🏦 Loan Default Predictor

Predict whether a LendingClub borrower will fully repay or charge off their loan — powered by 5 ML models with a Streamlit web app.

🧠 Tech Stack & ML Models

Scikit-learn • ONNX • Streamlit • Decision Tree • Random Forest • AdaBoost • Gradient Boosting • XGBoost

---

📌 Problem Statement

LendingClub is the world's largest peer-to-peer lending platform. When a borrower applies for a loan, the platform must assess the risk of default — will this person repay or charge off?

This project builds an end-to-end ML pipeline that:

• Analyses 396,000+ historical LendingClub loans
• Engineers features from 27 raw columns
• Trains 5 different ML models and compares their performance
• Deploys the best models in a Streamlit web app for instant risk assessment

---

🖥️ Live Demo

→ Open App on Streamlit Cloud

---

📁 Project Structure

loan-default-predictor/
│
├── loan-defaulters-prediction.ipynb   ← Full EDA + preprocessing + model training
│
├── app.py                             ← Streamlit web application
├── dt_preprocessor.pkl                ← Fitted preprocessing pipeline
│
├── dt_model.onnx                      ← Decision Tree (ONNX)
├── rf_model.onnx                      ← Random Forest (ONNX)
├── ada_model.onnx                     ← AdaBoost (ONNX)
├── gb_model.onnx                      ← Gradient Boosting (ONNX)
├── xgb_model.json                     ← XGBoost (native format)
│
└── requirements.txt

---

📊 Dataset

Property	Value
Source	Kaggle — Lending Club Dataset
Rows	396,030 loans
Features	27 columns
Target	loan_status — Fully Paid / Charged Off
Class split	~80% Fully Paid · ~20% Charged Off

Key Features Used

Feature	Description
loan_amnt	Loan amount applied for ($)
int_rate	Interest rate on the loan
grade / sub_grade	LendingClub assigned risk grade (A–G)
emp_length	Borrower employment duration
annual_inc	Self-reported annual income
dti	Debt-to-income ratio
revol_util	Revolving line utilization rate
purpose	Stated reason for the loan
open_acc / total_acc	Number of open / total credit lines
pub_rec	Number of derogatory public records

---

🔍 Exploratory Data Analysis

Key findings from the notebook:

• Interest rate is the strongest predictor — charged-off loans cluster at higher rates
• Grade F & G loans default significantly more than A & B loans
• Higher DTI correlates with default, especially above 20
• Debt consolidation is the most common loan purpose across both classes
• Employment length has weak predictive power — 10+ year employees still default

---

⚙️ Preprocessing Pipeline

Raw CSV (396k rows, 27 cols)
        ↓
Missing value imputation (mean for numeric, mode for categorical)
        ↓
Feature encoding:
  • term           → integer (36 / 60)
  • grade          → ordinal (A=0 … G=6)
  • sub_grade      → ordinal (A1=0 … G5=34)
  • emp_length     → ordinal (< 1 year=0 … 10+=10)
  • emp_title      → top-20 label encoded, rest → "Other"
  • issue_d        → split into issue_year + issue_month
  • earliest_cr_line → earliest_cr_year
  • pub_rec / mort_acc / pub_rec_bankruptcies → binarised
  • verification_status / home_ownership
    application_type / purpose → label encoded
        ↓
Train / Test split (67% / 33%)

---

🤖 Models & Results

All models trained on the same preprocessed split. Metrics on the test set (130k rows):

Model	Accuracy	Precision	Recall	F1	AUC
Decision Tree	80.4%	0.755	0.804	0.728	0.695
AdaBoost	80.5%	0.765	0.805	0.733	0.720
Random Forest	80.5%	0.764	0.805	0.729	0.715
Gradient Boosting	80.7%	0.766	0.807	0.748	0.729
XGBoost	80.7%	0.767	0.807	0.747	0.731

Why ONNX for deployment?

Streamlit Cloud has a ~1 GB RAM limit. Large ensemble models (Random Forest at ~400 MB pkl, Gradient Boosting at ~200 MB pkl) crash the app. Exporting to ONNX reduces file sizes by 5–10×, making all models deployable.

Model	pkl size	ONNX / JSON size
Decision Tree	~1 MB	~0.5 MB
AdaBoost	~50 MB	~8 MB
Random Forest	~400 MB	~50 MB
Gradient Boosting	~200 MB	~30 MB
XGBoost	~100 MB	~15 MB (.json)

---

🚀 Run Locally

# 1. Clone the repo
git clone <repo-url>
cd Loan-Default-Prediction-System

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

# 3. Launch the app
streamlit run app.py

The app loads the preprocessor from dt_preprocessor.pkl and whichever model you select from the sidebar.

---

🗺️ Future Improvements

• SHAP explainability — show which features pushed each prediction toward default or repayment, making the model interpretable for end users
• Threshold tuning — add a slider to adjust the decision boundary (0.5 default) so lenders can tune precision vs recall based on their risk appetite
• Batch prediction — upload a CSV of multiple applicants and download a results file with predictions and probabilities for each row
• Hyperparameter tuning — use Optuna or GridSearchCV to squeeze additional performance out of XGBoost and Gradient Boosting
• LightGBM model — benchmark against XGBoost; LightGBM is often faster to train and competitive in accuracy on tabular data
• Class imbalance handling — experiment with SMOTE oversampling or class_weight='balanced' to improve recall on the minority (default) class
• Model monitoring — track prediction drift over time if the app receives real loan applications

---