Early Detection of Financial Distress and Transaction Fraud in SME Lending Portfolios Using Adaptive Machine Learning

Abstract

Small and Medium Enterprises (SMEs) form the backbone of most emerging economies, yet lending institutions continue to struggle with two interlinked risks when serving this segment: the gradual build-up of financial distress in a borrower's cash-flow behaviour, and the presence of anomalous or fraudulent transaction activity within the borrower's account. Traditional credit-scoring frameworks, which rely on static, periodically-updated financial statements, are often too slow to flag these risks before they materialise into non-performing exposure. This article proposes an adaptive machine learning framework that continuously monitors transaction-level data to generate early-warning signals of both financial distress and transaction fraud in SME lending portfolios. Building on a supervised learning methodology, the study benchmarks Logistic Regression against an ensemble Random Forest classifier on a large, highly imbalanced transaction dataset that serves as an analytical proxy for SME account activity, given the scarcity of publicly available labelled SME lending data. Exploratory analysis is used to identify behavioural and balance-based indicators that separate distressed/fraudulent accounts from healthy ones, and these indicators are subsequently engineered into predictive features. The results show that while Logistic Regression provides an interpretable baseline (recall of approximately 51%, precision above 90%), the Random Forest classifier substantially outperforms it, achieving recall and precision above 99% on held-out test data, with consistent performance across cross-validation folds, indicating no overfitting. The findings suggest that tree-based ensemble methods, combined with carefully engineered balance-discrepancy and time-based features, offer a practical and scalable route for lenders to build adaptive early-warning systems for SME credit portfolios.