Home Loan Default Prediction (PROP-1006-HomeLoanDef)

Project Overview

This project aims to predict loan defaults by analyzing historical customer data from Home Credit. The goal is to identify key factors influencing loan repayment behavior and build a predictive model to classify eligible vs. ineligible borrowers. The project includes:

• Data Analysis: Comprehensive exploration, cleaning, and feature engineering across 7 datasets.
• Predictive Modeling: Development and comparison of multiple machine learning models (XGBoost, LightGBM, CatBoost, Random Forest).
• Model Deployment: Selection of the best-performing model for production use.

Key Tasks

1. Data Analysis Report: Assess data quality, handle missing values/outliers, and derive actionable insights.
2. Predictive Model: Identify customer segments eligible for loans using engineered features.
3. Model Comparison: Evaluate performance metrics (accuracy, recall, F1-score) to select the optimal model.
4. Challenges Report: Document technical hurdles and solutions during data processing and modeling.

---

Dataset Description

The data comprises 7 interconnected CSV files with loan application history, credit bureau records, and transactional behavior:

Dataset	Description	Rows	Key Features
application_train	Main dataset with loan status (Target: 1=Default, 0=Non-Default)	307,511	DAYS_BIRTH, AMT_CREDIT, EXT_SOURCE_1/2/3
bureau	Clients' previous credits from other institutions	1,716,428	DAYS_CREDIT, AMT_CREDIT_SUM
bureau_balance	Monthly balances of previous credits	27,299,925	MONTHS_BALANCE, STATUS
POS_CASH_balance	Monthly POS/cash loan snapshots	10,001,358	SK_DPD, CNT_INSTALMENT
credit_card_balance	Monthly credit card balance history	3,840,312	AMT_BALANCE, CREDIT_UTILIZATION_RATIO
previous_application	Clients' prior loan applications	1,670,214	AMT_ANNUITY, NAME_CONTRACT_TYPE
installments_payments	Repayment history for previous credits	13,605,401	AMT_INSTALMENT, DAYS_ENTRY_PAYMENT

Download Link: PRCP-1006-HomeLoanDef.zip

---

Project Structure

├── data/                        # Raw and processed data
│   ├── 1.1 raw/                # Original CSV files
│   └── 1.2 processed/          # Cleaned and aggregated data (final_data.csv)
├── docs/                       # Problem statement and dataset documentation
├── notebooks/                  # Jupyter notebooks for analysis and modeling
│   └── PRCP-1006-HomeLoanDef.ipynb  # Main workflow
├── reports/                    # Final reports (Markdown/PDF)
├── results/                    # Analysis outputs, figures, and saved models
│   ├── 365csv pre-anlysis/     # Statistical summaries and visualizations
│   ├── figures/                # Charts and graphs
│   └── models/                 # Serialized models (e.g., final_xgb_model.pkl)
├── scripts/                    # Utility functions and helper scripts
└── requirements.txt            # Python dependencies

---

Installation

1. Clone the Repository:

   git clone [https://github.com/dhaneshbb/HomeLoanDef](https://github.com/dhaneshbb/HomeLoanDef).git
   cd HomeLoanDef

2. Install Dependencies:

   pip install -r requirements.txt

3. Download Data:
   Extract PRCP-1006-HomeLoanDef.zip into data/1.1 raw/.

---

Usage

1. Run the Jupyter Notebook:

   • Execute notebooks/PRCP-1006-HomeLoanDef.ipynb for end-to-end analysis and modeling.
   • Key steps:
     • Data preprocessing (missing value imputation, outlier handling).
     • Feature engineering (e.g., CREDIT_UTILIZATION_RATIO, WEIGHTED_EXT_SOURCE).
     • Model training/evaluation and threshold optimization.

2. Generate Reports:

   • The notebook exports:
     • reports/Final Report.md: Detailed analysis and model comparison.
     • results/figures/: Visualizations (e.g., ROC curves, feature importance plots).

3. Use the Trained Model:
   Load the saved XGBoost model for predictions:

   import joblib
   model = joblib.load('results/models/final_xgb_model.pkl')

---

Results

Model Performance

Model	Accuracy	Recall	ROC AUC	Training Time (s)
XGBoost	83%	53%	0.785	98.10
LightGBM	74%	68%	0.787	88.67
CatBoost	79%	59%	0.776	147.94

Optimal Threshold: 0.60 (balances risk detection vs. false positives).

Key Insights

• Top Features: External credit scores (EXT_SOURCE), demographic factors, and credit utilization ratios.
• Business Impact: At a threshold of 0.60, the model detects 53% of defaulters while minimizing false positives.
• Full details: Final Report.

---

Challenges Faced

Challenge	Solution
High Memory Usage	Downcasted numerical dtypes
Class Imbalance (8% defaults)	Adjusted scale_pos_weight in XGBoost
Data Leakage	Removed post-application features (e.g., DAYS_LAST_PHONE_CHANGE)
Multicollinearity	VIF analysis + dropped correlated pairs (e.g., AMT_CREDIT vs AMT_GOODS_PRICE)

Full Challenges Report

---

References

1. Dataset Source: Home Credit Default Risk
2. Custom Library: insightfulpy
3. Technical Guides: Memory optimization, feature engineering, and model tuning best practices.

---

License

This project is licensed under the MIT License. See LICENSE for details.

This README provides a structured, overview of the project, linking to key files and highlighting critical insights. It balances technical detail with accessibility for collaborators.