Fuzzy matching is approximate string-matching that scores how similar two names are instead of demanding an exact match. It lets screening catch misspellings, name variants, reordered words, and transliteration differences, because sanctioned names rarely arrive perfectly formatted.
What is fuzzy matching, in plain English?
Fuzzy matching compares two names and produces a similarity score rather than a yes-or-no answer. Exact matching only fires when strings are identical, which is nearly useless for screening because real names come in with typos, missing middle names, reversed word order, and different spellings. Fuzzy matching asks how close two names are and flags anything above a set threshold.
It exists because sanctioned names rarely arrive perfectly formatted. A payment message might spell a name slightly differently than the list does, put the surname first, or use one of several valid romanizations of a non-Latin name. A rigid exact match would sail past all of those. Fuzzy matching is what bridges the gap between messy input and clean list data.
The catch is that it runs on a threshold. Set the similarity cutoff and everything above it becomes an alert. Where you set that line decides the whole balance between catching real hits and drowning in false ones.
Loose vs tight thresholds
What changes | Loose threshold | Tight threshold |
Catches | More real hits, plus lots of near-misses. | Only very close matches. |
Queue impact | Floods analysts with false positives. | Quieter queue. |
Main risk | Alert fatigue and slow response. | Real hits fall below the line, false negatives. |
Best for | High-risk flows where misses are unacceptable. | Cleaner data where variants are less likely. |
What it catches
Variation | Example of what fuzzy matching bridges |
Misspellings | A single transposed or dropped letter that exact matching would treat as a different name. |
Reordered words | Surname and given name swapped, common across naming conventions. |
Transliteration | Different valid romanizations of a name written in another script. |
Extra or missing tokens | A middle name, initial, or title present in one version and absent in the other. |
What it looks like in practice
In practice
A wire names a beneficiary with the surname first and one letter off from the list spelling. Exact matching returns nothing. The fuzzy matcher scores the name against the list entry, accounts for the reordered tokens and the single-character difference, and lands above the threshold, so an alert fires.
An analyst confirms the date of birth against the entry and holds the payment. The team also notes that the threshold catching this was validated against a test set, so they trust it is tuned to catch this kind of near-miss without flooding the queue with distant, unrelated names.
Why it matters to operators
Fuzzy matching is a tuning tradeoff, and getting the threshold right is one of the most consequential calibration decisions a screening program makes. Looser settings catch more real hits but flood the queue with false positives. Tighter settings do the reverse: a quieter queue, but real near-matches start slipping below the line into false negatives.
Operators do not set it once and walk away. They calibrate the scoring, often blend fuzzy matching with phonetic and token-based methods, and validate the settings against known-positive and known-negative test sets. The error to avoid is setting and forgetting the thresholds, because the right balance shifts as your data quality, customer base, and the lists themselves change.
What to watch in the data
- Threshold validation. Test settings against known-positive and known-negative sets, not just against queue volume.
- Blended methods. Combine fuzzy scoring with phonetic and token-based matching to catch variants a single method misses.
- Drift over time. The right balance moves as data and lists change; a set-and-forget threshold degrades quietly.
- Data normalization. Clean, consistent input names make fuzzy scores meaningful; messy fields distort them.
- Both error types. Watch false positives and false negatives together; tuning one without the other misleads you.
Quick questions
Why not just use exact matching?
Because real names arrive with typos, reordered words, missing tokens, and different transliterations. Exact matching would miss almost all of them. Fuzzy matching scores similarity so screening can catch names that are close but not identical to a list entry.
What does the threshold control?
The similarity cutoff above which a comparison becomes an alert. Loose thresholds catch more real hits but generate more false positives; tight thresholds cut noise but risk false negatives. Setting it is a deliberate balance, not a default.
How is fuzzy matching related to phonetic matching?
They are complementary. Fuzzy matching scores string similarity; phonetic matching compares how names sound. Programs often blend them, along with token-based methods, so that names varying in spelling, sound, or word order are all caught.
Can you set the thresholds once and be done?
No. The right balance shifts with your data quality, customer base, and the lists. Set-and-forget thresholds degrade quietly over time, so programs revalidate periodically against test sets rather than leaving the settings untouched.
How do you know the settings are right?
By validating against known-positive samples, which should match, and known-negative samples, which should not. If seeded true positives clear or unrelated names flood the queue, the thresholds need recalibration.
Go deeper
- OFAC, US Treasury ↗ — Administers US sanctions programs, the SDN list, and licensing.
- European Banking Authority ↗ — EU banking regulator. Strong Customer Authentication under PSD2 and AML guidance.