Revolut Trains AI Model on Its Own Data
The model outperformed Revolut’s existing baselines across multiple tasks.
Hey, it’s Matt. You’re reading AI Street, where I report on how Wall Street uses AI.
Kim Posnett, the co-head of investment banking at Goldman, argued in the FT last year that AI may turn overlooked corporate data into a newly valuable asset:
Imagine how a textbook company might use its archives of technical manuals and coursework to train an AI system to do complex scientific processes.
AI models are only as good as the data they’re trained on. Hard-to-replicate, legacy data is more valuable in the age of AI. And legacy companies are generally the ones with the legacy data. Many corporations are sitting on valuable intellectual property and, I suspect, don't even know it.
But you don't need decades of data. Scale works. Revolut, the UK-based neobank with 70 million customers across 40 countries, has been collecting billions of data points. Its users generate a continuous stream of timestamped card transactions, peer-to-peer transfers, in-app navigation events, and communications.
Revolut researchers and Nvidia say they have used that stream of banking activity to train PRAGMA, a foundation model for financial event data. The model is designed to analyze a user’s event history of transactions, app activity, communications and profile data, allowing one underlying model to be adapted for tasks such as credit scoring, fraud detection and product recommendations.
A NOTE FROM OUR SPONSOR
When your agent reads Bloomberg or Reuters, is it finding an edge?
On Tesla’s Q1 earnings, Goldman held at $375, TD Cowen reiterated Buy at $490, JPMorgan stayed at $145. You and your competitors are reading the same call.
The dispersion across sell-side targets. The reasoning behind each one. The hedge buried in the fifth paragraph of an operator quote.
That’s where the analytical signal is. It lives in the paragraphs your agent isn’t getting.
Typical retrieval looks fine. The agent doesn’t know what it’s missing, and neither do you.
Seltz returns full context in hundreds of milliseconds, every result traceable to source. Built for workflows where deep research matters more than the headline.
If you’re running agents on financial news, Seltz will run an eval on your setup.
Email CEO Antonio Mallia at antonio@seltz.ai or ask me for an introduction.
What Revolut Did
PRAGMA was trained on 26 million anonymized user records from 111 countries, covering 24 billion events, according to a research paper posted to arXiv. The model is not a chatbot. It is designed to make predictions from banking histories, not generate text.
“Most ‘foundation model for finance’ discussions still default to text. But bank data is not text,” said Revolut’s head of AI Pavel Nesterov on LinkedIn.
His point is that financial activity has its own structure. Customers generate long sequences of transactions, app actions, communications, trading activity and profile changes. Turning all of that into text for a generic language model, Nesterov wrote, means “you lose too much structure and waste too many tokens.”
PRAGMA keeps the records closer to their original form. Each event is represented by what happened, the value attached to it and when it occurred. A card payment, for example, can include the transaction type, amount, currency, merchant category and time. The model then looks for patterns across long sequences of customer activity.
The authors say PRAGMA beat Revolut’s internal task-specific baselines across credit scoring, external fraud detection, product recommendation, communication engagement, recurrent-transaction detection and lifetime-value prediction.
Results
The biggest reported gains came in credit scoring and customer communications. Compared with Revolut’s existing models, PRAGMA improved one credit-scoring measure by 130% and one customer-communications measure by 79%. It also improved fraud recall by 65% and product-recommendation performance by 41%.
For Revolut, the operational goal is to reduce the need for separate models for every use case. Nesterov said the company is trying to move away from “a separate model stack for every narrow use case” and toward one shared model that can be adapted for tasks such as credit scoring, fraud detection and product recommendations.
PRAGMA fell short on one task: anti-money laundering, where it significantly underperformed Revolut's existing system. The authors say that is because money-laundering detection often depends on relationships among accounts, counterparties and transaction networks. PRAGMA analyzes one customer history at a time.
Revolut joins Netflix and Stripe, which both trained models on their own internal data. Netflix built a foundation model on hundreds of billions of user interactions that now underlies its personalization across search and recommendations. Stripe’s payments foundation model, trained on tens of billions of transactions, increased its detection rate for card-testing attacks from 59% to 97%.
That’s the fun part of AI for me. It reveals patterns that were always there but we didn’t have the computing power to see.
Thanks for reading!
I’m always happy to receive comments, questions, and feedback.
Connect with me on LinkedIn, or
Send an email to matt [at] ai-street.co