Atala Alqtish

Type of the article: Research Article

Abstract
Banks rely on ownership, transaction, and institutional information to support anti-money-laundering decision-making, yet these sources differ in decision proximity, uncertainty, and evidentiary strength. This multi-study article evaluates three external anti-money-laundering signals within a decision-governance framework emphasizing provenance, auditability, and bounded interpretation. Study 1 examines United Kingdom Persons with Significant Control records as ownership-verification workload signals. Of 6.2 million active companies, 745,200 (12.0%) met at least one verification-priority rule. Among flagged companies, 58.3% were assigned to multi-layered control structures, 25.0% to non-United Kingdom or foreign-linked controllers, 11.2% to high officer/controller turnover, and 5.5% to circular ownership patterns. Study 2 compares extreme gradient boosting, graph convolutional networks, Graph Sample and Aggregate, and graph attention networks on the Elliptic transaction-network benchmark. Graph attention networks recorded the highest point estimates (AUROC = 0.960; AUPRC = 0.740), followed by Graph Sample and Aggregate (0.930; 0.610), graph convolutional networks (0.910; 0.560), and extreme gradient boosting (0.820; 0.370). Study 3 identifies a reported −6.8% posterior mean deviation in the Persons with Significant Control Anomaly Index after the late-June 2020 filing-regime boundary (95% credible interval: −10.9% to −2.4%), interpreted descriptively rather than causally. Overall, the findings establish an evidentiary hierarchy for governing heterogeneous anti-money-laundering signals while preserving boundaries between decision support, adjudication, and causal inference. The empirical objects are companies, transaction nodes, and register observations rather than banks.