PRAGMA is a family of foundation models for multi-source banking event sequences. Encoder-only transformer pre-trained with masked modeling on heterogeneous banking events. Supports credit scoring, fraud detection, lifetime value, communication engagement, recurrent transaction detection, and product recommendation via linear probes or lightweight fine-tuning.
Architecture: profile-state encoder, per-event encoder, history encoder fusing static and sequential signals. Scales from 10M to 1B parameters.
Warning
Paper reports 47.1% F-0.5 drop on AML downstream task — per-user encoder limitation; authors recommend graph-aware complements for relational fraud.
Submitted April 9, 2026. Presented at GTC 2026.