The Problem with Conventional Screening#
Conventional watchlist screening systems generate excessive false positive rates — often exceeding 95% of all alerts produced — because they rely on simple string matching. String matching cannot distinguish between a sanctioned party and a customer who happens to share a similar name. When your screening system produces a hundred alerts for every genuine match it finds, the consequences are predictable: compliance teams are overwhelmed, legitimate transactions are delayed, and the quality of alert review declines as reviewers work through high volumes with insufficient time to assess each case properly.
In emerging markets, this problem is significantly worse than it is in developed market contexts. Name diversity across Sub-Saharan Africa, South Asia, and Southeast Asia is enormous. Transliteration from Arabic, Hindi, Swahili, Bahasa Indonesia, and dozens of other scripts into Latin characters is inconsistent — the same name can appear in multiple romanised forms across different documents, databases, and screening lists. Common names — both given names and family names — are shared by very large proportions of the population in many of our focus markets. A screening system that cannot account for this will generate alert volumes that are unmanageable for any compliance team, regardless of its size.
Anqa AML Smart Screen was built to solve this problem.
The Anqa Approach#
Anqa AML Smart Screen uses a multi-algorithm approach that combines several distinct matching techniques with contextual data analysis. Rather than applying a single method and accepting its limitations, Smart Screen runs multiple algorithms simultaneously, aggregates their outputs, and combines the results with demographic and geographic context to produce a composite match score that is substantially more accurate than any single method alone.
The result is a screening system that identifies genuine sanctions matches with high reliability while eliminating the majority of false positives that burden conventional systems — reducing alert review workload by up to 65% compared to string-matching approaches.
Technical Methodology#
Algorithm 1: Levenshtein Distance#
Levenshtein Distance calculates the minimum number of single-character edits — insertions, deletions, or substitutions — required to transform one string into another. In screening terms, this means the system can identify likely matches even where there are minor spelling variations, typographical errors, or inconsistencies in transliteration.
A sanctioned individual whose name is romanised as “Mohammed Al-Rashid” in the screening database may appear as “Mohamed Alrashid”, “Mohammad Al Rasheed”, or “M. Al-Rashid” in customer records. Simple string matching will not connect these variants. Levenshtein Distance analysis will, identifying the edit distance between each variant and the listed name and flagging those that fall within the configured threshold for human review.
Algorithm 2: Soundex#
Soundex is a phonetic algorithm that encodes names based on how they sound rather than how they are spelled. It was originally developed for English-language name matching, but its core insight — that names which sound the same should be treated as potential matches regardless of spelling — is directly applicable to the transliteration challenges common in emerging market screening.
Names from Arabic, South Asian, and African language traditions can be romanised in ways that look very different on paper but are phonetically near-identical. Soundex captures these matches by reducing each name to a phonetic code and comparing codes rather than character strings. This approach is particularly effective for identifying matches across different romanisation conventions for the same underlying name.
Algorithm 3: Metaphone#
Metaphone is an enhanced phonetic technique that models the phonological patterns of names with greater precision than Soundex. While Soundex applies relatively simple encoding rules, Metaphone accounts for more complex phonological patterns — consonant clusters, vowel elision, and pronunciation conventions that vary across language families.
For the name diversity encountered in emerging market screening — Arabic, Swahili, Hindi, Urdu, Bahasa, Sinhala, and many others romanised into Latin script — Metaphone provides significantly improved accuracy over Soundex alone, particularly for longer names and names with complex consonant structures. Smart Screen runs both algorithms in parallel, using their combined output to strengthen the phonetic matching signal.
Algorithm 4: Contextual Data Analysis#
Name matching alone, even with sophisticated phonetic and edit-distance algorithms, cannot eliminate false positives when names are genuinely common. The fourth component of Smart Screen supplements name matching with additional data points that significantly reduce ambiguity.
Contextual data used in Smart Screen analysis includes date of birth, nationality, country of residence, and entity type. Where customer records contain this information and the sanctions list entry also contains it, Smart Screen applies it to refine the match score. A customer named “Ibrahim Hassan” born in 1985 who is a Kenyan national resident in Nairobi scores very differently against a sanctions entry for “Ibrahim Hassan” who is a Somali national with a 1978 birth year — even though the name match is strong. Contextual analysis makes that distinction explicit and translates it into match scoring.
The Six-Step Screening Process#
Step 1: Data Normalisation#
Before matching begins, customer data is normalised to a consistent format. This includes standardising name presentation, removing diacritical marks and special characters, resolving common transliteration variants to a canonical form, and handling name component ordering differences across different naming conventions. Normalisation ensures that the matching algorithms operate on comparable data and reduces the artificial mismatch that occurs when different data entry conventions produce superficially different records for the same individual.
Step 2: Multi-Algorithm Screening#
All four matching algorithms run simultaneously across the consolidated watchlist database — which combines UN, OFAC, EU, HM Treasury, and applicable regional sanctions designations. Each algorithm produces its own match score independently. Running the algorithms in parallel rather than sequentially ensures that each method’s output is uncontaminated by the others, and that the combination step in Step 4 is working with genuinely independent signals.
Step 3: Contextual Analysis#
Where contextual data is available — in customer records, in sanctions list entries, or in both — it is applied at this stage to refine the raw match scores from Step 2. The contextual analysis layer uses demographic and geographic data to calculate the probability that a given name match represents the same individual as the listed person, taking into account the prevalence of the name in the relevant population and the consistency of other identifying information.
Step 4: Intelligent Scoring#
The outputs from Steps 2 and 3 are combined into a single composite match score. The combination model is calibrated to the institution’s risk appetite and the characteristics of the screened population — a higher weight on phonetic matching for institutions screening predominantly Arabic-origin names, for example, or a lower contextual data threshold for institutions where customer records contain limited demographic information. The resulting composite score reflects all available evidence and is calibrated to produce a meaningful signal-to-noise ratio rather than simply maximising sensitivity at the cost of specificity.
Step 5: Human Review#
Alerts above the configured threshold are routed to the compliance team through a structured decision support interface. Each alert presents the match rationale — which algorithms flagged the match, the specific scores, the contextual factors applied — alongside the customer record and the relevant sanctions list entry. This transparency enables reviewers to make informed decisions efficiently, rather than working from an opaque system output that provides no basis for independent assessment.
Step 6: Continuous Learning#
Smart Screen incorporates reviewer decisions — both confirmations of genuine matches and disposals of false positives — into its ongoing scoring calibration. Over time, the system learns the specific name patterns and contextual factors that generate false positives in the institution’s particular customer population, and adjusts its scoring accordingly. This continuous learning loop progressively improves accuracy without requiring manual reconfiguration.
Performance#
Anqa AML Smart Screen reduces alert review time by up to 65% compared to conventional string-matching screening systems, while maintaining robust detection of genuine sanctions matches. For compliance teams managing significant screening volumes with limited staff, this represents a material improvement in both compliance quality and operational efficiency.
Configurable Thresholds#
Institutions can set matching sensitivity thresholds appropriate to their specific risk profile and regulatory obligations. The appropriate threshold will vary by context: correspondent banking relationships warrant lower thresholds — higher sensitivity, more alerts — because the consequences of a missed match are severe. Retail transaction screening in a high-volume, low-value context may appropriately use higher thresholds to manage alert volume while maintaining adequate coverage. Smart Screen supports this differentiation, allowing institutions to apply different threshold configurations to different screening contexts within the same platform.
All threshold configurations are documented within the platform, providing the audit trail that regulators expect when examining an institution’s screening controls.