Inside Revolut's PRAGMA: The Foundation Model Trained on 40 Billion Banking Events 🧠

Why It Matters

By turning massive behavioral data into universal embeddings, PRAGMA could redefine risk assessment, fraud prevention and customer engagement across fintech, giving Revolut a strategic edge while forcing rivals to match its data scale.

Key Takeaways

• PRAGMA trained on 40 billion banking events from 25 M users.
• Single backbone boosts credit scoring PR‑AUC by 130 %.
• Fraud recall improves 64.7 % versus Revolut’s production baseline.
• AML detection performance drops 47 % with the new model.
• Stripe, Mastercard, Visa also launch financial foundation models.

Pulse Analysis

Foundation models have moved beyond language processing into domain‑specific arenas, and the financial sector is now the latest frontier. Revolut’s PRAGMA, described in an arXiv pre‑print, is built on an encoder‑only transformer that ingests 40 billion event records—transactions, app navigation, trading actions and push‑notification interactions—from a global user base. By converting this longitudinal behavior into dense embeddings, the model sidesteps the traditional pipeline of hand‑crafted features for each use case. The approach mirrors similar initiatives from Stripe’s PFM, Mastercard’s LTM and Visa’s TransactionGPT, but PRAGMA is unique in its consumer‑neobank origin and its multi‑source fusion.

The reported metrics illustrate both the promise and the pitfalls of a single‑backbone strategy. Credit‑scoring PR‑AUC jumps 130 %, and fraud‑recall climbs nearly 65 %, suggesting that richer representations can dramatically sharpen risk models. However, the same architecture yields a 47 % degradation in anti‑money‑laundering detection, highlighting that one size may not fit all regulatory‑heavy tasks. These mixed results will draw close scrutiny from supervisors, especially in Europe and the United States, where explainability and bias mitigation are mandatory for any AI‑driven underwriting tool.

From a market perspective, PRAGMA raises the bar for data‑centric competition. Fintechs that can amass comparable event volumes will be able to launch thin‑file credit products, while incumbents may struggle to retrofit legacy data pipelines to the scale required. Start‑ups focused on embedding explainability, model monitoring or domain‑specific fine‑tuning stand to benefit from a burgeoning ecosystem of financial foundation models. As Revolut, Stripe, Mastercard and Visa race to monetize these embeddings, the industry is likely to see a shift toward partnership models, data‑sharing consortia, and new regulatory frameworks that govern AI‑driven financial services.