Far-Out-Of-Sample Financial Misconduct Prediction

Abstract

Most jurisdictions around the world do not have strong institutions in place to identify and enforce financial misconduct. In this paper, we investigate whether financial accounting misconduct models established in the literature can be applied to far-out-of-sample prediction. We use the latest and most powerful models that are trained, tuned, and tested on U.S. data, which contains hundreds of identified misconduct instances. We apply these predictors to a European sample of firms that contains 21 accounting misconduct cases. We find that simple Logistic Regression models as well as sophisticated RUSBoost Ensemble Learning assign high ex-ante probabilities of misconduct to a large portion of European cases during misconduct years. Comparing the out-of-sample classification performance of European cases to U.S. benchmarks shows our approach to be highly predictive. These results have important implications for global enforcement institutions on how to design a preselection model for overseen firms prior to investigations.

Keywords: misconduct prediction, accounting fraud, Machine Learning, ensemble learning JEL codes: C53, M41