AI for financial fraud detection: Empirical evidence from benchmarking deep learning architectures to case studies

Financial statement fraud poses a significant threat to market integrity, but its detection across different institutional environments remains a major challenge. This study develops and validates a deep learning framework for cross-context financial fraud detection, leveraging a novel dataset of confirmed cases in Germany and a purposive sample of high-risk firms in Brazil. Our empirical investigation proceeds in two parts. First, we conduct a systematic benchmark of three neural network architectures—Feed-forward Neural Network (MLP), Convolutional Neural Networks (CNNs), and Gated Recurrent Units (GRUs)—trained on six key financial ratios from 2778 company-year observations. By learning from sequential financial data, the AI model directly addresses the fundamental problem of information asymmetry between corporate insiders and external stakeholders. The sequential GRU model demonstrated clear superiority, achieving a 94.3% Area Under the Curve (AUC) in the German market and maintaining robust performance with an 89.7% AUC in the high-risk Brazilian context. Second, a detailed case study analysis reveals a stark contrast in fraud prevalence and patterns: a 3.6% rate characterized by subtle, multi-year manipulation in Germany, versus a 41.7% probability marked by acute financial distress in the Brazilian sample. A key finding unifying both parts is the identification of the “profitability-efficiency mismatch"—a systematic divergence between accrual-based earnings and cash flow—as a primary, universal fraud signature. These findings provide regulators and auditors with a powerful, interpretable tool for proactive risk assessment and underscore the critical need for context-specific detection strategies in global markets.