Firms, large and small, need to navigate a set of increasingly complex compliance rules and regulations as regulatory bodies clampdown on loopholes in the financial regulatory framework. With tighter regulation comes the need to seek out more advanced and cost effective compliance solutions.
Despite the introduction of tougher legislation over recent years, money-laundering and financial scandals continue to dominate global news. It is estimated by the Financial Action Task Force that over one trillion dollars is laundered annually. As the financial services sector falls under increased scrutiny, banks are mandated to implement full measures to prevent financial crimes. Regulators increasingly require greater oversight from institutions, including closer monitoring for anti-money laundering (AML) and know your customer (KYC) compliance.
Unlike fraud, where historical data is available, there are not many verified historical cases of money laundering; furthermore, even if these historical cases did exist, they would not prove to be particularly useful in detecting money laundering. The methods and tactics used to launder money are constantly evolving, from loan-back schemes and front companies, to trusts and black market currency exchanges, there is no “typical” money laundering case.
The unique nature of money laundering requires a new generation of solutions. Business rules, a technology familiar to regulators all over the world, can be empowered with fuzzy logic, supervised and unsupervised learning, and other machine learning technologies to increase detection and reduce false positives. To implement a next generation solution for BSA/AML, firms must look towards updated machine learning tools that allow finer grain resolution at the scale needed to detect AML.
Current legacy technologies for detecting money laundering are based on business rules, neural networks, or statistical approaches (e.g. data mining). Business rules written, tested, and deployed by experts in the field create a first-pass detection method for known types of malicious behavior. In the case of fraud prevention, a well-defined classification model can be designed by learning from the patterns of former fraudulent behaviors. These technologies need massive amount of historical data to extract some usable information for future uses. Due to the nature of fraud, large repositories of confirmed fraudulent behavior are available to train these supervised technologies; however, with money laundering there is no labeled data to learn from.
In addition to the lack of historical information, each money laundering case is unique; however, legacy approaches apply the same logic to every entity. Relying on only these approaches results in low detection and high false positive rates.
The high false positive rates associated with current AML approaches result in compliance costs that can quickly spiral out of control, due to the amount of manpower required to investigate the large number of alerts generated by legacy technology. Additionally, deploying legacy models often requires expensive hardware and databases, and the resulting solutions are unable to provide real-time performance.
Unsupervised Learning Technology
Unsupervised learning is learning from unlabeled data, where particularly informative privileged variables or labels do not exist. As a result, the greatest challenge is often to differentiate between what is relevant and what is irrelevant in any particular dataset. In the context of classification, the goal is to divide a set of unlabeled data into classes, or clusters. Unsupervised learning also encompasses dimensionality reduction, feature selection, and a number of latent variable models. While first-pass solutions often use business rules, the combination of these painstakingly tested and verified rules with the power of unsupervised learning technology empowers these initial solutions with far greater accuracy.
As historical data related to money laundering is scarce and unreliable, it is vital to utilize unsupervised learning technologies which have the ability to gain insight from the data without any prior knowledge of what to look for. Unsupervised learning utilizes temporal clustering, link analysis, associative learning and other techniques to allow customers to track transaction volatility, entity interactions, behavioral changes, etc.
The power of unsupervised learning for detecting money laundering shines when data from a multitude of sources can also be ingested by the system. Having a system flexible enough to accept multiple data points across a variety of sources is essential in tracing the full behavior of the individuals and the money/assets laundered.
The complexity of money laundering makes it important to utilize a technology that has the ability to track the behavior of each individual involved and link them to one another. One such AI technology is Smart Agents. Smart Agents are virtual representations of each entity in the system and work collaboratively to gather collective intelligence. These agents are in charge of analyzing the behavior of each specific entity (card holder, merchant, account, computer, etc.) and its activities over time. The various Smart Agents will communicate between one another to provide a 360 degree view of all the activity involved in a laundering scheme.
In conclusion, a comprehensive AML solution must combine the benefits of Business Rules, as well as Fuzzy Logic and Smart Agent technology, to reduce false positives and also incorporate supervised and unsupervised learning to increase detection.