Implementation of anti-money laundering (AML) and fraud detection rules is always a challenge for financial organisations, not only from a technical, but also from the client-facing perspective. Monitoring needs to strike a balance between strict inspection of suspicious transactions, and allowing genuine transactions through. Failure to do so would potentially endanger the overall client experience.
To maintain this balance, and a positive overall UX, financial institutions employ algorithmic approaches, rule engines, manual inspection, artificial intelligence solutions and a mixture of these approaches in various proportions. Unfortunately, the perfect solution, which would detect all fraudulent transactions and only fraudulent transactions does not yet exist. One could only aim at achieving the best possible balance. How, then? Our suggestion to getting one step closer is to employ NLP (natural language processing) for a better and more robust understanding of the real transaction geography involved.
Step One – Identify / Discover any Geographical Mismatch
An overwhelming majority of transactions are executed using SWIFT messaging. Almost any bank branch in the world is uniquely named with a structured SWIFT code, which includes information about the bank code, domicile country and branch code. With this efficient and unambiguous encoding, each transaction is typically delivered to the recipient without delay.
Figure 1 Source: https://marketbusinessnews.com/FINANCIAL-GLOSSARY/SWIFT-CODE/
As great as SWIFT codes are, they do not offer the full picture of the transaction geography, as there is no information of the actual country of the counterparty residence. There is no guarantee that the recipient company or individual resides in the same country as their bank. In fact, from our experience of working with such data, a residence mismatch can happen in as many as 60% of of transactions. This poses a great challenge for financial organisations striving to stay compliant with AML, cross-border and other international regulations, as it is actual country of residence that matters most for these rules. An ‘imperfect’ or facile approach would be to use the easily extractable bank domicile from the SWIFT code and claim that this is “the best effort”.
Step Two – Detect Country from Address
As you might have guessed, it is not the best effort, and there is a better one. In 2016, stricter requirements were introduced for transaction messages accompanying wire transfers. Namely, they now need to include the name and domicile address of the beneficiary. So, the specific detail of what organisations need to properly implement AML rules. And precisely where Natural Language Processing (NLP) can add value.
We have recently discussed the capacity of NLP for text analysis and for the extraction of important information. However, there are two challenges here. First, transaction information is entered by a person initiating the transaction. Consequently, it is not fully reliable, and can contain typos and misspellings. Second, addresses are a very special type of natural language expressions: they don’t have nouns, verbs, adjectives and other sentence members that are familiar to most NLP models.
In our upcoming webinar we will show how we solve these issues, to be able recognise country information from a transaction message. What we achieved is annotation of Geo-Political Entities (GPE) in addresses and consequent correct country recognition in 95% of the addresses.
In this blog, we have explored how data analytics in the form of NLP analysis of transaction messages can enhance the understanding of transaction geography, and in turn improve AML compliance. In the third of this three part blog series, will focus on high-level transactional behaviour analysis and account clustering.
Join us for a two-part webinar series in which we explore the concrete application of Artificial Intelligence to KYC and AML. Select the most convenient session for you and register using the links below. See you there!
Register here Part One – Using the power of semantic search to build comprehensive client risk profiles and automate continuous KYC monitoring
Register here Part Two – Applying advanced data analytics to implement effective transaction monitoring and improve the quality of KYC files