Auditors traditionally use sampling techniques to examine general ledger (GL) data, which suffer from sampling risks. Hence, recent research proposes full-population testing techniques, such as suspicion scoring, which rely on auditors’ judgment to recognize possible risk factors and develop corresponding risk filters to identify abnormal transactions. Thus, when auditors miss potential problems, the related transactions are not likely to be identified. This paper uses unsupervised outlier detection methods, which require no prior knowledge about outliers in a dataset, to identify outliers in GL data and tests whether auditors can gain new insights from those identified outliers. A framework called the Multilevel Outlier Detection Framework (MODF) is proposed to identify outliers at the transaction level, account level, and combination-by-variable level. Experiments with one real and one synthetic GL dataset demonstrate that the MODF can help auditors to gain new insights about GL data.

Data Availability: The real dataset used in the experiment is not publicly available due to privacy policies.

JEL Classifications: M410, M42.

This content is only available via PDF.