Problem Statement
Financial institutions face significant risk when issuing loans. Accurately predicting whether a customer will default allows banks to minimize losses, optimize lending strategies, and improve risk management.
This project focuses on predicting the probability of default rather than a simple binary classification, enabling more flexible decision-making.
Dataset Overview
- train.csv: Customer data with target (default or not)
- test.csv: Same features without target
- submission.csv: Final predicted probabilities
The dataset includes financial, demographic, and credit-related features, totaling 30+ variables.
The target variable is highly imbalanced, with significantly fewer default cases compared to non-default cases.
Exploratory Data Analysis
- Analyzed class imbalance in default vs non-default
- Visualized key risk indicators
- Focused on high-impact features (EXT_SOURCE variables)
External risk scores (EXT_SOURCE_1, 2, 3) show strong separation between defaulters and non-defaulters, making them critical predictive features.
Feature Engineering
Raw financial data was transformed into meaningful indicators to better capture customer risk.
- Credit-to-Income Ratio
- Annuity-to-Income Ratio
- Income per Family Member
- Employment Duration Ratio
Behavioral Features:
- Credit Bureau Inquiry Count
- Document Submission Count
Additional interaction features were created from external risk scores to capture combined effects.
Modeling Approach
A LightGBM-based ensemble approach was used to improve prediction stability and performance.
- Model: LightGBM (Gradient Boosting)
- Validation: Stratified K-Fold (5 folds)
- Ensemble: Multiple random seeds
- Technique: Early stopping to prevent overfitting
Predictions from multiple folds and seeds were averaged to reduce variance and improve generalization.
Model Evaluation
The model was evaluated using ROC-AUC, which is suitable for imbalanced classification problems.
- Cross-validation with out-of-fold predictions
- ROC Curve analysis
The model achieved an approximate AUC score of 0.75, indicating strong ability to distinguish between defaulters and non-defaulters.
Results
Final predictions were generated by averaging outputs from 15 model runs (3 seeds × 5 folds).
- Stable and robust predictions
- Reduced variance across folds
- Output saved as submission.csv
The final output represents the probability of default for each customer, enabling flexible threshold-based decision making.
Business Impact
- Improves loan approval decision accuracy
- Reduces financial losses from defaults
- Enables risk-based pricing strategies
- Supports scalable credit risk assessment
This system can be integrated into financial pipelines to provide real-time credit risk scoring.
Back to Portfolio