Explainable Machine Learning for Predicting Late Payments in Enterprise Accounts Receivable Systems

Abstract

Late payment is a routine source of working-capital friction, yet an accounts-receivable model is useful only when its prediction time, feature availability, and explanation are aligned with the work of a collector. This paper develops an explainable dual-task framework for predicting whether an invoice will be paid late and, conditional on the information available at invoice issuance, estimating the number of days late. The intended benchmark is the public IBM Watson accounts-receivable sample, documented as 2,466 simulated invoices with 12 fields. Three post-outcome variables---settlement date, days to settle, and the late-days target---are explicitly blocked from the issuance-time feature set; dispute status is reserved for a later collection-stage sensitivity analysis. Customer-history variables are calculated from preceding invoices only. Logistic regression, random forests, histogram gradient boosting, XGBoost, and a transparent validation-weighted ensemble are compared for classification, while Ridge and three nonlinear regressors are compared for delay estimation. SHAP (Shapley Additive explanations) is used as a model-dependence diagnostic rather than a causal explanation. Because the historical IBM download could not be retrieved in the build environment, the archive does not manufacture IBM performance numbers: it includes an executable real-data mode and a deterministic 2,466-row schema-faithful surrogate used only to verify the pipeline. On that surrogate stress test, post-settlement leakage inflates XGBoost PR-AUC from 0.6949 to 0.9921, illustrating why timing discipline can matter more than algorithm choice. The principal contribution is therefore a reproducible, timing-aware and explanation-aware design that can be rerun unchanged when the original IBM CSV is supplied.