Sanctions name matching: aliases, fuzzy matches, and review

On this page

What fuzzy name matching can and cannot tell you

Fuzzy matching looks for similarities beyond identical text. It can help retrieve a candidate when names differ in spelling, order, spacing, or script. That candidate still needs identity review. Two people can share the same full name, and one person can appear under several names.

OFAC’s own search tool uses fuzzy logic for its name field and describes its score as name similarity. SanctionsKit has its own retrieval and scoring implementation; its scores are not interchangeable with OFAC scores. A SanctionsKit score ranks evidence under a versioned policy, not the probability that someone is sanctioned or involved in wrongdoing.

Keep source names separate from derived forms

A source-provided alias is an authority-published fact. A normalized token, reordered name, or generated transliteration is a search aid. Preserve that distinction in a reviewer’s evidence: a Latin spelling generated by software should not appear as though the authority supplied it.

SanctionsKit accepts a full name without requiring Western first-name and last-name fields. It retains original names and uses derived forms during matching. Transliteration support is limited, so evaluate the scripts and spelling variations relevant to your own subjects. The methodology explains the path from source record to evidence.

Synthetic examples: similarity is only one part of identity

All names in these examples are invented. Mira Calder and Calder, Mira illustrate reordered tokens. North Lantern Trading and North Lantern Trading Ltd illustrate a legal suffix. A retrieved candidate in either pair is a reason to compare available identifiers and source facts; neither example establishes that the underlying parties are the same.

Now consider two fictional people both called Rowan Ellis. Exact name agreement can coexist with reliably different birth dates or identifiers. Conversely, a source that supplies only a birth year may still be consistent with a subject’s full date. The false-positive review checklist shows how to record those comparisons without dismissing uncertainty.

Treat weak aliases and missing data deliberately

OFAC labels some broad or generic aliases as weak because they can produce many false hits. Such an alias can still help corroborate other evidence. Preserve the publisher’s label and explain what actually caused a candidate to appear instead of presenting every alias as equally specific.

Known identifiers can help distinguish candidates when their type and issuer are comparable. Missing identifiers are simply missing. Do not infer nationality from a name, invent a full birth date, or silently replace a native-script name to make a request look more complete. The screening request documentation describes supported inputs.

Measure retrieval and review outcomes separately

A scoring function cannot rank a candidate that retrieval omitted. Evaluate whether expected candidates were retrieved, whether the returned evidence supports review, and how many unrelated candidates require attention. SanctionsKit uses bounded candidate retrieval and returns an error when a request exceeds its safe bound rather than presenting a truncated search as complete.

Create an independently labeled sample with known matches and known non-matches. Include common names, spelling changes, reordered names, scripts you actually encounter, partial dates, and organizations with similar trading names. Keep examples used to tune a process separate from those used to evaluate it. Synthetic examples are useful for behavior checks but do not establish recall for every real population.

Name-matching evaluation checklist

Record the source, engine, and policy versions with each evaluation so a change can be compared against the same expectations. Count missed expected candidates as well as unnecessary alerts; a lower alert count alone does not demonstrate better matching.

Preserve original names and identify every transformation used in the comparison.
Include positive and negative examples with independently established identities.
Check weak aliases, common names, and partial dates as separate cases.
Inspect candidate-bound errors and coverage failures rather than counting them as no matches.
Review changed outcomes after source, engine, policy, or input changes.
Document the limits of the sample before applying its findings to another population.

There is no universal score that clears a subject

OFAC does not recommend one match threshold for all searches. Its FAQ 250 ties that decision to the facts, risk assessment, and established compliance practices. A vendor score should likewise not be copied into an automatic approval rule without understanding what it measures.

SanctionsKit applies its matching policy and returns explicit outcomes. Your application handles those outcomes and your reviewers investigate the evidence. Version information and retained results make that separation visible as the workflow evolves.