Academic journal article Auditing: A Journal of Practice & Theory

Assessing the Risk of Management Fraud through Neural Network Technology

Academic journal article Auditing: A Journal of Practice & Theory

Assessing the Risk of Management Fraud through Neural Network Technology

Article excerpt

Key Words: Analytical auditing, Management fraud, Neural networks.

Data Availability: A list of the public companies used to develop the matched fraud and nonfraud sample is available from the authors upon request. All other data sources are described in the text.

INTRODUCTION

Numerous studies have examined techniques to assess the risk of errors and irregularities in financial statements (Calderon and Green 1994; Green and Calderon 1995; Loebbecke and Steinbart 1987; Wheeler and Pany 1990; Wilson and Colbert 1989). Significant post-expectation gap research has stemmed from revised auditing standards that extend the auditor's responsibility for the detection of material misstatement (AICPA 1988a, 1988b). In conflict with the goal of increased effectiveness of fraud detection are market pressures to increase audit efficiency. The National Commission on Fraudulent Financial Reporting (NCFFR), the Treadway Report, describes market pressures that include increased competition for services, reduced market fees, and a dynamic product service mix offered by the industry (NCFFR 1987).

Advanced analytical procedures (APs) possess both efficiency and effectiveness potential. Statement on Auditing Standard (SAS) No. 56, "Analytical Procedures," requires the use of APs in the planning and review stages of all audits. The Standard defines APs as "consisting of evaluations of financial information made by a study of plausible relationships among both financial and non-financial data" (AICPA 1988b). APs range from basic scanning to multifactor regression. Researchers have examined the use of both endogenous and exogenous data analytical procedures (Calderon and Green 1994; Green and Calderon 1995; Loebbecke and Steinbart 1987; Wheeler and Pany 1990). While prior research has indicated that simple and advanced planning techniques aid auditors when assessing primary risk of misstatement (Biggs and Wild 1984; Loebbecke et al. 1989; Loebbecke and Steinbart 1987; Wilson and Colbert 1989), results have not been promising because of high error rates.

Methods used to signal errors have also been applied to fraud. Unlike errors, fraud is intentional and may be intentionally hidden. While developing the study's sample, it was noted that most fraud cases affected multiple accounts, financial statements and transactions. Individual APs are reported to be minimally effective in the detection of fraud (Green and Calderon 1995). Increased effectiveness may be achieved if APs survey an entire transaction cycle, recognizing changes in aggregate cycle and statement relationships as opposed to individual account changes. Neural networks (NNs) can simultaneously examine the changes and relationships between multiple accounts or groups of account balances. They are also a classification technique superior to earlier methods under varying situations (Salchenberger et al. 1992; Tam and Kiang 1992; Widrow et al. 1994). By examining a simultaneous relationship, NNs can then classify or predict the need to further investigate account balances represented on financial statements.

The purpose of this study is to develop a NN model for fraud detection employing endogenous financial data. Three models, using different expectation methods to develop data input, act as an investigation rule to classify financial statement data. The paper will focus on the preliminary audit stage, where resources are allocated to areas based on screening tools indicating the increased risk of misstatement. The task of predicting financial statement fraud can be viewed as a classification problem, mirroring the technique used in the NN application to financial institutions' failure prediction (Tam and Kiang 1992; Salchenberger et al. 1992). The NN learns the pattern of input data in a learning fraud and nonfraud sample. The learned behavior pattern is then applied to a test fraud and nonfraud sample. Using the predictive model, each financial statement is categorizes as either fraud or nonfraud. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.