Financial Fraud Detection in Business Entities: A Comparative Analysis of Machine Learning Methods

The rapid development of digital financial services has led to a significant increase in the volume of electronic transactions conducted by business entities, while simultaneously making the effective detection and management of financial fraud a growing concern. This issue becomes particularly challenging when fraudulent transactions represent only a very small proportion of overall operations, making the appropriate evaluation of classification models essential. This study comparatively investigates Logistic Regression, Random Forest, and XGBoost machine learning methods for financial fraud detection using highly imbalanced transactional data. To examine the impact of class imbalance on model performance, experiments were conducted under two conditions: with and without the application of the Synthetic Minority Over-sampling Technique (SMOTE). Model performance and consistency were evaluated using 3, 5, and 10 fold Stratified K-Fold cross-validation based on Precision, Recall, F1-score, ROC-AUC, and PR-AUC metrics. The final test results demonstrate that the effectiveness of SMOTE is not uniform across all machine learning models. In particular, Random Forest without SMOTE achieved an F1-score of 0.8743 and a PR-AUC of 0.8734. The findings provide a practical basis for improving automated monitoring of financial transactions and developing more effective financial fraud detection systems for business entities.