FRAUD DETECTION USING MACHINE LEARNING IN FINANCIAL TRANSACTIONS: A SYSTEMATIC REVIEW

Fraud in online financial transactions is a constant and growing threat as digital commerce and real-time payments expand. Since fraud is rare, labels are often delayed or unclear, and attackers change tactics quickly, detection systems must handle class imbalance while meeting strict requirements for speed, privacy, security, and oversight. This review summarizes peer-reviewed research from 2020 to 2025 on machine learning methods for detecting fraud in card-present and card-not-present payments, account-based transactions, bank transfers, and multi-channel monitoring. After a structured screening process, 1,021 records were reviewed, duplicates were removed, and 77 studies were included for qualitative analysis. To look beyond accuracy, we introduce the Operational Capability Triad (ORT), which focuses on three main areas: data realism and leakage control, resilience to changing fraud patterns, and understandability and governance for human investigators. The literature shows that gradient-boosted decision trees are still strong for tabular data, deep learning is used more for behavioral sequences, and graph-driven methods help find organized fraud. The review ends with practical recommendations, including drift-aware evaluation, privacy-respecting benchmarking across institutions, and standard metrics to measure loss prevention and analyst workload.