What if you treated every bank transaction like a word in a sentence?
That's exactly what Revolut did. They just published PRAGMA on arXiv — a foundation model trained on 40 billion financial events from 25 million users. Instead of feeding text into a Transformer, they fed banking events: transfers, payments, currency exchanges, subscriptions.
The result: 20% better fraud detection, plus credit scoring and churn prediction — all from a single pre-trained base.
How it works (surprisingly simple idea)
The core insight: financial event sequences have the same structure as language. They're sequential, contextual, and predictive.
PRAGMA tokenizes financial events the way GPT tokenizes words. Then it uses masked modelling (like BERT, not autoregressive) to learn patterns from billions of real transactions.
"If a user paid at a coffee shop Monday, received salary Tuesday, and sent an international transfer Wednesday — what happens Thursday?"
PRAGMA learned these patterns from 40 billion events.
Three models, one pre-trained base
| Model | Parameters | Use Case |
|---|---|---|
| PRAGMA-10M | 10M | Real-time fraud detection (ultra-low latency) |
| PRAGMA-100M | 100M | Credit scoring, cross-sell prediction |
| PRAGMA-1B | 1B | Precision analysis (latency-tolerant) |
All three share the same pre-trained weights, fine-tuned per task. Stack a simple linear model on top of PRAGMA's embeddings and you get strong performance. No task-specific architecture needed.
Already running in production
This isn't a paper exercise. PRAGMA runs on 200+ NVIDIA H100 GPUs and powers AIR (Revolut's AI assistant), currently rolling out to 13 million UK customers.
The inference stack runs on Nebius (formerly Yandex Cloud) — a European fintech using European AI infrastructure. GDPR compliance matters.
Why this is different from BloombergGPT
| Model | Approach | Training Data |
|---|---|---|
| BloombergGPT | Text LLM + financial docs | News, reports |
| IndexGPT | Text LLM + financial QA | Advisory text |
| PRAGMA | Event sequence model | 40B real transactions |
BloombergGPT is "AI that knows about finance." PRAGMA is closer to "AI that has experienced finance."
The real moat
The architecture is published on arXiv. Anyone can read it. But nobody can replicate the dataset — years of financial behavior from 25 million users. That's the moat.
Any large neobank sitting on similar transaction volumes could build their own version. The technique isn't secret. The data barrier is.
Top comments (0)