CREW: Credit Risk Early Warning System. Banking Data Science, XGBoost, Agentic AI & Explainable AI

CREW: Credit Risk Early Warning System | Banking Data Science Portfolio
An end-to-end credit risk system built with XGBoost, dual explainability (SHAP + LIME), isotonic calibration, and an autonomous multi-agent surveillance system using the ReAct framework, deployed as a live Streamlit dashboard.
The business context:

Credit risk deficiencies account for 40% of all ECB supervisory findings (ECB SREP, 2025), driven by monitoring failures and outdated data
Basel IV tightens capital requirements for IRB models, increasing demand for transparent, auditable risk systems
IFRS 9 requires forward-looking Expected Credit Loss estimation with defensible model logic

Key results:

AUC-ROC 0.7778 | Gini 0.5556 on 307,511 Home Credit loan applications
Business-optimal threshold of 0.79 derived from real loan economics (Verbraken et al., 2014)
Statistical threshold baseline established via Youden's J (Fluss et al., 2005)
Dual explainability: SHAP global feature importance + LIME local case-level explanations (Ribeiro et al., 2016)
Autonomous AI agent monitors model drift (PSI, Yurdakul & Naranjo, 2020), triggers alerts, and generates audit-ready reports using the ReAct framework (Yao et al., 2023)
XGBoost selected via gradient boosting benchmarks for sparse, mixed-type banking data with high class imbalance (Bequé et al., 2017)
Regulatory-aligned with Basel III/IV, IFRS 9, SR 11-7, GDPR Art. 22, EU AI Act, ECOA/Reg B, and FINMA Circular 2023/1

What this video covers:

The business case: credit risk monitoring failures in the current supervisory landscape
Model development: XGBoost pipeline with calibration, threshold optimization, and drift detection
Live dashboard demo: 5 tabs serving CEOs, Data Scientists, and Regulators
Explainability deep dive: SHAP global analysis and LIME individual case explanations
Agentic AI surveillance: autonomous ReAct agent monitoring model health in production
Academic foundation (MSc Data Science & AI, University of Liverpool) mapped to production banking requirements

Links:
Dashboard: https://credit-risk-early-warning-system.streamlit.app
GitHub: https://github.com/JuanCRuizA
LinkedIn: https://linkedin.com/in/juancarlosruizarteaga72
Built by Juan Carlos Ruiz Arteaga | Banking Data Scientist | 2026

Видео CREW: Credit Risk Early Warning System. Banking Data Science, XGBoost, Agentic AI & Explainable AI канала JC Ruiz - Banking Data Scientist