The Efficacy of Red Flags in Predicting the SEC's Targets: An Artificial Neural Networks Approach

Abstract

As a proximal precursor to the Sarbanes-Oxley Act of of 2002 (Section 704: Study of Enforcement Actions), this paper tests the ability of the 'red flags' that are defined in terms of public information to predict which firms will be singled out as reporting violators by the enforcement program of the Securities and Exchange Commission (SEC). The SEC enforcement program consists of investigations and subsequent injunctive actions or administrative proceedings against offending registrants and auditors (Feroz et al.1991). The agency hopes "to...anticipate emerging problems" (SEC, 1989, p.1). Potential SEC enforcement actions provide incentives for corporate officers and independent certified public accountants (CPAs) to avoid unacceptable practices whose "effective prosecution is essential to preserving the integrity of the disclosure system" (SEC, 1989,p.8).

Feroz et al. (1991) report that the owners of enforcement targets experience abnormal negative returns in the 40% range for the year ending on the day subsequent to the disclosure of reporting violation. Given the magnitude of these losses, investors typically sue managers and auditors (Feroz et al. 1991; Palmrose, 1991). Because of the potential for legal sanctions, CPAs need to anticipate if potential or current clients are likely enforcement targets.

The official literature provides guidance to assist auditors in predicting problem clients. Statement on Auditing Standards (SAS) No. 53 lists several categories of red flags including personnel, financial and audit oriented red flags. These red flags reflect the judgment and collective experience of the members of the Auditing Standards Board. However, there is little evidence concerning their ability to predict problem clients (Feroz 1997). Thus, a major goal of this paper is to provide feedback to regulators on how well, models using publicly available information, predict the reporting violators.

We compare the reporting violators with control samples of industry and size matched firms whose reporting was not challenged by the SEC. We then test the ability of artificial neural networks (ANNs) to distinguish between members of the target and control groups. We also compare ANN results with conventional logistic regression (logit) in order to provide comparative performance measures across prediction models.

Consistent with the expectations of the Auditing Standards Board, we find that even the publicly available red flags can discriminate between the reporting violators and control groups. The ANN model successfully distinguish reporting violators from matched control firms in 81% of the possible cases. The prediction success of the ANN model substantially exceeds that achieved by conventional statistical (logit) models.

Finally, these results have practical implications for auditors and regulators. In recent years, some CPA firms accepted future SEC targets within one year of the commencement of the SEC investigation. The use of red flags along with the ANN models may reduce the number of similar client disasters. Regarding the goal of providing feedback to the regulators who developed the red flags in SAS No. 53 and reinforced in SAS No. 82, we find that the reported results support regulators' efforts. Similarly, the mandate of SAS No. 56 that analytical procedures be used at the planning stage and the review stage of the audit is supported by our results.

Keywords: SEC, Financial reporting, Fraud, Red Flags, Regulations, Earnings management, ANN, SAS, Sarbanes-Oxley Act of 2002, Political Economy

JEL Classification: C40, C45, G18, G28, G38, K22,K23, L51,M40, M41, P16