Mining corporate annual reports for intelligent detection of financial statement fraud – A comparative study of machine learning methods

Abstract

Financial statement fraud has been serious concern for investors, audit firms, government regulators, and other capital market stakeholders. Intelligent financial statement fraud detection systems have therefore been developed to support decision-making of the stakeholders. Fraudulent misrepresentation of financial statements in managerial comments has been noticed in recent studies. As such, the purpose of this study was to examine whether an improved financial fraud detection system could be developed by combining specific features derived from financial information and managerial comments in corporate annual reports. To develop this system, we employed both intelligent feature selection and classification using a wide range of machine learning methods. We found that ensemble methods outperformed the remaining methods in terms of true positive rate (fraudulent firms correctly classified as fraudulent). In contrast, Bayesian belief networks (BBN) performed best on non-fraudulent firms (true negative rate). This finding is important because interpretable ``green flag” values (for which fraud is likely absent) could be derived, providing potential decision support to auditors during client selection or audit planning. We also observe that both financial statements and text in annual reports can be utilised to detect non-fraudulent firms. However, non-annual report data (analysts’ forecasts of revenues and earnings) are necessary to detect fraudulent firms. This finding has important implications for selecting variables when developing early warning systems of financial statement fraud.