Financial Statement Fraud Detection via Large Language Models

ABSTRACT With the widespread adoption of Internet‐based AI technologies, addressing financial fraud has become increasingly critical, particularly within the realm of machine learning. In this case, deep learning and natural language processing (NLP) techniques offer powerful means of detecting fraudulent activity by analyzing financial documents, thereby enhancing both the efficiency and precision of such assessments and supporting financial security. In this study, we introduce deep representation learning‐based approaches relying mainly on large language models (LLMs) for identifying fraud in financial statements by examining temporal changes in the Management Discussion and Analysis (MD&A) sections of corporate disclosures. Departing from conventional techniques that rely only on word frequency analysis, we propose D eep F raud that combines time‐evolving financial LLM embeddings, such as FinBERT, FinLlama, and FinGPT embeddings, of paragraphs and uses long short‐term memory (LSTM) to predict frauds via historical textual embeddings. In addition to LLM embeddings, we also integrate (1) time‐evolving word frequencies of words relevant to fraud detection, such as those expressing sentiment or uncertainty, and (2) time‐evolving financial ratios. Trajectories of paragraph‐level embeddings, frequencies, and ratios are used to construct a fraud detection model, which we evaluate against machine learning methods and deep time‐series models. Using 30 years of financial report data (from 1995 to 2024), our experiments demonstrate that D eep F raud on average enhances fraud detection performance across a number of scenarios and on average outperforms the competing approaches as well as conventional word frequency approaches. Our framework introduces a novel direction for deep feature engineering in the field of financial statement fraud detection.