Financial fraud represents one of the most complex phenomena in capital markets, driven by multifaceted factors and diverse actors. This study aims to develop a dynamic model of financial reporting fraud from an ethical perspective by integrating evolutionary game theory and deep neural networks. Primary data were gathered using a researcher-designed questionnaire and the Delphi technique from three key stakeholder groups in Iran: corporate managers (n = 152), auditors (n = 123), and capital market regulator (n = 31). To mitigate uncertainty in expert judgments, the initial model parameters were fuzzified using Type-3 fuzzy logic in logistic mode. Subsequently, Bayesian calibration and Convolutional Neural Network (CNN) training were executed simultaneously within a feedback loop. The CNN architecture which demonstrated superior performance in capturing behavioral dynamics compared to Long Short-Term Memory (LSTM) and Recurrent Neural Networks (RNN) simulated player strategic interactions, whereby neural-network-generated game outputs updated the Bayesian posterior until convergence. Leveraging the calibrated parameters, the tripartite evolutionary game was simulated to identify evolutionary stable strategies (ESS) and conduct scenario analyses by varying ethical parameters. The findings indicate that the adoption of ethical policies exerts the most pronounced influence on audit firms, followed by listed companies and the regulator, respectively resulting in more rigorous auditing, a reduction in financial fraud, and enhanced regulatory oversight.