PORTFOLIO CREDIT RISK MODELING: FROM THEORY TO EMPIRICAL APPLICATION

This Master's thesis provides a comparative analysis of portfolio credit risk modeling, tracing the methodological evolution from traditional parametric frameworks to Data-Driven architectures. Accurately estimating systemic tail risk is vital for financial stability; however, the 2008 crisis exposed how classical models—constrained by rigid assumptions of symmetry and linear correlation—fail to capture extreme default clustering during severe economic downturns. A systematic literature review examines Structural Models, Reduced-Form models, and Actuarial approaches. The focus then shifts to Copula-based dependency, illustrating the limitations of the Gaussian Copula and exploring advanced methodologies for asymmetric tail dependence, such as the Clayton Copula and stochastic recovery rates (Hull & White). To overcome the constraints of closed-form equations, the research introduces Machine Learning paradigms (XGBoost, LSTM), addressing their black-box nature through Explainable AI (SHAP values) to ensure regulatory auditability. Empirically validated on 593 U.S. companies, the study contrasts individual default probabilities from Merton's structural model with those estimated by XGBoost. Findings reveal that the structural benchmark exhibits severe probability polarization driven by financial leverage, often classifying averagely capitalized firms as entirely risk-free. Conversely, XGBoost simultaneously evaluates 25 financial indicators, distributing risk in a continuous and granular manner to uncover latent vulnerabilities. Monte Carlo simulations across six quantitative frameworks demonstrate that under asymmetric stress (Clayton Copula with stochastic LGD), the XGBoost-initialized portfolio captures a massive "double-blow" effect. The lethal combination of simultaneous defaults and collateral fire-sales generates VaR and Expected Shortfall (ES) metrics that drastically exceed traditional baselines, whereas the CreditRisk+ actuarial framework yields highly mitigated tail losses due to its inter-sectoral independence assumption. Finally, sensitivity analysis demonstrates that extreme asymmetric tail dependence structurally overwhelms the impact of individual portfolio concentration (Wrong-Way Risk). Ultimately, the study provides critical insights for the Basel III/IV macroprudential framework, proving that the integration of Machine Learning and asymmetric dependence structures is essential to prevent the underestimation of regulatory capital requirements.