Developing automated applications for clustering and outlier detection: data mining implications for auditing practice

Occupational fraud is viewed as a growing, global problem, and solutions are thus needed. Furthermore, since passage of Statement on Auditing Standards (SAS) 99, auditors have been held to a higher standard relative to audit quality. More specifically, auditors are now required to consider the risks of material misstatement due to fraud throughout the entire audit process. Interestingly, clustering has emerged as one method for addressing this challenge. Unfortunately, a set of difficulties exists in implementing data mining in practice, such as complexities relative to data pre-processing, algorithm selection, and model evaluation schemes. Given this, the traditionally trained auditor is ill-equipped to effectively perform clustering in the context of the financial statement audit. Given the likelihood that clustering will become ubiquitous in the auditing and accounting domains of the future, accounting professionals should be positioned to effectively use data mining in fulfillment of their responsibilities. One possibility for achieving this involves substantial automation of the clustering routine. In this way, many of the historically manual decision points within the process can be eliminated, thus making it a more user friendly task. In so doing, practitioners could then focus on problem investigation and resolution, instead of being burdened with technical nuances of clustering operations. In this dissertation, efforts are made to progressively automate clustering and outlier detection. This is done via auditing credit card customer data. First, cluster analysis is performed to generate an initial set of partitions. Next, each group is evaluated using various mechanisms to note whether nested clusters exist. Following this, a method for identifying irregularities is proposed and implemented. Overall, results demonstrate clustering and outlier detection can provide utility in the auditing of organizational assets. In conclusion, findings are synthesized and two distinct applications are created. These are provided as implementable artifacts as well as proofs of concept demonstrating feasibility of automating clustering and outlier detection routines. It is hoped auditors see value potential in this type of software, and ultimately find such programs to offer both ease of use and perceived usefulness when investigating fraud in audit engagements.Ph.D.Includes bibliographical referencesby Paul Eric Byrne