Machine Learning for Nonperforming Corporate Loan Recovery: A Systematic Review and Meta-Analysis of Predictive Performance

Abstract

Nonperforming corporate loans pose persistent challenges for financial institutions because recoveries depend on interacting borrower, collateral, contractual, and macroeconomic conditions, while published evidence on machine learning forecasting remains fragmented. This study examines whether machine learning improves recovery rate and loss given default prediction, identifies influential determinants, and assesses validation, explainability, and institutional variation. A qualitative systematic literature review used a cross-sectional evidence snapshot and comparative case-based synthesis of published corporate loan, corporate debt, and related credit recovery investigations. Published studies, rather than cloud enterprises or newly surveyed respondents, constituted the units of analysis. Evidence was organized across six results domains and five hypotheses using structured extraction of model classes, predictor characteristics, evaluation metrics, and validation procedures, followed by qualitative appraisal informed by Statistical Learning Theory. Six illustrative publications included comparative analyses of 12 machine learning methods, 20 candidate algorithms, 768 corporate bond observations, and 2,636 defaulted corporate bonds. One United States small-business investigation examined 89,903 loan records, including 39,943 defaults; its preferred hybrid model reported mean absolute error of 0.1835, root mean squared error of 0.3108, and R² of 0.4577. Another investigation reported up to 30% improvement in quantile forecasting precision, while chronological validation improved out-of-time stability. Recovery predictors included collateral, borrower characteristics, loan structure, collection activity, and economic conditions. Provisional five-point evidence ratings were 3 for machine learning superiority, 4 each for predictor relevance, validation, and explainability, and 3 for contextual variation. Findings suggest conditional predictive advantages rather than universal algorithmic superiority. Because recovery definitions, financial instruments, and validation schemes differed across source studies, no pooled effects or independent significance tests were produced. For banks and credit risk managers, the synthesis supports transparent, context-sensitive recovery models, disciplined temporal validation, and careful separation of forecast accuracy from realized collection outcomes across diverse lending institutions and changing macroeconomic and legal environments.