I think you all know that correspondent banks are required to meet specific regulatory obligations while maintaining their correspondent relationships, as well as meet general compliance obligations to report suspicious activity, prevent money laundering, and comply with economic sanctions. A mouthful but true, nonetheless.
Although the Legacy AML and traditional AI systems, in place at banks monitor customer activity, this approach is not effective for monitoring flow of funds that aren’t related to the bank’s own customers. Moreover, these systems are based on rules that incorporate preset scenarios with conditions and thresholds looking for patterns that are known.
We do it differently. Our IntuitiveAI for correspondent banking analyzes SWIFT messages without setting any predefined condition or threshold. Historical data is used to learn what is “normal” in terms of transactions and data flow and then detect anomalous activity with respect to that normal. This approach enables complex patterns of behavior to be caught which otherwise would have been missed by legacy or traditional AI systems.
At the core of the solution is the data scientist
It’s up to them to set the analysis strategy. The strategy includes selecting the relevant data sources and calculation of “features” — the key parameters and ways of deriving information specifically relevant for correspondent banking.
ThetaRay’s IntuitiveAI algorithm then analyzes these features altogether to define what’s normal and what’s abnormal activity, without setting any predefined patterns or thresholds. The key data sources utilized are the SWIFT messages (MT103, 202, 202C), KYC information, Country Risk and bank risk information. These data sources are extracted over a historical period of a few years and are then utilized to calculate an entity called the “Full Path.”
The Full Path covers the end-to-end payment flow and banks participating in the payment. Once the Full Path is derived from the SWIFT messages, several features are calculated. The features designed are related to volume and value changes in the Full Path activity over time, various combinations of risk indicators of the sources and destinations of funds, banks involved and the structure and location of the bank codes in the path.
Stay with me now.
An example of such a feature is the total amount of SWIFT 103 messages that have been sent over a Full Path during a specific time period and with a specific currency. We then take that total amount and compare it to similar time frames in the past in a form of a statistical calculation. Then, this calculation is factored with another calculation which indicates how significant this total amount is compared to other similar paths during the same time period. These two calculations are then combined into a single feature value. Several of these smart features are calculated and then injected altogether into the algorithms.
Still here?
Once features are created, an analysis is conducted and fine-tuned. Results worthy of further investigation are turned over to analysts to then take that investigation forward.
On top of all this we have to ensure that we do what we say we’re going to do:
• How many and what the quality of investigation worthy alerts were generated?
• How efficient was the detection effort?
• Operationally how effective was the system in helping to determine alert resolution?
• How flexible was the system to adjust to changing requirements?
Not to brag, and speaking honestly, I’ve had the opportunity to work with many systems but ThetaRay is the only one that catches financial crime in a correspondent banking network as well as it does within the “walls” of the bank.
Not to brag, and speaking honestly, I’ve had the opportunity to work with many systems but ThetaRay is the only one that catches financial crime in a correspondent banking network as well as it does within the “walls” of the bank.