Catch Me If You Can: In Search of Accuracy, Scope, and Ease of Fraud Prediction

Abstract

We offer two new fraud prediction metrics: the AB-score, which is based on Benford’s Law, and the ABF-score, which combines the AB-score with the well-known F-score model from the seminal work by Dechow, Ge, Larson, and Sloan (2011). Multiple performance evaluation metrics show that the ABF-score provides the highest accuracy while the AB-score substantially expands the scope over which misreporting can be predicted. Additionally, both of these models are easier to estimate than machine learning and textual analysis models while delivering similar accuracy. Our models perform well in- and out-of-sample and across alternative misstatement proxies. Back of the envelope calculations suggest that our improved precision (over the F-score model) could save stakeholders about $16.24 billion (in 2016 dollars) annually. Finally, in a case study approach using a sample of notorious financial frauds, we show that our models offer sharper identification of fraud with an expanded scope that correctly identifies far more fraudulent firm-years.

Keywords: fraud prediction; Benford’s Law; F-score

JEL Classification: G20, G23, M41