Using data analytics and automation for testing journal entries in financial auditing

Entries in accounting journals may contain various types of anomalies, ranging from unintentional errors to sophisticated fraudulent schemes, which pose significant risks to companies and their auditors. Traditional approaches to testing accounting data, primarily based on rule-based controls, face limitations when analysing complex data structures, as they are unable to effectively capture relational and structural patterns between individual records. This thesis investigates the use of Graph Neural Networks (GNN) for unsupervised anomaly detection. Accounting data are represented as a single global graph, in which nodes correspond to accounts and edges to transactions. Within a systematic ablation study, the thesis compares selected architectures (VGAE, DOMINANT, and CoLA) and evaluates the impact of different input feature configurations on detection performance. The results show that GNN effectively identify anomalies arising from the structural characteristics of the data. The CoLA model achieves the best performance, while the analysis suggests that simpler topological features may be more effective for anomaly separation than more complex combinations of attributes. The models also demonstrate the ability to detect anomalies independently of transaction volume, thereby complementing traditional audit procedures focused on materiality. The findings suggest that GNN represent a promising approach for extending analytical procedures in the field of financial auditing.