Sanctions name screening fuzzy matching validation of Arabic names is one of the least tested parts of a screening programme in the Kingdom. A tool usually performs well in a vendor demonstration, because the demonstration uses clean examples. Real customer and payment data is different. One Arabic name may arrive in several Latin spellings, with the words in a different order, with a part omitted, or with a prefix joined or separated. This article describes a validation method that a head of sanctions compliance or a model-validation lead can run and defend.
Why Arabic names are a distinct test problem
Arabic has no single standard for writing names in Latin letters. Passports, bank records, SWIFT messages and sanctions lists each follow their own habits. The same name can appear as Mohammed, Muhammad or Mohamed, and the same family name can appear with or without the definite article, as a single token or as two.
Several features make this harder for a matching engine:
- Vowels are often not written in Arabic, so Latin vowels vary freely.
- Some Arabic sounds have no Latin equivalent, so letters and digraphs differ (kh, gh, q, dh, th).
- Long names use a chain of given name, father's name, grandfather's name and family name, and any link may be dropped.
- Prefixes such as al-, bin and abu may be attached, detached or omitted.
- Source lists may carry names in Arabic script, in Latin script, or both.
Banks in the Kingdom must screen against UN, local, OFAC and EU regimes, and the sanctions requirements sit within the SAMA AML/CTF rules and the FATF Recommendations. A programme that screens thoroughly but matches poorly on Arabic names carries a risk that its own reporting will not show.
Step 1: Define scope before testing
Write down what is being validated. Name the lists in scope, the customer and payment flows screened, the fields compared (full name, aliases, original script), and the system's current matching configuration. Record which list sources are loaded in Arabic script and which in Latin script only. This scope statement becomes the first page of the validation record.
Step 2: Build the known-positive test file
The core of the method is a file of names that should match, because you already know the answer. Draw the base entries from the lists you screen against, then create controlled variants of each. The test file should include:
- Transliteration variants: alternative Latin spellings of each name, including common vowel and consonant substitutions.
- Word-order variants: surname first, given name first, and the order used in payment messages.
- Partial-name variants: names with a middle link dropped, or only two or three of four components present.
- Prefix and spacing variants: al- attached, detached or removed; compound names joined or split.
- Script variants: the Arabic-script form against the Latin-script list entry, where your data holds both.
- Noise variants: typing errors and truncation of the kind seen in real payment data.
Label every record with its base entry, the variant type and the reason it should match. Have an Arabic-literate reviewer confirm that each variant is plausible, since a variant no human would write inflates the miss rate unfairly.
Add a set of known-negative names too: common Arabic names that resemble list entries but belong to different people. Without them you can measure detection but not the alert volume that a lower threshold creates.
Step 3: Run the file at every threshold
Run the full file through the screening engine at each threshold setting you could reasonably adopt, not only the current one. For each run, record for each variant type whether the base entry was returned, at what score, and in what rank among the results.
Report detection by variant type rather than as a single figure. A blended result can hide the fact that transliteration variants are found while partial names are not. Then run the known-negative set at the same thresholds to count the false alerts each setting produces. The two results together give the trade-off the business is actually choosing.
Step 4: Separate vendor settings from engine limits
When a name is missed, the cause is one of two things, and the remedy differs.
A configuration cause is something the institution can change. Examples are an algorithm that is disabled, a name-type setting that is wrong, a word-order option that is switched off, a transliteration or phonetic module that is not enabled, or a threshold set too high for a given field.
An engine limit is something the tool cannot do under any available setting, for instance no handling of Arabic-script comparison or no way to match a name when a component is absent.
To separate them, re-run each missed record after changing one setting at a time and note whether it is recovered. Keep a log of each change. Misses that no setting recovers are engine limits and belong in the model's limitations register, with a compensating control such as list enrichment with additional aliases, a secondary screening pass, or manual review of specific payment types. Misses that a setting recovers become tuning recommendations.
Step 5: Check the data the engine receives
A matching engine can only compare what it is given. Test whether names are truncated on ingestion, whether fields are concatenated, whether Arabic script is preserved or converted, and whether aliases from the lists are loaded. Poor results are frequently a data problem rather than a matching one, and the test file will not reveal this unless the input path is part of the test.
Step 6: Decide the threshold and document it
The threshold decision should be a documented risk decision, not a vendor default. The record should contain:
- the test file design and how variants were created and reviewed;
- detection results by variant type at each threshold;
- false-alert volumes from the known-negative set at each threshold;
- the configuration changes tested and their effect;
- the engine limits found and the compensating controls;
- the chosen threshold, the reasons for it, the residual risk accepted and who approved it;
- a date for re-testing, and the events that trigger an earlier review, such as a list change or a system upgrade.
Where an institution uses an above-the-line and below-the-line approach to tuning, test below the proposed threshold as well, so that the evidence shows what is being left unalerted and why that is acceptable.
Common weaknesses in the method
Validators often test with names the vendor supplied, which the engine has effectively been tuned on. Others test only exact Latin spellings from the list, or report a single detection rate. Some change several settings at once and cannot say which one worked. Each of these produces a result that will not hold up when SAMA or an internal audit function asks how the threshold was chosen.
How this connects to our work
Our AML Monitoring and Screening Model Validation service is an independent validation of transaction monitoring scenarios and name-screening tools, covering data quality, thresholds, fuzzy-matching performance and coverage of your risk assessment. The work follows four stages: data, coverage, performance and governance. Screening match-quality testing is one of the service's listed features, alongside data lineage and completeness testing. It is delivered as a project of 6–10 weeks, and the deliverables are a validation report, tuning recommendations with evidence, and a model inventory and governance pack.
On access, we work on extracts in your environment or a secure copy, so data never leaves your control. Validation includes evidence-based tuning recommendations, and we can support implementation. If your concern is wider than the screening model, our Sanctions Compliance Program service reviews screening, trade and payment controls, escalation and reporting across UN, local, OFAC and EU regimes.
If you are preparing to validate your screening tool or need to document a threshold decision, we can discuss scope and timing with your sanctions and model-risk teams.
