Detecting Accounting Fraud with Noisy Labels

Partial observability is a prominent feature of accounting fraud. Not only the act of committing fraud is not directly observable, but the detected frauds are often revealed to the public with substantial time delays. In this article, we specify an underlying fraud data generating process that gives rise to noisy labels. It comprises two latent processes: fraud commission and fraud detection. We model fraud probability and detection probability separately, and combine them optimally in an ensemble machine learning algorithm. We show that our new fraud prediction model significantly outperforms various competing models, and generates more accurate fraud prediction