Machine Learning and Corporate Fraud Detection

The purpose of this dissertation was to study why corporate fraud detection models are often met with skepticism by industry practitioners despite a vast literature supporting their use. This dissertation examined the parsimonious standards in the academic literature for corporate fraud detection and included the latest studies that introduced ideas from Benford’s Law and machine learning algorithms. The study of corporate fraud detection models is important because academic literature is relied upon by industry practitioners and government regulators including the Securities and Exchange Commission. This paper starts with a critique that was recently published in Econ Journal Watch. This critique examined the results of a paper recently published in the Journal of Accounting Research applying machine learning to the detection of accounting fraud. Afterwards, I applied the most popular ensemble boosting algorithm in machine learning known as XGBoost to a comprehensive sample of financial ratios and variables. In addition to this model, I ran a horserace with the other models from the extant literature. Results showed that the F-Score (Dechow, et al. 2011) stood up quite well against the machine learning models. Interestingly, a univariate screen on sales growth performed about as well as more complicated methodologies at the top of the probability distribution. Finally, I provided a discussion based on a Bayesian analysis that illustrated why practitioners find fraud detection difficult.