Revolut Deploys PRAGMA, a 40-Billion-Event Foundation Model for Finance
Revolut has published PRAGMA, a family of Transformer-based foundation models built specifically for banking event sequences, marking one of the more ambitious in-house AI efforts from a European fintech to date.
The model is described in a research paper posted to arXiv by a team of Revolut researchers led by Maxim Ostroukhov, Ruslan Mikhailov and Vladimir Iashin, alongside Pavel Nesterov and colleagues. Rather than adapting an off-the-shelf large language model, Revolut's team designed PRAGMA from the ground up for the messy, discrete, variable-length nature of transactional and event-level banking data.
A foundation model for transactions, not text
PRAGMA is pre-trained using a masked modelling, self-supervised objective, the same broad recipe that powers modern language models, but adapted to financial event logs instead of sentences. Each customer's history of transactions, logins, card actions and other events is treated as a sequence that the model learns to reconstruct when parts of it are hidden.
According to the paper, PRAGMA was trained on a large-scale corpus of roughly 40 billion behavioural events drawn from Revolut's customer base. The architecture uses a key-value-time tokenisation scheme and a two-branch design, with profile-state and event encoders feeding into a history encoder that handles heterogeneous financial records.
Downstream tasks
The point of a foundation model is that one pre-trained backbone can be reused for many jobs. Revolut's paper highlights three in particular:
- Credit scoring
- Fraud detection
- Customer lifetime value prediction
The team reports that strong results can be obtained by training a simple linear model on top of PRAGMA's embeddings, and that performance improves further with lightweight fine-tuning. In other words, once the foundation model exists, adding a new risk or growth use case becomes a much smaller piece of work than training a bespoke model from scratch.
Why it matters
Most retail banks still run a zoo of separate machine learning models, each trained on a narrow slice of customer data for a single task. A single foundation model that ingests raw event sequences and produces general-purpose embeddings is a very different operating model, closer to how big tech companies run recommendation and ranking systems.
For Revolut, the strategic logic is clear: with 50 million plus customers and a very high volume of transactional and in-app activity, the company has exactly the kind of data that benefits from self-supervised pre-training at scale. PRAGMA is also a recruiting signal to ML researchers that Revolut is prepared to do original work rather than just wire up third-party APIs.
The paper is authored by Maxim Ostroukhov, Ruslan Mikhailov, Vladimir Iashin, Artem Sokolov, Andrei Akshonov, Vitaly Protasov, Dmitrii Beloborodov, Vince Mullin, Roman Yokunda Enzmann, Georgios Kolovos, Jason Renders, Pavel Nesterov and Anton Repushko, and is available on arXiv as paper 2604.08649.
About Revolut
Revolut is a global financial technology company that offers banking, payments, investing, crypto and business services across more than 35 countries. Founded in 2015 and headquartered in London, the company now serves more than 50 million customers worldwide.
