Feedzai has recently introduced RiskFM, a specialized foundation model designed for financial crime detection and risk decision-making.

This model is crafted to cover fraud detection, AML (Anti-Money Laundering), and broader risk assessments throughout the financial crime lifecycle, moving beyond traditional rule-based and manually engineered machine learning approaches that have long dominated the sector.

Unlike existing methods which are limited to card network data, RiskFM has been trained on a comprehensive dataset encompassing various aspects such as onboarding, digital activities, payments, transfers, and AML processes across different regions and institutions.

Facing the challenge of transactional behavior

The primary hurdle in applying foundation model techniques to financial data is the unpredictability of transaction patterns. Unlike structured text or visual data where adjacent elements follow recognizable patterns, financial transactions vary widely based on consumer behavior, payment methods, and fraud tactics. Moreover, financial risk operates within a competitive environment where criminals continuously adapt their strategies to bypass detection systems.

This complexity makes it significantly more challenging to apply large language model principles to financial crime compared to text or image data analysis. RiskFM addresses this challenge by utilizing native tabular structures common in financial systems rather than adapting architectures from other domains.

Performance and deployment implications

In tests, RiskFM is expected to match the performance of fine-tuned specialized models when used within a single institution without requiring manual feature engineering. Training across multiple institutions simultaneously reportedly outperforms traditional gradient boosting and deep learning techniques, with performance improving as more data is ingested.

For financial institutions, this means faster deployment times and lower costs compared to custom-built solutions. Feedzai has developed RiskFM as a unified layer covering various use cases from mule account detection to AML, designed for scalability across an institution’s full risk infrastructure.

IDC, an industry analyst firm, noted that RiskFM represents a significant effort to integrate foundation model capabilities into a domain that has historically been resistant to such approaches. The capability to match bespoke supervised models without manual feature engineering has clear advantages in terms of deployment speed, cost, and coverage across the financial crime lifecycle.

Collaborative development and future outlook

Lloyds Banking Group, a UK-based financial institution, is cited as a collaborative partner in RiskFM’s development. A company representative referred to RiskFM as part of ongoing AI-focused initiatives aimed at preventing economic crimes.

Feedzai states it is currently working with early adopters to validate initial RiskFM frameworks and plans to scale these methodologies across larger datasets. Future integration into its full suite of use cases is also in the works.