Adaptive Machine Learning for Detecting Emerging Financial Fraud Patterns in Digital Payment Networks

Abstract

This study aimed to evaluate static and adaptive machine-learning models for detecting fraud in digital payment transactions under conditions of severe class imbalance and evolving fraud behaviour. The study compared Random Forest, conventional XGBoost, and Adaptive XGBoost across multiple performance measures, and further assessed static and adaptive XGBoost across five sequential evaluation periods to examine temporal robustness under a rolling-window learning strategy applied to a large simulated transaction dataset. Random Forest and XGBoost achieved high overall accuracy but limited recall, reflecting weak minority-class detection despite strong aggregate performance. Adaptive XGBoost recorded lower accuracy and precision but achieved higher recall and stronger discriminatory performance. Sequential evaluation showed static XGBoost maintained more stable precision, while Adaptive XGBoost achieved higher recall in most periods, with fluctuating precision but a stronger overall balance between precision and recall. Adaptive XGBoost demonstrated greater sensitivity to emerging fraud patterns and improved recall over time, whereas static XGBoost offered more consistent precision, indicating a trade-off that should guide model selection in real-world fraud-monitoring systems requiring both detection sensitivity and operational reliability.