Detecting Accounting Frauds in Publicly Traded U.S. Firms: A New Perspective and a New Method
48 Pages Posted: 8 Oct 2015 Last revised: 1 Sep 2019
Date Written: August 29, 2019
Prior accounting research often develops fraud prediction models based on theory-motivated financial ratios. We propose a new fraud prediction model that makes use of raw financial data as its input. In addition, we employ one of the most powerful machine learning methods, ensemble learning, rather than the commonly used method of logistic regression. Our ensemble learning model, which is based on a small set of raw data derived from theory-motivated financial ratios, outperforms two benchmark models: Dechow et al.’s  logistic regression model based on financial ratios derived from the same raw data, and Cecchini et al.’s  support-vector-machine model with a financial kernel that maps the same raw data into a broader set of ratios. We find no evidence that an ensemble learning model based on the same financial ratios, or on the combination of the financial ratios and raw data, outperforms our ensemble learning model. There is also no evidence that an ensemble learning model based on several hundreds of raw data from the three financial statements outperforms our ensemble learning model. Overall, our results suggest the importance of selecting both machine learning method and model input in developing powerful fraud prediction models.
Suggested Citation: Suggested Citation