Publikation
Machine-Learning-Based Anomaly Detection in Accounting Data: A Systematic Literature Review
Jan Gronewald; Peter Fettke
SSRN preprint, 2026.
Zusammenfassung
Machine learning (ML) is increasingly proposed for detecting anomalies in accounting data, yet the research remains fragmented and lacks a synthesis that reflects the specific characteristics of this domain. This study presents a systematic literature review of 22 studies on ML-based anomaly detection in financial accounting data, focusing on journal entries and the general ledger. Following the PRISMA 2020 guidelines, the corpus is analyzed from three perspectives: theory, methodology, and empirical evidence. Theoretically, the literature conceptualizes anomalies as audit leads that indicate potential misstatements rather than as direct evidence of fraud or error. However, audit semantics such as materiality are rarely integrated into model design and evaluation, leaving a semantic gap between statistical deviations and audit-relevant irregularities. Methodologically, the field is concentrated: autoencoder-based architectures dominate unsupervised detection and tree-based classifiers dominate supervised detection, while the central burden lies in transforming structurally rich, high-cardinality, and highly imbalanced journal data into ML-compatible representations without losing audit-relevant meaning. Empirically, the evidence rests largely on single-company, single-fiscal-year case studies using proprietary SAP-derived data; only 7 of 38 datasets and 4 of 22 code bases are publicly available, which limits reproducibility and comparability. We conclude that the principal bottleneck of the field is not a lack of modeling ideas but the absence of a shared semantic and empirical foundation. Based on these findings, we derive a research agenda comprising conceptual foundations, preprocessing guidelines, hybrid detection systems, human-AI interaction, and a shared benchmark infrastructure to advance ML-based anomaly detection toward audit-useful practice.