Proposing JournalGuard a Graph Neural Network Algorithm for Continuous Audit of ERP Journal Entries with Explainable Control-Risk Scoring

Abstract

Enterprise Resource Planning (ERP) systems generate massive volumes of journal entries that support financial reporting, regulatory compliance, and internal control monitoring. Traditional audit procedures rely on rule-based validation, statistical sampling, or anomaly detection models that often struggle to capture complex relational dependencies among accounts, users, cost centers, and transaction flows. These limitations reduce the effectiveness of continuous auditing frameworks and increase the risk of undetected financial manipulation or control weaknesses. This study proposes JournalGuard, a novel Graph Neural Network (GNN)–based algorithm designed for continuous audit of ERP journal entries with explainable control-risk scoring. The proposed method models ERP journal data as a heterogeneous financial transaction graph in which nodes represent accounts, users, vendors, cost centers, and journal entries, while edges encode transactional relationships such as debit–credit mappings, approval hierarchies, and posting sequences. JournalGuard integrates Graph Attention Networks (GAT) with a temporal message-passing mechanism to capture both structural dependencies and sequential posting behaviors across accounting cycles. A novel Control-Risk Propagation Function (CRPF) is introduced to quantify the probability that a journal entry contributes to internal control violations. The algorithm further incorporates an Explainable Risk Attribution Layer (ERAL) that identifies the dominant graph features influencing risk scores, enabling auditors to interpret detected anomalies in relation to specific financial control pathways. To evaluate the effectiveness of JournalGuard, the model is benchmarked against widely used anomaly detection approaches including Isolation Forest, Autoencoder-based reconstruction models, and conventional Graph Convolutional Networks (GCN). Experiments are conducted on simulated ERP transaction datasets reflecting real-world financial control scenarios such as unusual posting times, circular journal adjustments, unauthorized account pairings, and abnormal transaction sequences. Performance is evaluated using classification accuracy, F1-score, Area Under the ROC Curve (AUC), and explainability fidelity metrics. Results demonstrate that JournalGuard achieves superior detection performance, improving anomaly identification accuracy by up to 18–24% compared with Isolation Forest and 12– 17% compared with deep autoencoder models, while also outperforming baseline GCN approaches in capturing relational control patterns. Graph-based risk visualization further enables auditors to identify clusters of suspicious entries and trace risk propagation paths across accounting entities. Comparative performance graphs show that the proposed algorithm maintains stable detection accuracy even under increasing transaction volumes and complex journal structures.
The findings indicate that integrating graph neural networks with explainable risk scoring significantly enhances the capability of continuous auditing systems within ERP environments. JournalGuard provides a scalable and interpretable framework that bridges the gap between advanced machine learning techniques and practical financial auditing requirements. The proposed approach has potential applications in automated internal control monitoring, regulatory compliance assurance, and intelligent financial fraud detection within modern enterprise accounting infrastructures.