Ai-Driven Detection of Industry-Specific Financial Fraud: A Theory-Informed NLP Approach

This study applies natural language processing (NLP) and machine learning techniques to classify and analyze distinct types of financial fraud based on narrative disclosures from SEC Accounting and Auditing Enforcement Releases (AAERs). Moving beyond the binary classification common in fraud detection research, the study introduces a hybrid framework that integrates LDA-based topic modeling and supervised machine learning to identify five major fraud categories: financial misstatements, bribery, tax fraud, investment fraud, and related-party transactions. These fraud types are then mapped to specific industry sectors to reveal contextual patterns and vulnerabilities. The empirical results demonstrate that logistic regression offers the most stable and interpretable performance across classification tasks, while deep learning models like RNN+LSTM and biLSTM capture sequential and procedural aspects of fraudulent behavior. Chi-square tests and clustering analyses confirm significant industry–fraud associations, underscoring the importance of sector-specific detection strategies. By aligning classification outcomes with criminological theories—specifically Routine Activity Theory and Crime Script Analysis—the study provides a theoretically grounded, empirically validated contribution to fraud analytics, regulatory enforcement, and risk governance in financial markets. The findings offer practical applications for auditors, regulators, and compliance officers in developing targeted fraud detection systems tailored to industry-specific risks.