Self-Organizing Fuzzy and MLP Approaches to Detecting Fraudulent Financial Reporting

This paper predates the Sarbanes-Oxley Act of 2002 (Section 704: Study of Enforcement Actions) by three years. The purpose of this study is to compare a class of neural networks and fuzzy controller approaches, more specifically, a multi-layered perceptron (MLP) and a self-organizing fuzzy approach, in determining the efficacy of selected Statement of Auditing Standard No. 53 red flags in predicting the targets of the Securities and Exchange Commission's (SEC) investigations. The motivation for studying these two approaches is provided in part by our earlier work with conventional tools such as logit which generally leads to inferior prediction accuracy for such classification problems. In direct comparison of the performance on the present SEC investigation problem, the MLP network outperformed both logit and fuzzy approaches. We believe that the performance of the fuzzy approach can be significantly improved if the center vectors are fine tuned using some adaptive algorithms. Our empirical results demonstrate that both MLP and fuzzy approaches can be powerful tools, especially in the pattern classification or detection problems, such as those related to fraudulent financial reporting