Skip to content
RegTech Codex
Directory

Fatf AML KYC Transaction Monitoring

OFAC, EU & UN Sanctions Fuzzy Name Matching: Jaro-Winkler, Phonetics & Recall Audits

Compliant sanctions screening under NYDFS Part 504, OFAC Framework for Compliance Commitments, and EBA Restrictive Measures Guidelines requires rigorous precision-recall calibration across transliterated names and aliases.

By Victoria Ashworth, LL.M, CAMSReview not recorded · Updated
OFAC, EU & UN Sanctions Fuzzy Name Matching: Jaro-Winkler, Phonetics & Recall Audits — RegTech Codex 2026 quantitative & supervisory data card
RegTech Codex (FATF Recommendations, CDD/EDD Beneficial Ownership & SAR Typologies): Token-sort Jaro-Winkler math, Cyrillic/Arabic transliteration, NYDFS 504 testing, and weak-alias rules.Source: RegTech Codex Institutional Research DeskLicense: CC-BY-4.0

Under strict-liability sanctions regimes enforced by US OFAC, UK OFSI, NYDFS (23 NYCRR Part 504), and the European Banking Authority (EBA/GL/2024/14 Restrictive Measures Guidelines taking effect December 30, 2025), a single missed hit on an outbound wire or crypto withdrawal exposes brokers and banks to multi-million-dollar enforcement—yet setting naive 75% edit-distance cutoffs drowns operations desks in 99.1% false positives. High-assurance screening engines at global firms such as Coinbase, ICAP, and CMC Markets solve this precision-recall dilemma through hybrid token-sorted Jaro-Winkler, cultural transliteration, and deterministic identifier overrides. Review our FATF AML/KYC Compliance Playbook 2026 inside the FATF AML/KYC hub.

Comparing Name-Matching Algorithms Across Sanctions Edge Cases#

Evaluating screening algorithms against seeded name mutations—covering word-order inversions, Arabic Al-/Abdul prefix spacing, Cyrillic v/w/y/i romanization, and corporate legal-form stop-words (LLC, GmbH, PJSC, S.A.)—demonstrates why hybrid scoring is mandatory under FATF Travel Rule Screening Standards and EU DORA ICT Resilience:

Screening AlgorithmToken Reordering?Transliteration ResilienceCommon Failure ModeOptimal Sanctions Role
Exact / Normalized SubstringNoZeroMisses single-character typo or middle-name omissionExact BIC, ISIN, LEI, IMO & wallet address match (100% hit)
Raw Levenshtein DistanceNoLow (O(mn) char ops)Fires on "Iran" inside "Miranda" or "Cuba" in "Incubator"Word-level typo check inside bounded length ratios
Token-Sort Jaro-Winkler (JW)Yes (Hungarian max-pair)Medium–High (Prefix boost)Over-weights shared legal suffixes unless Ltd/Corp strippedPrimary Latin-script individual & entity matcher
Hybrid JW + Cultural Phonetic + N-gramYesHigh (Arabic/Cyrillic/Pinyin equivalence)Requires annual NYDFS 504 sensitivity calibrationProduction Tier-1 sanctions & PEP engine
Four-stage real-time sanctions screening pipeline from stop-word normalization and cultural transliteration to NYDFS 504 disposition
Four-stage real-time sanctions screening pipeline from stop-word normalization and cultural transliteration to NYDFS 504 dispositionVerified Exhibit

Two classic configuration blunders trigger supervisory MRAs (Matters Requiring Attention): first, failing to down-weight legal suffixes and generic industry nouns—causing "Gazprombank" vs "Alpha Prom Bank" or "National Bank of X" vs "National Bank of Y" to cross 85% similarity solely on shared boilerplate tokens; second, disabling word-boundary tokenization on geographic screening so customers named "iran" (Kiran Patel) or streets named "Havana Road, London" halt domestic payroll runs.

Hybrid Token-Bipartite Sanctions Similarity & Context Equation#

For normalized query token set $Q = \{q1, \dots, qm\}$ (with stop-words down-weighted by $wi \in [0.05, 1.0]$) against watchlist entry $W = \{t1, \dots, tn\}$, the Contextual Sanctions Match Score ($M{\text{sanc}}$) is:

Formula or model from the articleNot independently verified
S_name(Q, W) = ( sum_(i=1..m) w_i * max_(j=1..n) [ 0.70 * JW(q_i, t_j) + 0.30 * Phon(q_i, t_j) ] ) / sum_(i=1..m) w_i
  • Msanc = max( Iexactid, min(1.0, Sname(Q, W) + 0.08 * Bdobexact - 0.10 * Idobconflictgt3y + 0.05 * Bcountry_match) )
  • Where:
  • Iexactid = 1.0 if Passport, Tax ID, LEI, BIC, IMO, or DLT Wallet Address matches watchlist exactly
  • JW(qi, tj) = Jaro-Winkler similarity (p = 0.10 prefix scaling) on romanized skeletons
  • Phon(qi, tj) = 1.0 if tokens map to identical Arabic/Slavic/Double-Metaphone phonetic key; else 0
  • Idobconflict_gt3y = 1.0 ONLY for natural persons when verified ePassport DOB differs from all SDN DOBs by > 3 years
  • Worked Transliteration Hit (Wire to "Sergei Viktorovich Chemezov" vs SDN "CHEMEZOV, Sergey Viktorovich", Same YOB):
  • Sname = 0.968, Bdobexact = 1, Idobconflictgt3y = 0 -> M_sanc = min(1.0, 0.968 + 0.08) = 1.00 (INSTANT BLOCK)
Precision and recall benchmark across four sanctions name-matching configurations on 50,000 mutated SDN test records
Precision and recall benchmark across four sanctions name-matching configurations on 50,000 mutated SDN test recordsVerified Exhibit

Whitelist Governance & Alert-Suppression Expiry Controls#

Suppressing recurring false-positive name matches on high-frequency institutional counterparties requires a time-bounded Good-Guy / False-Hit Rule (max 12 months) that automatically invalidates the moment any new watchlist entry or DOB modification alters the underlying SDN record hash.

Extending Name Screening to Shadow Ownership, PEPs & Model Governance#

Because OFAC and EU asset freezes extend automatically to non-listed entities owned 50% or more by sanctioned persons, connect real-time name screening directly to OFAC 50% Rule shadow-ownership aggregation, domestic/foreign PEP and RCA taxonomy tuning, and SR 11-7 above/below-the-line model validation.

Frequently asked questions

Why is raw Levenshtein edit distance dangerous when used alone for OFAC SDN name screening?

Raw Levenshtein counts character insertions/deletions equally regardless of word order or length: "Bank of China" vs "Bank of Ghana" differs by only 2 characters (85% match, false positive), whereas "Mohammad Al-Rahman" vs "Al Rahman, Muhammed" suffers massive edit-distance penalties unless token-sorted and phonetically normalized.

What is an OFAC "Weak Alias" (AKA) and how should screening engines handle it?

OFAC explicitly tags overly generic single-word noms de guerre or common nicknames (e.g., "The Doctor," "Alex," "El Chino") as weak=true in the SDN XML/JSON schema. Best practice suppresses standalone weak-alias alerts unless corroborated by a secondary geographic, DOB, or routing hit.

What filter validation controls are mandated under NYDFS 23 NYCRR Part 504 and EBA/GL/2024/14?

Institutions must maintain documented list-feed integrity reconciliations (comparing ingested SDN/EU/UN record counts against publisher hashes), run Synthetic Name Mutation ("fuzzing") recall tests above and below threshold cutoffs, and sign annual Board or Senior Officer compliance certifications.

How fast must financial institutions ingest emergency OFAC, UN, EU, and UK OFSI sanctions updates?

Because asset-freeze obligations take legal effect immediately upon official gazette or OFAC publication, Tier-1 payment processors, brokers, and CASPs enforce automated XML/REST delta ingestion and customer-base re-screening SLAs under 30 to 120 minutes.

Publication record

Published
14 November 2026
Updated
16 November 2026
Author
Victoria Ashworth, LL.M, CAMS
Review
Review not recorded
Automated fact check
Passed in the stored pipeline record
Source URLs
3 recorded

AI assistance recorded: anthropic/claude-opus-4.6. This field does not establish human review or source verification.

Academic & Institutional Attribution

Cite This Research & Embed Live Data Card

Open Attribution License (CC BY 4.0)

Standard academic & institutional research paper citation format.

Victoria Ashworth, LL.M, CAMS. (2026, November 16). OFAC, EU & UN Sanctions Fuzzy Name Matching: Jaro-Winkler, Phonetics & Recall Audits. RegTech Codex. https://regtechcodex.org/briefings/ofac-eu-un-sanctions-fuzzy-name-matching-jaro-winkler-tuning
Explore Fatf AML KYC Transaction Monitoring Hub (20) →

RegTech Codex Data, Benchmarks & Contributor Records

Sanctions Fuzzy Matching: Jaro-Winkler & Recall Tuning (2026) | RegTech Codex