Detecting Accounting Frauds in Publicly Traded U.S. Firms: New Perspective and New Method

Abstract

We propose a new perspective and a new method to detect accounting frauds out of sample. We show that a logistic regression that directly uses raw accounting data as regressors outperforms the traditional logistic regression that uses expert-identified financial ratios. Using the same raw data as inputs, ensemble learning, a state-of-the-art machine learning method, further outperforms the logistic regression model. The ensemble learning method also outperforms a support vector machine (SVM) with a financial kernel that maps the same raw accounting data into a broader set of ratios and changes in ratios. Overall, our results suggest that the existing fraud prediction models haven’t fully utilized the information from publicly available financial statement data. In addition, we show that it is possible to extract such useful information by adopting better fraud prediction models that rely on raw accounting data rather than financial ratios as model inputs.