Detection and Prediction of Financial Fraud Using Deep Learning Methods: A case of the Companies Listed in the Amman Stock Exchange

Introduction: The study examined the ongoing issue of identifying financial fraud in emerging economies, concentrating on companies listed on the Amman Stock Exchange (ASE). Methods: A panel of 176 ASE-listed enterprises was studied from 2011 to 2021. Starting with a preliminary analysis of Beneish M-Score constituents and associated metrics, a supervised neural network (FNN) had been trained, and an ordinary least-squares (OLS) analysis was computed. The performance study was executed using reliability, recall, reliability, F1-score, and ROC-AUC. Results: The FNN achieved an accurate identification rate of 0.9844 with a recall of 1.0, indicating it accurately identified all fraudulent transactions in the experimental dataset. The ROC-AUC was 0.97. The OLS model, albeit less precise, demonstrated statistically significant correlations—particularly for GMI, SGAI, and LVGI—with the Beneish M-Score, thereby providing interpretable risk indicators. Conclusions: The study revealed that deep learning, namely a feedforward neural network (FNN), surpassed a traditional ordinary least squares (OLS) method in detecting fraud among ASE enterprises, whereas OLS offered contextual information about the factors associated with fraud. An integrated analytical framework was proposed to assist regulators and investors in achieving improved transparency and early warning in the Jordanian market.