Interpreting Machine Learning Models of Corporate Default Predicition - A Forensic Triage of the Italian Distressed Cohort

Post-pandemic monetary normalization has exposed structural fragilities in European SMEs, creating a "Refinancing Cliff" that renders linear credit scoring obsolete. This thesis departs from standard default prediction by isolating a "Distressed Cohort" of Italian firms with negative equity. The objective shifts from generic risk screening to a "Solvency Restoration" triage, distinguishing between turnaround-capable firms and "Zombie" entities destined for liquidation. Adopting a "Domain-Aware" framework, the study employs a Cost-Sensitive XGBoost architecture. Calibrated via Basel III parameters (LGD and NIM), the model minimizes Total Expected Loss to act as a conservative gatekeeper against capital destruction. The analysis introduces a "Hybrid Target Variable" encoding "Technical Failure" alongside legal bankruptcy to mitigate survivorship bias. To resolve the regulatory "Black Box" dilemma, the research integrates SHAP and ALE within the DALEX framework. This approach decodes non-linear "Cliff Effects" and financial dynamics like the "Scissor Effect" of debt. Empirical results show the ensemble model significantly outperforms linear benchmarks in identifying "Silent Killers", firms with low leverage but deep structural insolvency. Finally, the study operationalizes a "Schivardi Swap" simulation, identifying € 13 billion in "Trapped Capital." Surgical reallocation to resilient entities is estimated to generate a € 2 billion net EBITDA delta, offering a roadmap for credit portfolio "sanitization."