In a world inundated with data, traditional methods of financial crime detection are becoming inadequate.
Financial crime remains a challenge for financial institutions (FIs) worldwide, evolving in complexity and scale with each passing day. This shifting threat landscape means that advances and innovations designed to give detecting, preventing and combating financial crime a shot in the arm are increasingly necessary.
Wolfgang Berner, the co-founder and CPO of Hawk, unpacked for PYMNTS the opportunities that large transaction models (LTMs) — generative artificial intelligence (AI) models adapted to financial crime — represent in establishing more robust, accurate and comprehensive detection and prevention mechanisms.
“The core idea is we treat transactions as sentences, teaching the transformer model the language and grammar of transactions, similar to how large language models like GPT-4 are trained on the text of the web,” Berner explained. “And by doing that, it develops a very good understanding of the transactions, how transactions relate to each other, and what is genuine or possibly suspicious with them.”
The model is trained so that it excels at finding novel criminal activities and reducing false positives, identifying authentic activities and illicit ones alike by detecting patterns and anomalies.
“The LTM is a booster for what we’ve been doing over the years, enhancing our ability to reduce false positives and find more crime,” Berner added.
As financial criminals become more sophisticated, traditional rule-based or machine learning approaches are insufficient. Berner highlighted two advantages of LTMs: higher resolution and natural language understanding.
“LTMs see transactions in their entirety without aggregation, maintaining a broad view and understanding language data within transactions, such as reference texts indicating intent,” he emphasized.
These capabilities allow LTMs to detect complex relationships and patterns, providing an edge over traditional models.
“The LTM achieves a clearance and sharp vision akin to a bird of prey, seeing transactions in their original form with all attributes intact,” Berner said.
Understanding the technical foundation of LTMs sheds light on their effectiveness. Transactions, though structured, contain various attributes that must be analyzed in detail. The attention mechanism, a key feature in both large language models and LTMs, plays a critical role.
“The attention mechanism helps the model focus on relevant transactions and elements within them, similar to how an LLM connects words and phrases across sentences,” Berner explained.
This mechanism enables the LTM to detect correlations between transactions, such as those indicating money mule behavior, while ignoring genuine transactions. Additionally, LTMs work on an individual transaction level, maintaining high resolution and clarity, which traditional models, often reliant on data aggregation, can miss.
Read more: Hawk Updates AI Platform to Enhance Financial Crime Detection
Seven in 10 FIs are now using AI and machine learning (ML) to fend off fraudsters, according to the PYMNTS Intelligence and Hawk collaboration, “Financial Institutions Revamping Technologies to Fight Financial Crimes.”
Testing the efficacy of such a sophisticated model as an LTM is crucial. Berner shared Hawk’s approach, emphasizing the importance of practical application and benchmarking against traditional models. One test involved comparing Hawk’s LTM-based false positive reduction model with a traditional one.
The results were promising, with the LTM achieving a 30%, or 15 percentage point, improvement in false positive reduction, boosting the rate from 50% to 65% while maintaining the same level of precision.
Maintaining the accuracy and reliability of predictions is paramount in financial crime detection.
“We track model output KPIs [key performance indicators] and validate results through automatic sampling. This process, while similar to traditional models, is enhanced by the LTM’s higher resolution and anomaly-detection capabilities,” Berner said.
Looking ahead, LTMs are poised to evolve further, addressing new challenges and use cases in financial crime prevention.
“LTMs will serve as a foundational model, evolving through specific use cases and continually improving existing applications,” Berner predicted, noting that Hawk’s LTM can act as a core infrastructure layer for new applications to be built atop.
One exciting development is the application of generative AI in screening for sanctions and identifying sanction circumvention through semantic matching.
“By understanding context and intent, the LLM can relate different attributes, enhancing our ability to detect sophisticated financial crimes,” Berner said.
As the strategies and tactics leveraged by financial criminals continue to evolve, so too must the technologies designed to thwart them, ensuring a more secure financial landscape.
We’re always on the lookout for opportunities to partner with innovators and disruptors.
Learn More