Sanctions Screening: Fuzzy Name Matching Without False Positives

Sanctions screening sits apart from adverse media and PEP checks: a match against OFAC SDN, the EU consolidated list, or UK OFSI is not a risk signal to weigh — it can be a hard legal prohibition on transacting. That asymmetry makes screening a matching problem, not a judgment problem. The engineering question is how to catch every genuine listed party while keeping analysts from drowning in near-identical names.

Why Exact Matching Fails

Names on sanctions lists arrive in inconsistent forms. A single individual may appear as a primary name plus a dozen aliases, low-quality and strong. Arabic, Cyrillic, and Chinese names are transliterated differently across lists — Mohammed, Muhammad, Mohamed all reference the same phonetics. Date-of-birth fields may be a year, a range, or absent. If you match only on exact strings, you miss listed parties who used a variant spelling on their application form.

The lists themselves are structured for this. OFAC publishes name components, aka types (a.k.a. vs. f.k.a.), and secondary identifiers such as passport numbers and DOB. Ignoring those fields collapses your precision. A defensible screening engine parses each entry into components rather than treating the full name as one blob.

Tuning the Match Threshold

Most engines score name similarity using algorithms like Jaro-Winkler, Levenshtein distance, or token-based comparison, then apply a threshold above which a candidate becomes an alert. Set the threshold too high and you miss transliteration variants; set it too low and every common surname generates hits. There is no universal number — the correct value depends on your customer base and list coverage.

• Segment by name frequency. A rare full-name match at 85% deserves more suspicion than a common one at 92%.
• Weight secondary identifiers. A name at 80% with a matching DOB and nationality is stronger than a name at 95% with a contradicting birth year.
• Handle name order. Token-based matching that ignores word order catches reversed given/family names common in cross-border data.
• Suppress known false positives. A whitelist of previously cleared candidates, tied to a specific list version, cuts repeat noise without hiding new entries.

Document why each threshold was chosen. When a regulator asks why a name at 78% did not alert,