Financial Audit Assistance using Misinformation Detection and Explanation

Financial auditing is a high-stakes, knowledge-intensive process aimed at ensuring the integrity and accuracy of financial statements (FS). Because companies have incentives to manipulate financial data to boost investor confidence or reduce tax liabilities, auditors must manually scrutinize vast amounts of data to detect potential misinformation. This paper introduces an AI-assisted framework designed to automate the detection of misinformation in balance sheets, income statements, and cash-flow statements, while providing actionable explanations to guide auditors.

The researchers utilize a large-scale dataset of 11,460 financial statements from 2,292 companies over a five-year period. The system operates through three primary components:

1. Misinformation Detection: Using unsupervised anomaly detection techniques, the system flags financial statements that deviate significantly from expected patterns.
2. Silver Label Generation: Since ground-truth labels for fraud are rarely available, the authors use generative language models (such as Mistral-7B-Instruct) to analyze audit reports and automatically assign 'silver labels' to statements containing adverse remarks.
3. Explanation Generation: The system maps detected anomalies to specific financial variables using XBRL (eXtensible Business Reporting Language) categories. This allows the system to provide a human-readable explanation of which financial metrics contributed to a statement being flagged as anomalous.

Manual auditing is time-consuming and prone to human error or bias. By integrating data from multiple financial statements and leveraging past audit reports, this system provides a scalable, objective tool that helps auditors prioritize their efforts. Instead of reviewing every document manually, auditors can focus on the specific variables and business processes highlighted by the AI, potentially increasing the efficiency and quality of the audit process while reducing the risk of missing material misstatements.