UNMASKING CREATIVE ACCOUNTING USING MACHINE LEARNING MODELS A CASE STUDY OF SONATRACH COMPANY, ALGERIA (2017-2022)

Abstract

This research employs integrated Beneish M-Score and machine learning methodologies to detect creative accounting at Sonatrach (Algeria's state-owned energy enterprise) during 2017-2022. All six years exceeded the -2.22-fraud risk threshold, with 2020 reaching peak risk (M-Score: -1.835; XGBoost probability: 0.89) during the oil price collapse. XGBoost achieved 93.7% accuracy, substantially outperforming traditional 76% baseline detection. Accrual management (TATA: 0.28 importance) and receivables manipulation (DSRI: 0.21) emerged as primary manipulation channels. Convergent evidence from traditional and machine learning methods strengthens confidence that state-owned enterprises in commodity-dependent sectors face endemic manipulation incentives driven by multi-principal agency conflicts and political cost pressures. Findings advance fraud detection theory, validate hybrid traditional-ML approaches, and provide auditors with mechanism-specific guidance for enhanced oversight in emerging market energy SOEs.