Settings

The two lists

Both are opened by this tab. Neither is uploaded, which is the only reason a list of real customers or real subjects has any business going into a tool like this.

How close counts

Lower it and you will read more that is nothing. Raise it and you will not be told about something you should have seen. Eighty-two is a reasonable place to argue from, not a right answer. It does not go below sixty-five: Jaro-Winkler gives two entirely unrelated names something near half by its nature, so a threshold under that matches everybody and tells you nothing.

82%

What was compared

What came back

The record of the run

The decisions are the point. A screening run nobody wrote a reason on is a run that cannot be defended a year later.

How it works

MATCH screens a list of names against a watchlist or sanctions list you hold. It uses stated fuzzy-matching methods to catch spelling variations and transliterations, lets you set the threshold, and shows every candidate for a person to decide. It does not make the decision for you.

The guide: Fuzzy name matching, and the threshold trap

Questions people ask about MATCH

What lists can I screen against?

Any CSV you hold. The consolidated sanctions list published by OFSI, the OFAC SDN list, a politically-exposed-persons list you subscribe to, an internal list of people you will not deal with, a list of former staff, a debarment list: anything with a name column. You map which column is the name and, where they exist, the date of birth, country, reference and any listing detail you want carried through onto the hits. Only the name column is required.

Which algorithms does it use, exactly?

Jaro-Winkler with Winkler’s own constants, a prefix of at most four characters scaled by a tenth. Levenshtein edit distance expressed as a ratio of the longer string. A token alignment that pairs each part of the shorter name with its best unused partner in the longer one and averages, so name order is irrelevant by construction. Soundex as Russell and Odell specified it. And an equivalence table of transliteration families, written out in the source so it can be read and argued with rather than trusted. The implementations are written out in full rather than pulled in, because the window and the transposition count are where Jaro-Winkler implementations quietly differ, and a threshold set against one means nothing against another.

Why is Soundex weak here, if you use it?

Soundex was designed for English surnames in American census records. It is poor on exactly the names that matter most in sanctions work, because it was never meant for transliterated Arabic, Russian or Chinese. That is why it contributes a little and never carries a hit on its own, why the equivalence table exists beside it, and why the page tells you when Soundex is the only thing that agreed.

How big a list can it handle?

Comfortably tens of thousands of entries. Below two million possible pairs it compares every one; above that it indexes the list first, scores only the candidates that share a block with the name, and says on the page that it did and what the rule for skipping was. If you need every pair compared on a very large run, split it into smaller runs; the page gives you the numbers to know whether that matters.

Does it decide anything for me?

No, and it should not. It produces a queue ordered by how close the match is and how much the date of birth and country corroborate it, with the reason attached to every entry, and you mark each one a match, escalate it, or clear it, with a note. Those decisions are saved and come out in the download. Screening is a judgement with a record behind it; a tool that makes the judgement for you is one you cannot defend making.

Is any of it sent anywhere?

No. Both files are read by the browser’s own file reader, all the scoring happens in the page, and the downloads are written by the tab. There is no account and no server. Clearing the run leaves neither list in the browser.

Why does the threshold not go below 65?

Because a lower number would be dishonest. Jaro-Winkler is a similarity between two strings, and two entirely unrelated names of similar length routinely score somewhere between forty and fifty-five simply from sharing letters of the alphabet: “Robert James Wilson” against “Patricia Okonkwo” comes out near fifty. A threshold below sixty-five therefore does not mean “be more careful”, it means “return everything”, and a screening tool that offers a setting which quietly matches the whole list is offering a trap. Sixty-five is where the number starts carrying information; the useful working range is seventy-five to ninety.

Doing this often? OBSCURA Practice checks every new matter against all its clients, parties and other sides before it is opened. Everything is US$5 a month or US$48 a year.