflow.

---

📌 Problem Statement

Banks and NBFCs need to assess whether a loan applicant is likely to default before disbursing funds. Manual credit review is slow and inconsistent. This project builds an automated, explainable ML pipeline to classify loan risk as Default or Non-Default.

📊 Key Results

Metric	Value
Dataset size	404,800 rows × 27 columns
Best model	XGBoost Classifier
Test F1 Score (weighted)	0.988
CV mean F1 (5-fold)	0.88
Class imbalance handled	SMOTE
Experiment tracking	MLflow
Explainability	SHAP summary plots

🛠️ Tech Stack

Python · XGBoost · scikit-learn · imbalanced-learn (SMOTE) · SHAP · MLflow · Streamlit · Pandas · Matplotlib · Seaborn

📁 Project Structure

EMIPredict_AI/
│
├── EMIPredict_AI.ipynb          ← Main notebook (run this)
├── emi_prediction_dataset.csv   ← Dataset (place here)
├── requirements.txt
└── README.md

▶️ How to Run

# 1. Clone or download this folder
# 2. Install dependencies
pip install -r requirements.txt

# 3. Place emi_prediction_dataset.csv in this folder
# 4. Open the notebook
jupyter notebook EMIPredict_AI.ipynb

# 5. Run All Cells (Kernel → Restart & Run All)

📂 Dataset

• Source: Kaggle — EMI Prediction Dataset
• Size: 404,800 rows, 27 columns
• Target: Loan default (binary classification)

🔍 Key Findings

• XGBoost with SMOTE outperformed Random Forest and Gradient Boosting
• Top default risk indicators (SHAP): Credit history length, loan amount, interest rate, income
• 5-fold CV confirms results are stable, not overfitted to one split
• All runs logged in MLflow for reproducibility

⚙️ Pipeline Steps

1. Exploratory Data Analysis (EDA)
2. Data cleaning and feature engineering
3. SMOTE for class imbalance
4. Benchmarking: Logistic Regression → Random Forest → Gradient Boosting → XGBoost
5. Hyperparameter tuning (5-fold CV)
6. SHAP explainability
7. MLflow experiment logging
8. Streamlit app generation

---

Project completed as part of AI/ML Internship at Labmentix Pvt. Ltd. (2025–2026)