Harnessing Machine Learning to Detect Financial Statement Fraud in Real Time

Abstract

Financial statement fraud remains difficult to detect because manipulators exploit accrual discretion, related-party structures, and delayed audit cycles. Classical ratio models such as the Beneish M-score and the Dechow F-score remain useful screens, but they use fixed coefficients estimated on older samples. This article reviews machine-learning approaches to financial statement fraud detection and specifies a real-time screening design suitable for a Nigerian college of education research project. Evidence from cross-country and market-specific studies shows that ensembles combining gradient boosting with distance-based classifiers, and models that reuse the inputs of the Altman, Beneish, Montier, and Dechow scores, can exceed the roughly 78% average accuracy reported across earlier algorithm comparisons. A 2025 cross-country ensemble reached 85.69% overall accuracy on a held-out set and flagged part of the Wirecard record retrospectively. The article argues that “real time” should mean continuous ingestion of interim filings, board packs, and bank-reconciliation exceptions, not a claim of perfect contemporaneous detection. Implications for auditors, audit committees, and teacher-education programmes in accounting are discussed. This paper does not report primary data collected at the college.