Mining Financial Statements: Comparative Performance of Adaptive and Ensemble Models in Discovering Financial Statement Fraud

• Activity Number: 521
• Type: Contributed
• Date/Time: Wednesday, August 5, 2009 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): David G. Whiting*+ and James V. Hansen and James B. McDonald and Conan Albrecht and Steve Albrecht
• Companies: Brigham Young University and Brigham Young University and Brigham Young University and Brigham Young University and Brigham Young University
• Keywords: rule ensembles ; financial statement fraud ; random forests ; stochastic gradient boosting ; adaptive models ; data mining

Abstract

Financial statement fraud is a topic of critical interest to investors, analysts, regulators, auditors, and the general public. Major financial statement frauds at corporations such as Enron, WorldCom, Tyco, etc., has led to a loss of confidence in the integrity of American business, bankruptcies of major organizations, significant stockholder losses, and a severe decline in stock markets worldwide. The identification of this type of fraud using publicly available financial statements has resisted accurate detection. Previous research has at best achieved forecast accuracy approaching 70% using probit and logit regression methods. We compare the performance of partially adaptive estimators, stochastic gradient boosting, random forests, and rule ensembles in discovering financial statement fraud. The best of these, Friedman's RuleFit, achieved an average error rate of only 18.6%.