Machine Learning–Based Transaction Risk Scoring Models for Financial Compliance Monitoring in Foreign Exchange Operations

Abstract

This study provided a quantitative cross-study synthesis of machine learning–based transaction risk scoring models for financial compliance monitoring in foreign exchange operations, with emphasis on measurable modeling practices, evaluation rigor, and governance instrumentation. A total of 124 analytic records derived from 89 publications were coded using a structured extraction protocol that converted heterogeneous reporting into standardized variables across model family, feature construction, validation design, labeling strategy, evaluation metrics, and governance controls. Descriptive results showed that ensemble models were the most frequently evaluated approach (44.4%), followed by logistic regression and generalized linear models (37.1%), decision tree models (33.9%), neural architectures (29.8%), and unsupervised or semi-supervised methods (26.6%). Customer-profile variables (69.4%), geographic corridor indicators (62.1%), and temporal aggregation features (57.3%) were the most commonly engineered feature groups, while network-based variables appeared in 41.9% of records. Evaluation practices were dominated by discrimination metrics (82.3%), with lower reporting of ranking metrics (61.3%), calibration measures (34.7%), and cost-sensitive analyses (28.2%). Governance and auditability constructs were underreported, with access control indicators documented in 29.8% of records and traceability artifacts in 22.6%. Reliability testing demonstrated strong internal consistency for governance maturity (α = 0.86) and documentation completeness (α = 0.84) indices. Logistic regression analysis showed that ensemble models (OR = 2.27, p = 0.008), neural models (OR = 1.99, p = 0.041), and out-of-time validation (OR = 2.83, p = 0.004) were significantly associated with high predictive performance reporting. Linear regression indicated that operational studies were strongly associated with higher governance maturity scores (β = 1.12, p < 0.001). Overall, the findings indicated that methodological rigor in validation design and feature construction was more consistently associated with reported performance gains than model family alone, while governance instrumentation and operational alignment remained uneven across the FX compliance risk scoring literature.