Predictive Gains and Governance Tradeoffs in Socioeconomic Automobile Repossession Modeling

Abstract

Automobile repossession is a resource-constrained process in which lenders, forwarders, and recovery agencies must prioritize delinquent vehicle assignments with limited operational capacity. Prior work shows that assignment-level operational data can support repossession prediction, but less is known about whether geographic and socioeconomic augmentation changes how model-guided prioritization is distributed across socioeconomic strata. This study uses the Automobile Repossession Dataset (ARD), which contains 203,436 repossession assignments from June 2021 through December 2024, and links assignment records to ZIP/ZCTA-level socioeconomic indicators from the American Community Survey. The analysis compares an ARD-only CatBoost baseline with socioeconomic-augmented models and ZIP sensitivity models that retain raw ZIP code as a categorical predictor. Rather than evaluating predictive performance alone, the study introduces a decile-based socioeconomic audit of top-10% model selections, including selection concentration, threshold-specific performance concentration, and comparisons across income, poverty, SNAP participation, and unemployment deciles. Results show modest improvements in AUC-ROC, AUPRC, and lift score across the augmented and ZIP sensitivity conditions relative to the ARD-only baseline. The baseline-to-augmentation performance pattern is also preserved in a separate distribution drift robustness test that trains models on 2021–2023 assignments and evaluates them on 2024 assignments. However, these gains are accompanied by larger increases in selection concentration, especially for unemployment, Supplemental Nutrition Assistance Program (SNAP) participation, poverty, and income-related audit variables. The findings identify a performance–governance trade-off: geographic and ZIP/ZCTA-level socioeconomic inputs can improve ranking performance, but they can also concentrate top-priority operational attention across socioeconomic strata. Socioeconomic augmentation in repossession modeling should therefore be treated as a governed decision-support design choice rather than a routine feature-engineering addition.