Credit Risk Model

Machine learning system simulating a bank's credit scoring decisions. Predicts whether a loan applicant will default or successfully repay. Built on Kaggle's Home Credit Default Risk dataset to replicate real-world credit scoring infrastructure.

🌐 Try the Application

Dataset & Methodology

Data: 307K loan applications, 7 related tables (bureau history, previous applications, payment records)

• Target: 8% default rate (imbalanced)
• Features: 122 initial → 95 engineered → 40 final

Pipeline:

1. Baseline models (Logistic Regression, LightGBM)
2. Incremental feature engineering (Bureau, Previous Apps, Installments, POS/CC)
3. Two-level filtering (statistical + importance-based)
4. Hyperparameter optimization (Optuna)
5. Model feature reduction (95→40 features)

Results:

• Validation AUC: 0.7610 (baseline) → 0.7780 (final) [+2.2%]
• Kaggle Score: Public 0.77284 | Private 0.77125

Documentation:

• Setup & Installation
• Data Overview - Dataset schema and relationships
• Baseline Models - Baseline modeling results
• Feature Engineering - Phase-by-phase feature creation
• Model Optimization - Hyperparameter tuning with Optuna
• API Deployment - FastAPI deployment guide

TODO

• Review Kaggle winning solutions writeups
• EDA on additional datasets
• Detailed feature engineering and model selection architecture
• Refactor repetitive code (MLflow, feature selection) into helpers
• Create final pipeline notebook after R&D phase
• Feature stability analysis (PSI monitoring)
• Business review of all features in final model
• Threshold optimization
• Scorecard conversion
• Build segmentation model for risk level tracking
• Limit calculation system design
• Add interest rate pricing to limit system
• Model monitoring dashboard (PSI-GINI)
• Add system design architecture to readme