Feedzai has unveiled RiskFM, a tabular foundation model purpose-built for financial crime detection and risk decisioning.
The model is built to span fraud detection, AML, and broader risk decisions across the financial crime lifecycle, moving away from the rules-based and manually engineered machine learning models that have dominated the sector for decades.
Unlike existing approaches that Feedzai says are limited to card network data, RiskFM has been trained on a dataset covering onboarding, digital activity, payments, transfers, and AML workflows across multiple geographies and institutions.
Addressing the challenge of transactional data
The fundamental obstacle in applying foundation model techniques to financial data lies in the unpredictability of transactional behaviour. Unlike language or image data, where adjacent tokens or pixels follow recognisable patterns, financial transactions vary continuously based on consumer behaviour, payment type, and fraud methodology. Feedzai also notes that financial risk operates in an adversarial environment, where fraudsters actively adapt their methods to evade detection systems in real time.
With this in mind, this makes the application of large language model (LLM) logic to financial crime substantially more complex than its application to text or visual data. RiskFM is positioned as a response to that challenge, using tabular data structures native to financial systems rather than adapting architectures from other domains.
Performance benchmarks and deployment implications
In testing, RiskFM is expected to match bespoke, fine-tuned supervised models when deployed for a single institution, without the need for manual feature engineering. When trained across multiple institutions and geographies simultaneously, the model is reported to outperform traditional gradient boosting and deep learning approaches, with performance improving as it ingests additional data.
The practical implications for financial institutions include faster deployment timelines and lower implementation and maintenance costs compared with custom-built models. Feedzai developed RiskFM as a unified layer covering use cases from mule account detection to AML, intended to scale across an institution's full risk infrastructure.
Industry analyst firm IDC noted that RiskFM represents a credible attempt to bring foundation model capabilities to a domain that has so far proved resistant to that approach. An IDC representative also observed that the ability to match bespoke supervised models without manual feature engineering has tangible consequences for deployment speed, cost, and coverage across the financial crime lifecycle.
Early adoption and next steps
Lloyds Banking Group, a UK-based financial institution, has been cited as a collaborative partner in the model's development, with a company representative describing RiskFM as a development in ongoing AI-focused work on economic crime prevention.
Feedzai states it is currently working with early adopters to validate initial RiskFM frameworks and intends to scale these methodologies to larger datasets, with plans to integrate the model across its full suite of use cases.
