This page may contain stale information. Last updated: 2026-06-28
Overview
Transaction foundation models are encoder-style transformers pre-trained on raw financial event streams (transactions, app behavior, notifications) to replace siloed task-specific ML in banking and payments.
Timeline
- 2025: Nubank publishes nuFormer (encoder for transaction ledgers)
- 2026-04: revolut submits PRAGMA paper (arXiv:2604.08649)
- 2026 (GTC): revolut + nvidia present PRAGMA production deployment
- 2026: Stripe, Mastercard, Visa converge on transaction FM strategies (2026-06-25-revolut-pragma-nvidia-transaction-fm-blog)
Key Players
- revolut — pragma (broadest event fusion among neobanks)
- Stripe — fraud blocking ~$112B/year with foundation model approach (stripe)
- Mastercard, Visa — payments-network transaction models
- nvidia — NeMo AutoModel developer tooling for transaction FMs
Analysis
Competitive layer emerging: behavioral representation quality vs individual model sophistication. Unified backbones reduce feature pipeline duplication across fraud, credit, and growth teams. Regulatory constraints may limit deployment in highest-value markets.