Abstract

Detecting anomalous journal entries in a company's general ledger is essential for external auditors. An increasing trend employs outlier detection (OD) methods, especially machine learning methods, for anomaly detection in journal entry data. Recent research often lacks comparative analysis of OD methods. Thus, this study provides a comparative analysis of OD methods for journal entry anomaly detection using real-world accounting data. Additionally, in the context of domain-specific data preprocessing, we give special consideration to the amount, due to its importance for auditors. This yields three different dataset variants. We conduct our analysis based on three example accounts manually labeled by external auditors. Autoencoders, clustering-based local outlier factor (CBLOF), and histogram-based outlier score (HBOS) consistently outperform other methods across different accounts and dataset variants. With the provided results, this research enhances the understanding and applicability of OD methods for journal entry anomaly detection.

Share

COinS