Building a foundation risk model
Real-world learnings from training an AI model for issuing.
Fraud models read a flat list of features. A customer's history has to be compressed into thirty-day counts and averages before the model ever sees it, and every aggregation throws away the order the transactions happened in. That order is exactly what a foundation model is built to read.
This whitepaper documents what happened when we pretrained a transformer on issuing card transaction sequences, with no fraud labels anywhere in the process, and used it to generate features for the fraud model we already run. The scoring model didn't change. Its inputs did.
Inside, you'll find:
- The full five-stage pipeline: how card transactions were tokenized, pretrained on with no labels, and turned into two feature sets for the XGBoost model already in production
- The numbers: a 68% AUC-PR improvement, what that means in additional fraud caught at review depths from 0.05% to 2%, and why the "surprise score" barely mattered
- How the representation performed on a client held out of pretraining entirely, a 29% improvement, and exactly where that advantage ran out
- Why the embeddings alone underperform conventional features, and why combining the two beats either one on its own
- Where we expect this approach to fail next: merchant data, onboarding, and anywhere the data isn't shaped like an issuing portfolio
Fill out the form to download your copy now.
By submitting this form, you agree that Sardine may contact you about our products and services. For more information, review ourprivacy policy.