RiskFM: From Custom Models to Foundation Intelligence

Imagine going up to a coffee shop counter and saying, “I’d like a…” It’s relatively easy to guess the next word. A latte, an espresso, or a tea are all safe predictions. This is how large language models (LLMs) like ChatGPT and Gemini are able to “solve” word and image patterns and predictions so quickly, because they have built a solid foundation that makes it easy to anticipate what’s next in the sequence. 

While language has a fixed alphabet, predicting what transaction a consumer will perform next is much harder to get right. You might be buying a coffee now, but your next purchase could range anywhere from a cookie to a car ride to the airport. A wide range of factors including spending habits, limits, or payment methods makes anticipating the following activity more challenging.

That’s why Feedzai has developed the Risk Foundation Model (RiskFM), the industry’s first Tabular Foundation Model purpose-built for risk decisioning across the entire entire financial crime lifecycle, covering multiple use cases including fraud, anti-money laundering, and identity. Currently in its research phase, RiskFM has already demonstrated the ability to autonomously learn behavior patterns across vast datasets, achieving parity with custom-built models on day one without the effort of labor-intensive feature engineering.

In this post, we’ll dive into how RiskFM provides banks with the next generation of AI-led fraud detection, and why this approach is the key to outsmarting the next generation of threats.

Key Takeaways

• While large language models (LLMs) can learn the foundation of language making it easy to understand and predict a sequence, financial risk is significantly more challenging. That’s because there are so many payment methods, spending habits, and options that could come next.
• Building traditional machine learning models often requires a significant investment of manual labor and time. Additionally, because these models are normally trained using a single institution’s data, they fail to see emerging global threat patterns. 
• Feedzai’s RiskFM model, currently in its research stage, is uniquely designed for tabular transaction data. It has achieved parity with highly-tuned, bespoke models, removing the typical months of data cleaning and feature engineering work usually required to build a custom model.
• RiskFM drastically reduces the “time to value” by eliminating operational bottlenecks. This allows banks to more effectively stop fraud with less initial effort.

A Race Against Adaptive Financial Criminals

Language follows a predictable script. Financial crime, on the other hand, is a chaotic, constantly moving target. Criminals don’t follow the rules of the law and, when discovered, they invent new strategies.

This adversarial reality means that a customer’s “normal” behavior isn’t just hard to guess; it’s being constantly mimicked or manipulated by sophisticated fraudsters. Because transactional data is so much noisier than a sentence, current defenses are forced to catch up against threats that evolve faster than the software meant to stop them.

“RiskFM proves our multi-year investment in foundation models is paying off. We’re not just part of the conversation; we’re defining how it applies to the complexities of global financial crime prevention.” — Pedro Barata, Chief Product Officer at Feedzai.

Building a reliable machine learning model today often requires considerable labor: data scientists must manually build features to look for specific patterns like transaction velocity or rapid geographic leaps. It’s also highly time-consuming. This manual feature engineering makes it difficult to swiftly adapt to shifting fraud patterns and consumer spending habits.

Beyond requiring extensive data science effort, these models also face the challenge of a narrow “field of vision” because they are typically trained in isolation on a single institution’s data and use case. This limited view prevents banks from seeing broader, emerging global threat patterns observed elsewhere. By the time new patterns are identified and deployed into systems, criminals could have already shifted tactics.

RiskFM Offers a New Paradigm for Decisioning

Put another way, traditional machine learning requires data scientists to identify, define, and hardcode patterns before the model can learn from them. If a pattern is not captured, the model remains blind to it. RiskFM learns those patterns autonomously from the data itself, surfacing signals that no individual person anticipated and no ruleset could have captured.

“The ability to match bespoke supervised models out of the box, without manual feature engineering, has real implications for how institutions think about deployment speed, cost, and coverage across the full financial crime lifecycle, from card fraud to AML.” — Sam Abadir, research director, risk, financial crime, and compliance, IDC

Because RiskFM understands the foundational structure of risk, it creates immediate, out-of-the-box value that eliminates months of manual feature engineering. The model is built to make complex, autonomous decisions ensuring that while the intelligence behind the scenes is massive, the legitimate customer experiences no friction.

How RiskFM Changes the Game

RiskFM is reimagining how we defend the financial ecosystem. Here is how this research-first approach is rewriting the rules:

• From Manual to Autonomous. RiskFM uses self-supervised learning to autonomously identify complex financial behavior patterns, autonomously discovering the deep, mathematical signatures of risk.
• Immediate Value. Early research shows that RiskFM achieves parity with highly tuned, bespoke models on Day 1, reducing the time usually required for feature engineering and model iteration. 
• Cross-Institutional Intelligence. RiskFM is trained across vast, multi-institutional datasets,  allowing it to uncover a wide range of sophisticated threat patterns. Because it shares the knowledge of the threat without ever sharing the underlying personal data, every bank in the network benefits from the model’s collective intelligence.
• A Lower Barrier to Entry. By stripping away the operational bottlenecks, RiskFM lowers the time to value allowing banks to stop more fraud with less initial effort. 

Just as we’ve learned to anticipate that a customer at a coffee shop is likely asking for a latte and not a lawnmower, RiskFM promises to enable banks to understand the natural flow of financial behavior. We are moving away from the era of building every defense by hand and toward a world where your systems already speak the language of risk.

RiskFM, currently in its research and development stage, is a breakthrough in innovation that we are actively refining alongside the world’s leading institutions. We aren’t just looking for users; we’re looking for partners. If you’re ready to help shape the future of autonomous intelligence, we invite you to become a RiskFM early adopter. By leaning in now, you can help us validate these frameworks against real-world risks and ensure your institution is at the forefront of the next great shift in financial crime prevention.

Additional Resources

• Blog: AI for Fraud Detection: How it Works and Why It Matters
• Report: The Future of Fraud Prevention: Predictions for 2026
• Analyst Research: Merging Fraud Detection, AML, and Data Orchestration
• Solution: AI that Goes Beyond the Ordinary