There’s growing consumer dissatisfaction with the cost and slow response times in the financial services sector. The industry understands these frustrations but they have many constraints. One of the global responses to the September 11th 2001 terrorist attacks was a series of efforts to exclude terrorists and other criminals from using the financial system. So called “know-your-customer” (KYC) rules have been successful in preventing criminals using standard banking, investment and insurance products. But this is a costly process. Perfecting solutions is a major challenge for practical data scientists and financial technology companies.

Creating clarity

The fight against money laundering, terror financing and sanctioned individuals is based on lists of millions of names of individuals and companies that financial companies must not deal with. Yet even if these extensive, ever changing lists were accurately complied it would be tricky to check them against client databases in near real time. But the names found in client databases often feature inaccuracies such as typos, and are often in non-standardised formats.

A particular challenge is that there are no globally accepted rules for transscription from words written using Chinese, Cyrillic, Arabic, Taiwanese etc. characters. For example, someone with British, Chinese and Egyptian heritage might have their name written “Mohammed Lee-Smith”, “Li-Smyth, Muhamet”, “Mohammed Leesmith”, and so on. Searching for company names can be harder still as there is even less standardisation.

What’s more, there’s no room for error. Nine Luxembourg banks and investment firms were fined and publically named last year by the local regulator the CSSF for having failed to apply anti money-laundering rules. The supervisor has warned the industry that it will continue to take a tough line, not least due to the risk of reputational damage to the country. Globally the punishment can be spectacular. For example, the California-based subsidiary of Coöperatieve Rabobank was hit with a $368.7 million fine in February this year for laundering Mexican drug cartel money. Thus it is no surprise that the industry is willing to throw substantial resources at the problem short term. However, AI is the long-term efficient solution.

300% more efficient

“We conducted experiments in our lab on the Machine Learning KYC solution iDETECT we are perfecting, and saw an improvement in accuracy by a factor of three,” explained Holger Pletsch, Head of Research and Development at LOGOS IT Services. “This is significant because every false positive adds cost as it requires human intervention, and it reduces service quality as processing times are slowed,” he added. They made this breakthrough by improving the accuracy which fuzzy name matches and non-matches are identified. “We did this by combining the best features of existing complementary algorithms with novel features  via Machine Learning, and then making further advances” he explained. “The great advantage of AI systems is that they don’t keep making the same mistakes time and time again,” he added.

Until now system designers have typically employed logics that score the syntactical match between names and hard-coded thresholds for above which names are considered as matching into the system. What makes iDETECT different is how its AI capabilities scan datasets to spot patterns and then automatically learn a model for improved fuzzy name matching.. To capture the globally different structures of names, ethnicity awareness has also been built into the system. “It calculates the most probable the origin of a name, and then applies specific rules for this particularity previously learned from similar data,” Mr Pletsch explained. This further boosts performance and response times.

Transactions and potential fraud

Banks must also keep an eye on “suspicious transactions”. Any bank transfer over €15,000 must be reported immediately to the Luxembourg public prosecutor’s financial intelligence unit. This can be a one-off transfer, or multiple smaller transactions to the same beneficiary. Failure to comply can result in a fine of up to €1.25m. It can be tricky to spot these movements using pre-programmed systems, so often a safety-first attitude is adopted and this can generate substantial false positives and thus costly human intervention.

So far most transaction monitoring system have used a finely tuned set of fixed rules, often which the capabilities to manually update these rules. “The AI in iDETECT is able to spot and dynamically learn transactional patterns of clients and thus can adjust the rules dynamically to massively reduce false alerts” noted Mr Pletsch. A common example is when businesses pay salaries, with some systems failing repeatedly to recognise the benign nature of these payments. The system also keeps an eye out of unusual activity and conducts AI-enhanced economic profiling to detect suspicious transactions which might be fraudulent.

This cutting edge research and development work requires the support of motivated, quality data scientists. Hence LOGOS’ partnership with the University of Luxembourg’s internationally respected Interdisciplinary Centre for Security, Reliability and Trust (SnT). Working with SnT teams helps them make the theoretical and applied research breakthroughs that are needed to drive out financial sector inefficiencies.

This is important work, as the sector strives to retain the trust and loyalty of clients in an ever-more competitive market. The new AI capabilities  of the iDETECT platform further enhance its innovation leading position, with the new commercial release version due for the fourth quarter of this year.