Kateryna Shcheglova

Type of the article: Research Article

Abstract
This study evaluates whether explainable machine-learning models can provide reliable and operationally interpretable short-horizon delinquency monitoring in U.S. auto lease securitization panels. Six public SEC ABS-EE trust-family panels were harmonized at the contract-month level. The primary outcome is one-month-ahead incident escalation to 30 or more days past due among contracts below 30 days past due at the feature month. Fully tuned penalized logistic regression (M1), unconstrained gradient boosting (M2), and governance-constrained gradient boosting (M3/X-LEASE) were assessed through chronological development, calibration, locked out-of-time testing, contract-held-out validation, six leave-one-issuer-out experiments, and later temporal evaluation. The locked test comprised 198,301 observations and 768 events. M2 produced the strongest discrimination (AUC-ROC 0.8151; PR-AUC 0.0269), while M3 exceeded the logistic benchmark by PR-AUC but did not outperform M2. At an exact 1% review capacity, each model generated 1,984 alerts; M2 detected 104 events, compared with 61 for M3, so the hypothesized recall advantage of X-LEASE was not supported. Full-sample TreeSHAP analysis showed stable feature rankings across tested issuers, later periods, and contract-level resampling. M3 relied more heavily on credit score and issuer controls and had a more concentrated explanation structure, but this does not establish superior auditability or fairness. The results support human-reviewed early-warning monitoring within the tested securitization panels, not lifetime default prediction, automated adverse decisions, or universal transferability.

Acknowledgments
The authors acknowledge the public availability of SEC EDGAR Form ABS-EE, Exhibit 102 asset-level data. No individuals or institutions outside the author team provided paid analytical, editorial, or funding support for this manuscript.