Directional Asymmetry in banking Journal Entries: A Machine Learning Perspective on Audit Risk Prioritization

Abstract

Banking journal entry testing needs to provide auditors the ability to evaluate large sets of transactions and assess which transactions should receive the most focused attention. The purpose of this paper was to determine if a journal-entry risk-ranking model (AARR-XGBoost), based on a directional asymmetry-aware representation of entry features, could produce better risk rankings than traditional methods. Thirty individual simulated audit environments were created, including 240 companies with varying characteristics, 231,773 reference-year documents, and 231,986 target-year documents. The features used to represent account roles, account-pair relationships, creator paths, timing patterns, and transaction amounts were all extracted from the historical ledger records of each company. With a limited review budget of 3%, AARR-XGBoost captured 24.83% of the scenario-positive documents (95% CI [20.79%, 29.07%]), outperforming conventional XGBoost, Isolation Forest, transparent practice-informed rules, and random selection. When feature-set analyses were conducted to identify the specific components that contributed to AARR-XGBoost's performance, directional role violation and account-pair surprisal were the largest contributors to ranking performance. Additionally, an independent evaluation using a publicly accessible ERP fraud dataset provided descriptive evidence of how well AARR-XGBoost performed relative to conventional XGBoost in a different environment. Specifically, AARR-XGBoost identified 30 of 32 labelled positive documents, while conventional XGBoost identified 24. Collectively, the results demonstrate that including directional asymmetry within the representation improved journal-entry risk-ranking performance in controlled simulations when audit review capacity was constrained. The results also support the development of a reproducible framework for assessing journal-entry analytics and suggest future research using either operational data or de-identified ledger data.