Anti-Money Laundering (AML) legislation has proved challenging for many financial institutions who are struggling to comply in a cost effective and efficient manner. With expenditure on AML activities for banks increasing significantly, can technology provide a solution?
It is estimated that money laundering accounts for a staggering 2-5% of global GDP per annum. While the digitalisation of banking services has undoubtedly brought a host of benefits, it has also made it much more complex to implement efficient AML strategies. But can the data associated with financial transactions be harnessed to better identify those involved in money laundering activity?
The cost of compliance
More stringent and comprehensive regulations against money laundering have left many financial institutions rethinking how they operate in relation to AML. Complying with the regulatory requirements, often in combination with legacy systems, has created inefficient, labour-intensive processes and demanded increased investment for little return.
And while banks have strong routines and experience in assessing credit risk, identifying non-financial risks – at the level required by the EU Money Laundering Directive – is proving to be a challenge. In recent years, for example, there have been a string of high-profile cases involving Scandinavian financial institutions failing to prevent major money laundering activities.
Traditionally, AML detection is based on flagging transactions for manual assessment according to a rules-based system. The criteria specified could be amounts deposited or withdrawn in cash, certain types of transfer (such as those to countries considered “high risk”) or the frequency of transactions.
Machine learning can increase AML effectiveness and reduce costs
So, how can we tackle money laundering and comply with legislation in a way that is more effective and efficient, ultimately allowing for a better use of resources?
Many banks are trying to answer this question by turning to more advanced methods of detecting suspicious behavior, particularly in the form of artificial intelligence and machine learning. There are opportunities to both improve the monitoring which is already in place using supervised machine learning models, while simultaneously discovering new trends in suspicious behavior by developing unsupervised machine learning models to explore the unknown.
However, financial institutions as well as regulators have historically been reluctant to use unsupervised learning technology. This is due to unsupervised methods (as well as many supervised methods) being of a “black box” nature that produces results that are hard to explain rationally. A natural first step therefore – to demonstrate the capabilities of machine learning in relation to AML – can be to employ explainable supervised machine learning to increase the efficiency of existing surveillance processes. More advanced techniques such as unsupervised machine learning and network analysis can then be implemented to increase the overall effectiveness of the surveillance.
The application of new technology in AML varies between FIs
Our experience from working with financial institutions to tackle the issue of money laundering shows that many are still in the phase of employing traditional rule-based engines. We are however seeing an advance in the use of more sophisticated techniques such as supervised machine learning to improve transaction monitoring. As advanced analytics maturity increases financial institutions grow from using only rule-based methods to deploying independent advanced models, including supervised machine learning. In the most mature institutions, artificial intelligence, rule-based methods and network analysis are integrated in a hollistic system for optimal results, thus reducing compliance costs.
Illustration: A natural progression in the deployment of advanced analytical models to aid in anti-money laundering.
What was the outcome when we applied machine learning to the AML process?
BearingPoint has assisted Nordic banks to explore the initial opportunities when applying supervised machine learning to transaction and customer monitoring. We find that when supervised machine learning is applied to the AML process, and we allow the algorithm to refine the criteria used to identify potentially fraudulent activity, we immediately see an improvement in process efficiency.
The success of the algorithm is based on its ability to create new rules using an iterative process based on the outcome of expert, manual evaluations and monitor for suspect transactions or customers. A user-friendly dashboard is then used to graphically present key information on the flagged individual, further streamlining the evaluation process.
Illustration: Flow diagram showing Supervised Machine Learning (SML) and the AML process
Our own experience indicates that when this model is applied, the number of relevant flags is increased while the overall number of alerts is reduced. This entails less manual processing of flagged customers and a higher rate of Suspicious Activity Reports to regulatory authorities per manual evaluation.
For a traditional rules-based system, let’s say all cash withdrawals above 50.000 NOK are considered suspicious, which implies that all customers above the dotted line for 50.000 NOK in the illustration below would be flagged. However, as we can see, only a limited share of the customers has been reported historically. Machine learning helps identify additional criteria that can be applied to the rule, and thus narrow the scope to capture suspicious customers in a much more efficient manner while we are avoiding plenty of false positive alerts.
Illustration: Graph showing identified risk groups and description
Having applied this methodology to banks in Scandinavia we have been able to achieve the following results. For the retail side of “Bank A” more than 18% of flagged customers were reported to the Norwegian Financial Intelligence Unit (Økokrim), 12% of which were new cases. It is also worth noting that 50% of those cases flagged were classified as relevant by the bank’s investigators.
In another case where supervised machine learning was applied to the process, a 4-fold increase in alerts classified as relevant was achieved while at the same time reducing the number of overall alerts by 25%.
In summary, our experience shows that there is a very clear opportunity, today, to use data and technology to increase the efficiency and effectiveness of AML processes. If you would like to find out more about how we can help your organization reduce costs in the short term and harness technology to get an even better return in the medium term, feel free to contact us!
