New challenges in internal auditing are created as all areas of companies are digitalized. These challenges are forcing internal auditing to implement more and more data-driven procedures. Auditing is increasingly using artificial intelligence methods such as neural networks to overcome these challenges. Since in internal auditing labels are usually not available at the beginning of an audit engagement, unsupervised methods have to be used. We used autoencoders as an unsupervised method, which we evaluated for its use in auditing in a practical case study with an international automobile manufacturer. For the case study, two real-world, non-financial data sets from production-related processes were provided. The results of the case study show that the use of autoencoders can support auditors in the audit execution and in the audit planning process step to improve the quality of the internal audit engagement.

Auditing has to adapt to the growing amounts of data caused by digital transformation. One approach to address this and to test the full audit data population is to apply rules to the data. A disadvantage of this is that rules most likely only find errors, mistakes or deviations which were already anticipated by the auditor. Unsupervised anomaly detection can go beyond those capabilities and detect novel process deviations or new fraud attempts. We conducted a systematic review of existing studies which apply unsupervised anomaly detection in an auditing context. The results reveal that most of the studies develop an approach for only one specific dataset and do not address the integration into the audit process or how the results should be best presented to the auditor. We therefore develop a research agenda addressing both the generalizability of unsupervised anomaly detection in auditing and the preparation of results for auditors.

No abstract

No abstract

Outlier explanation approaches are employed to support analysts in investigating outliers, especially those detected by methods which are not intuitively interpretable such as deep learning or ensemble approaches. There have been several studies on outlier explanation in the last years. Nonetheless, there have been no outlier explanation approaches for mixed-type data. In this paper we propose multiple approaches for outlier explanation on mixed-type data. We benchmark them by using synthetic outlier datasets and by generating ground-truth explanation for real-world outlier datasets. The results on the various datasets show that while there is no approach that dominates others for all types of outliers and datasets, some can offer a consistently high performance.