Abstract

The pervasive threat of internal fraud, amplified by the digitalisation of financial services, poses a significant challenge to the banking industry, frequently overwhelming traditional auditing and investigative methods. This study addresses the urgent need for more adaptive and technologically advanced fraud mitigation strategies by proposing and validating a synergistic framework that integrates forensic accounting principles with advanced data analytics. Utilising a sophisticated, simulation-based methodology, we generated a realistic dataset of 1.5 million banking transactions, embedding subtle and complex internal fraud scenarios. This research evaluates the efficacy of four distinct machine learning models—Logistic Regression, Multi-Layer Perceptron (MLP) Neural Network, XGBoost, and a Balanced Random Forest—in detecting these fraudulent activities. The findings reveal that the Balanced Random Forest model demonstrates superior performance, achieving an AUCROC of 0.975, an F1-Score of 0.915, and a Recall of 0.952, indicating its exceptional ability to identify over 95% of fraudulent transactions while maintaining a low rate of false positives. Feature importance analysis highlights that behavioural anomalies, such as deviations from an employee's average transaction patterns, are the most potent predictors of fraud, surpassing the significance of static transactional attributes like the amount. This study provides a practical, data-driven workflow that translates these analytical insights into an automated, real-time fraud detection and prevention system. The results confirm that the integration of forensic expertise with machine learning not only enhances detection accuracy and operational efficiency but also provides a proactive, resilient defence mechanism against the dynamic nature of internal financial crime.