Why Financial Institutions Are Turning to 'Transaction Foundation Models' to Build Their Own Intelligence

Financial institutions have long built separate models specialized for specific tasks such as fraud detection, credit models, and recommendation engines. However, these siloed systems hinder a unified understanding of consumer behavior, and as data scale grows, they only add to the complexity between models.

As a solution to this, Transaction Foundation Models are emerging. These models are trained on billions of financial events—payments, transfers, and behavioral signals—to transform raw data into intelligence. In particular, by applying the Transformer architecture to tabular data, they capture context such as time, device, and location, catching signals that existing algorithms have missed.

Revolut partnered with NVIDIA to build a foundation model suite called PRAGMA. Leveraging NVIDIA Hopper GPU, cuDF, and Nemotron open models, this system was trained on data from 26 million users and outperformed existing dedicated models across various domains, including credit scoring and fraud detection. It also cut feature engineering time that used to take weeks down to nearly zero.

Mastercard is also developing a large-scale tabular foundation model for payments, utilizing technology from NVIDIA, AWS, and Databricks. This model is designed to process hundreds of billions of data points, and it is demonstrating performance that surpasses existing machine learning techniques across various areas such as cybersecurity, personalization, and portfolio optimization.