From Static Labels to Event-Based Prediction: A Temporal Machine-Learning Framework for SME Business Dissolution in an Emerging Market

Abstract

Machine-learning models for SME failure prediction frequently report strong results that do not survive rigorous evaluation. Using a firm–year panel of 3796 Thai food-and-beverage SMEs (15,018 firm–years, 258 dissolution events, prevalence 1.72%) drawn from the Department of Business Development records for 2020–2024, this study develops a leakage-aware temporal machine-learning framework for predicting registered business dissolution within a 12-month horizon and decomposes two sources of performance inflation that the literature typically conflates. Entity leakage from ungrouped cross-validation inflates ROC-AUC by 0.085 and PR-AUC by 0.099—a pure protocol effect. Replacing static firm-level labels with event-based labels changes the estimand rather than the accuracy: prevalence falls from 5.79% to 1.72% and PR-AUC from 0.152 to 0.090 even as ROC-AUC rises. Under the corrected design, gradient boosting with temporal features attains ROC-AUC 0.833 and the strongest precision–recall performance, with ratio volatility and year-on-year change as the leading predictive dimensions. Isotonic recalibration reduces expected calibration error from 0.081 to 0.022, supporting risk ranking and budget-constrained triage, and a break-even analysis translates model output into deployment conditions for supervisory review budgets. The framework provides a transferable template for the honest evaluation of early-warning models on administrative panels in emerging markets.