Entity Resolution for Financial Crime

Financial-crime investigations often involve records with inconsistent names, addresses, identifiers, corporate structures, and transaction references. Entity resolution attempts to decide whether records refer to the same real-world person or organization. Methods range from exact identifier matching to probabilistic and machine-learning systems using names, dates, addresses, ownership, devices, or relationships. A linked graph can help analysts see patterns, but links are hypotheses that need evidence. False merges can combine separate people or businesses and cause unfair scrutiny, account restrictions, or incorrect reports. Missed matches can fragment evidence and hide relevant relationships. Namesakes, transliteration, shell companies, shared addresses, and stale data complicate matching. The FFIEC BSA/AML manual emphasizes understanding customer information and reviewing unusual activity in context; entity resolution supports that work but does not replace it. Teams should preserve match features, confidence, source data, and analyst decisions. Evaluate precision and recall against reviewed examples, and test by entity type, language, and data source. Use tiered thresholds: automatically link only high-confidence identifiers, send uncertain matches for review, and keep no-match possibilities visible. Do not treat graph proximity as proof of collusion. Maintain correction and appeal processes for entities incorrectly linked. Reviewers should be able to dispute a connection and correct downstream records. Store the reason for a merge and the strength of the supporting attributes, so later investigators can evaluate whether a cluster still holds.