Coverage and Protocol-Aware Evaluation of Risk Enrichment and Graph-Tabular Learning for Anti-Money Laundering and Fraud Detection
Reported gains from graph-based machine learning in anti-money laundering and fraud detection can vary substantially with the historical evidence available to a model and with the evaluation protocol used. This study introduces a coverage-aware and protocol-aware evaluation framework for assessing account-risk enrichment, graph neural networks, tabular gradient boosting, and graph-tabular stacking across synthetic banking anti-money-laundering data, a real Bitcoin transaction graph, and a complementary real credit-card fraud dataset. The analysis compares account-risk enrichment, four topology-only graph neural networks, XGBoost, LightGBM, and embedding-based stacking under warm-start, account-grouped, cold-start, coverage-sensitivity, and temporal evaluation settings. On the synthetic HI-Small benchmark, strict account grouping removes source-account historical coverage by construction and substantially limits the opportunity for enrichment to contribute useful signal. Under the warm-start protocol, enrichment produces only a small change in discrimination, while the coverage-sensitivity analysis shows no reliable improvement in ROC-AUC as historical coverage increases. PR-AUC results instead indicate a small but consistent performance cost at most tested coverage levels after correction for multiple comparisons. On the lower-coverage LI-Small benchmark, enrichment again shows a small negative effect that does not remain significant after manuscript-wide correction. External validation on the Elliptic Bitcoin graph shows that topology-only graph neural networks do not automatically outperform strong tabular models. GraphSAGE is the strongest graph neural network tested, but XGBoost and LightGBM achieve clearly higher ROC-AUC and PR-AUC. Graph-tabular stacking is also dataset dependent: it yields small positive gains on Elliptic, while tabular features alone are competitive with or superior to stacking on HI-Small. These results show that graph-based improvements should not be interpreted independently of historical coverage, split construction, and feature redundancy. Practical evaluation should report coverage and protocol alongside graph-derived claims, use strong tabular baselines, and complement ROC-AUC with PR-AUC and calibration-oriented metrics before concluding that a graph-based method provides a robust advantage.