The prevalence and price distorting effects of undetected financial misrepresentation: Empirical evidence

Abstract

We use a comprehensive database of regulatory enforcement actions for financial misrepresentation to estimate prediction models using logistic, machine learning, and bivariate probit classifiers. Our parsimonious logistic model and three versions of a Support Vector Machine learning model perform well, each with an average area under the ROC curve (AUC) of 0.78 in out of sample tests. The base logistic model implies that 22.3% of Compustat-listed firms are engaged in financial misrepresentation that is potentially sanctionable by regulators in an average year. The average violation period is 3.1 years, implying that 22.3%/3.1 = 7.2% of firms initiate financial reporting practices each year that are potentially sanctionable. Of these firms, 3.5% eventually are sanctioned by regulators. We use these findings infer the fraction of firms that misrepresent their financials and yet never face regulatory penalties, to estimate the size of the price distortions imposed by misrepresentation on the shares of both misrepresenting and non-misrepresenting firms, and to estimate the size of firms’ ex ante expected costs of engaging in financial misrepresentation that incorporate both the probability of getting caught and the penalties if caught.

Keywords: Misrepresentation, fraud, prediction, enforcement, social cost

JEL Classification: G38, K22, L51, M48