Most financial institutions agree that their current AML transaction monitoring systems false positive rates range from 85% to 95%.
These results beg the question of why do something that has such a horrendous result. False positives represent a real issue because they are expensive to manage and take your eye off the ball of investigating suspicious transactions indicative of financial crime.
Many AML solution providers are applying machine learning techniques to reduce the number of false positives and to filter alerts that can suggest an automated alert disposition, significantly reducing human effort.
One area in which machine learning can help is behavioral analysis, which establishes customer profiles based on a combination of indicative and transaction data. The algorithms can troll through much larger volumes of data than humans and establish a more significant number of profiles at a more granular level. This allows a higher degree of sensitivity in detecting potential deviations in behavior. Criminals know that their transaction behavior is being monitored and are constantly looking for techniques that help them avoid detection, like integrating their illegal transactions with many legal transactions.
The challenge with greater sensitivity is a higher likelihood of false positive generation, and as mentioned above, it could significantly degrade the benefit of using AI-driven behavioral analysis.
One novel area of research is Social Network Analysis (SNA), which calculates a client or counterparty’s predicted risk based on social network metrics. SNA analyzes the patterns of connections, interactions, and dependencies between individuals, groups, or organizations within a social network.
From a financial crime perspective, the nodes on the network are the client and institutional entities, and the edges are the transactions between the entities. A typical analysis technique evaluates the centrality of specific nodes and the clusters of nodes on the network. The challenge with this is these are often more likely to identify high volume transactors who are engaged in legal activities. The key is to find the illegal activity that is trying to ride alongside undetected.
A team in Italy took a different approach, looking at other factors at both the client and transaction levels. They combined transaction frequency, transaction size, regional demographic data, and beneficial owner data. They found that criminal transactions often involved a lower-volume transactor conducting higher-value transactions spanning different geographies and involving some commonality across the beneficial owners.
While this research used a small sample size in a single country, it shows the potential of using advanced social network analysis to improve the quality of machine learning transaction monitoring techniques.
