Machine learning approaches to detect financial statement frauds and accounting control issues

Financial statement fraud is a serious concern for investors and other stakeholders. According to the Association of Certified Fraud Examiners' 2022 Report on global occupational fraud, fraud costs an estimated 5 % loss relative to revenue each year. Among them, financial statement fraud accounts for only 9% of cases but is the costliest type of occupational fraud, with a median loss of $593,000 in 2022. Traditionally, a company relies on internal control activities to prevent and detect fraud. Firms with a failure in internal control are more likely to indulge in fraud and have a higher probability of filing for bankruptcy. The Sarbanes-Oxley Act (SOX) requires listed companies to disclose material weaknesses in internal controls over financial reporting. This requirement is expected to provide investors with an early warning signal regarding the reliability of reported financial information. However, regulators and practitioners have expressed concerns about the unreported control weaknesses and the reliability of the SOX 404 reports, making detecting control issues increasingly important to investor protection.Recent artificial intelligence (AI) developments have precipitated substantive changes in accounting education, research, and practice. As an application or subset of AI, machine learning allows machines to learn from data without being programmed explicitly. This dissertation aims to provide solutions to two extremely difficult and related accounting problems using advanced machine learning techniques: financial statements fraud detection and accounting control issues identification.The first study proposes a novel fraud detection model based on an ensemble machine learning algorithm known as Cost-sensitive Cascade Forest. The proposed fraud detection model significantly outperforms the baseline, and the performance is further enhanced with appropriate missing data treatment. The second study focuses on accounting control identification. The results demonstrate the great potential of online employee reviews in signaling accounting control issues, which has yet to be explored in the literature. In the study, online employee reviews were collected, manually labeled according to the COSO’s internal control framework (2013), trained with advanced data mining techniques, and, finally, a web application based on the experimental results was developed and deployed in a cloud server.