Revolut PRAGMA: foundation model boosts credit scoring and fraud detection

Revolut has developed PRAGMA, an AI foundation model trained on proprietary banking data. The model uses an encoder-only Transformer architecture and is pre-trained on 24 billion banking events from about 26 million users across 111 countries. PRAGMA’s training ran from 2023 to 2025, using custom tokenization for financial records and a masked modeling objective to learn missing parts of banking event sequences. Revolut also reports strong early results: PRAGMA delivers a 130% improvement in PR-AUC for credit scoring versus its prior machine-learning benchmarks, and a 65% jump in fraud detection recall. The project was a collaboration between Revolut Research and NVIDIA, which supplied the accelerated computing needed for training at this scale. The foundational research was first published as an arXiv preprint on April 9, 2026, with additional analysis released on May 3, 2026. A key technical feature is that PRAGMA can generate reusable user and behavior embeddings. Instead of training separate models for each banking task, the same learned representations can be fine-tuned for credit scoring, fraud detection, and potentially broader risk management use cases. For traders, the main takeaway is that PRAGMA signals continued acceleration of “AI on financial data” inside the fintech sector. While it is not directly tied to crypto markets, it can influence sentiment around digital finance and institutional AI capabilities over time.