Machine Learning for Italian Non-Performing Loan Valuation: A Methodology Framework with Purged Cross-Validation

Version 2 (2026-08-19). This is a methodology-protocol contribution, not an empirical study of Italian NPL portfolios, and this revision makes that boundary enforceable rather than merely stated. The deposited version calibrated its synthetic generator's recovery marginals to published aggregates; here they are held out against supervisory data, and the generator is corrected in four places the deposited version disclosed but did not repair. The correction that matters most: the loan-to-value feature had no causal path to recovery at all in the generator, so the protocol was being demonstrated on a dataset in which its most economically central variable was inert. With that path restored and the missing closure-channel dimension added, the held-out recovery-level error falls from 11.89 to 4.40 percentage points; a closure-timing layer that was refuted out of sample by 42.86 points is re-anchored, and a noise constant becomes a reported sensitivity rather than a hidden choice. Two things are stated plainly because they bound what any reader may conclude. First, a feature-DGP overlap circularity: the 24 features in the recovery model overlap one-to-one with the synthetic signal components, so the reported recovery R-squared measures the protocol's ability to recover wiring the author installed. Second, the SHAP ranking is unstable across generator revisions - the deposited version reported is_secured, region_efficiency and guarantee_type as the top signals, this one reports collateral_gbv_ratio, region_efficiency and vintage_years - which is itself the reason to read SHAP on synthetic data as DGP wiring rather than as economics. Section 6 is new and reports the held-out test; Section 7 is a numbered erratum listing every changed number. The real-data benchmark is not vendored: its licence does not permit redistribution, so a fetch script with a pinned SHA-256 ships instead. This is the first code deposit for this paper. Two statements in the version-1 description below are no longer asserted. Version 1 reports the recovery model at "R^2 = 0.73 OOT"; the corrected generator gives 0.533 out of time (0.529 in cross-validation), and the 0.73 figure is superseded. Version 1 also reports "Median recovery time North: 25 months vs South: 70 months"; that pair is withdrawn in full - the per-district duration constants were a modelling choice, not a Ministero della Giustizia measurement, and neither the manuscript nor the code attributes them to one any longer. The version-1 text is left standing below, unedited, as the record of what version 1 said. Version 1 description, retained unedited: Develops an integrated machine-learning framework for Italian NPL valuation, addressing three methodological gaps: (1) cross-validation leakage via purged CV with embargo periods; (2) high-cardinality categorical features (26 judicial districts) via LightGBM native categorical handling; (3) regulatory interpretability via SHAP global+local explanations. Six-model framework: AVM (R^2 = 0.50 OOT), recovery rate (R^2 = 0.73 OOT), Weibull recovery timing, share-of-wallet, unsecured-credit, and Nash-bargaining ZOPA. Synthetic dataset of 100,000 loans calibrated to Banca d'Italia, OMI, and Ministero della Giustizia aggregates. Median recovery time North: 25 months vs South: 70 months. Synthetic-data disclosure: results illustrative of the methodology; not empirical findings about actual portfolios. Path to real-data validation through Italian NPL servicer partnership documented. Clarification (2026-06): The synthetic-data disclosure and all headline metrics in this deposit are unchanged and accurate. The manuscript's reference to an "accompanying replication package" denotes the six valuation models (AVM, recovery-rate, share-of-wallet, unsecured, ZOPA, pricing), whose implementations reside in the author's CreditAI codebase; a standalone packaged release will be linked in a future version. Calibration targets are reconstructed from published Banca d'Italia, OMI, and Ministero della Giustizia aggregates.