Revolut Trains AI Model on 40 Billion Banking Events

Revolut has published PRAGMA, a foundation model trained on 40 billion financial events from 25 million users. The model treats banking transactions like words in a sentence and uses masked modeling to learn patterns, enabling fraud detection, credit scoring, and churn prediction.

Details

Revolut has developed PRAGMA, a foundation model trained on 40 billion financial events from 25 million users. The core idea is to treat banking transactions like words in a sentence and use masked modeling (similar to BERT) to learn patterns from the sequential, contextual, and predictive nature of financial event sequences. PRAGMA tokenizes financial events and learns to predict future events based on past patterns. Revolut has created three models of different sizes (10M, 100M, and 1B parameters) that share the same pre-trained weights and can be used for various applications like real-time fraud detection, credit scoring, and precision analysis. PRAGMA is already running in production, powering Revolut's AI assistant and deployed on 200+ NVIDIA H100 GPUs. Unlike text-based financial language models like BloombergGPT, PRAGMA is an 'AI that has experienced finance' rather than just 'knows about finance'. The key advantage is Revolut's data moat - the 40 billion transaction dataset from 25 million users, which would be difficult for others to replicate.