Skip to content
Open access

Machine Learning Prediction of Accounting Manipulation Risk in MENA Listed Firms

Oct 2026 · Journal of Risk and Financial Management · 0 citations · 49 references

Abstract

This study examines the predictive performance of machine learning models in classifying accounting manipulation risk among listed non-financial firms in the Middle East and North Africa (MENA) using the Beneish M-Score. Drawing on 8978 firm-year observations from 1165 firms across 16 countries over 2013–2024, the study evaluates whether machine learning models capture information beyond the accounting variables embedded in the outcome definition. Eight supervised algorithms are estimated, including Logistic Regression, Random Forest, Extra Trees, XGBoost, LightGBM, CatBoost, Support Vector Machine, and Multilayer Perceptron, using a chronological train-validation-test framework. Predictor variables are classified by their relationship to the Beneish components, allowing model performance to be examined under progressively restrictive information sets. The results show substantial differences across predictor specifications. The full predictor set achieves a test ROC-AUC of 0.9024, whereas removing direct Beneish inputs reduces the best performance to 0.7965. Further removal of close accounting proxies reduces ROC-AUC to 0.6017. A forward-looking specification yields lower but economically meaningful discrimination, with ROC-AUCs of 0.7095 and 0.6863 for the full and restricted sets, respectively. Panel-econometric diagnostics further examine heterogeneity, cross-sectional dependence, and parameter stability. The findings indicate that reported predictive performance is strongly influenced by accounting information embedded in the Beneish classification rule and highlight the importance of leakage-aware evaluation in machine learning applications to accounting risk.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.