< Back to Glossary

What is Fuzzy Matching?

Last updated: 26 October 2026

What is fuzzy matching?

Fuzzy matching is a technique that identifies text strings that are similar but not identical, using a similarity score instead of an exact comparison. In financial crime compliance, it lets screening and data systems recognize that “Mohammed Al-Rashid,” “Muhammad Alrashid” and “Mohamed Al Rasheed” may all refer to the same person.

It is the engine behind name screening, transaction screening and entity resolution. Without it, a sanctioned party could avoid detection simply by changing a letter.

Why is fuzzy matching important in AML and sanctions compliance?

Exact matching fails in the real world. Names are spelled differently across languages, transliterated from other scripts, abbreviated, reordered, misspelled and deliberately altered. Regulators expect institutions to screen against sanctions lists using logic that can handle these variations, and a program that only finds exact matches will miss genuine risk.

Fuzzy matching also creates the problem it is designed to solve. The wider the net, the more false positives it produces. In the FinCrime Frontier 2026-27 report, 37% of respondents reported false positive rates above 50%. Much of that volume comes from matching logic that is too broad, too blunt or poorly tuned.

How does fuzzy matching work?

Fuzzy matching compares two strings, calculates how similar they are and returns a score, typically from 0 to 100 or 0 to 1. The institution sets a threshold, and any pair scoring above it is treated as a potential match and raised as an alert. The process usually follows four steps:

  1. Normalize. Names are converted to a standard format, with case, punctuation, titles and spacing removed or standardized.
  2. Compare. One or more algorithms measure the similarity between the input and each list entry.
  3. Score. The results are combined into a single match score, often weighted by other attributes.
  4. Apply the threshold. Scores above the threshold generate alerts, and scores below it are discarded.

What are the main fuzzy matching algorithms?

Here are six common name-matching algorithms used in sanctions screening:

Technique How it works Example
Levenshtein distance Counts the single-character edits (insertions, deletions, substitutions) needed to turn one string into another “Smith” and “Smyth” differ by one edit
Jaro-Winkler Scores similarity by matching characters and their order, giving extra weight to a shared prefix “MARTHA” and “MARHTA” score about 0.96
Soundex and phonetic algorithms Convert names to codes based on how they sound “Robert” and “Rupert” produce the same code
N-gram matching Splits names into short character sequences and compares overlap Handles missing or swapped characters well
Token-based matching Compares names as sets of words, regardless of order “Ali Hassan” matches “Hassan Ali”
Transliteration-aware matching Maps names between scripts, such as Arabic, Cyrillic or Chinese, and the Latin alphabet Handles multiple valid spellings of one name

No single algorithm is enough. Production screening engines combine several, and add logic for aliases, nicknames, initials, name order and cultural naming conventions.

What makes names hard to match?

  • Transliteration: one name can have many valid Latin spellings.
  • Name order and structure: family names come first in some cultures and last in others.
  • Nicknames, abbreviations and initials: “Bill,” “William” and “W. Smith.”
  • Common names: many unrelated people share the same name.
  • Typos and poor data: manual entry errors and OCR mistakes.
  • Deliberate obfuscation: minor changes made to evade screening.
  • Business name variants: legal suffixes, abbreviations and trading names.

What is the trade-off between precision and recall?

Setting the match threshold is a balance. A lower threshold finds more true matches (higher recall) but also raises more false alerts (lower precision). A higher threshold produces fewer alerts but risks missing a genuine match.

The below table shows how sanctions screening thresholds affect false positives and detection accuracy:

Threshold Effect Risk
Too low Many alerts, most of them false Analyst overload, backlogs, alert fatigue
Too high Few alerts Missed sanctioned parties and regulatory breaches
Well tuned Alerts concentrated on genuine risk Requires testing, monitoring and documentation

The right setting depends on the institution’s risk-based approach, the list being screened and the consequences of a miss. Sanctions screening usually warrants a wider net than lower-risk checks, because a breach is a strict liability matter in many jurisdictions.

Where is fuzzy matching used in financial crime compliance?

  • Sanctions screening: matching customers, counterparties and payment parties to sanctions lists.
  • PEP and adverse media screening: finding people and entities across large, unstructured datasets.
  • Transaction screening: matching names, banks and locations in payment messages.
  • Entity resolution: linking duplicate and related records across systems.
  • KYC and onboarding: detecting duplicate customers and existing relationships.
  • Fraud and mule detection: finding accounts that share slightly varied identifiers.

How can you improve fuzzy matching accuracy?

  • Clean and standardize data first. Poor inputs weaken every algorithm.
  • Use secondary identifiers. Date of birth, nationality, address and ID numbers confirm or discount a name match.
  • Combine several algorithms, instead of relying on one.
  • Segment thresholds by risk. Apply different settings to different lists, customer types and name types.
  • Test and tune regularly. Use sample data, back-testing and below-the-line testing to confirm the engine finds what it should.
  • Resolve entities before screening, so each party is checked once, with full context.
  • Document the rationale. Keep evidence of the settings and why they were chosen, so you can explain them to a regulator.
  • Maintain the lists. Accurate matching depends on good watchlist management.

How is AI changing fuzzy matching?

Traditional fuzzy matching is rules-based and treats every similar pair the same way. Machine learning improves on it by learning which matches analysts clear and which they escalate, then scoring future alerts accordingly. This helps reduce false positives without lowering detection. Generative AI can structure unstructured text and summarize the evidence behind an alert, while agentic AI can investigate ambiguous matches and draft a documented recommendation. Explainability remains essential, because every match decision must be defensible. For more, see why regulators love agentic AI.

Fuzzy matching with SymphonyAI

SymphonyAI’s sanctions screening software combines configurable fuzzy matching, transliteration handling and entity resolution to find genuine risk across names, aliases and payment data. AI Augment for Screening adds generative and predictive AI on top of existing matching to significantly reduce false positives. To compare vendors, read the top 10 sanctions screening software.

Frequently asked questions

What is fuzzy matching in AML?
Fuzzy matching finds names that are similar but not identical, using similarity scores. It lets screening systems catch variations in spelling, transliteration and format that exact matching would miss.

What is the difference between fuzzy matching and exact matching?
Exact matching only flags identical strings. Fuzzy matching flags strings that are close enough, based on a threshold.

What algorithms are used for fuzzy matching?
Common ones include Levenshtein distance, Jaro-Winkler, Soundex and other phonetic methods, n-gram matching and token-based comparison. Screening engines usually combine several.

What is a good fuzzy matching threshold?
There is no universal setting. It depends on the institution’s risk appetite, the list being screened and testing results, and it must be documented and reviewed regularly.

Why does fuzzy matching create false positives?
Because it deliberately accepts near matches, common names and similar spellings can score highly even when the parties are unrelated.

Can AI improve fuzzy matching?
Yes. Machine learning can learn from analyst decisions to prioritize likely true matches and reduce false positives, as long as decisions remain explainable.