The source describes a model that evaluates a transaction in the context of an individual history rather than treating each authorization as an isolated event. Ramp reports that the model ranks fraud 14% better than its predecessor and caught 32% more fraud dollars while flagging the same number of transactions.
Thesis
Transaction history can make card-authorization fraud scoring more selective without increasing the number of flagged transactions.
Observed facts
Mechanism
A transformer processes each cardholder’s prior transactions at authorization time, allowing the model to score a new purchase in sequence context.
Why now
Ramp has published comparative outcome figures for this real-time approach, making the claimed trade-off between detection and flag volume concrete.
Implications
The relevant operational measure is not only how many transactions a model flags, but how much fraud value those flags capture.
Evidence boundary
These performance figures are reported by Ramp; the source does not provide independent validation or establish that the results generalize to other issuers or systems.
Watch conditions
Look for independently reported results, or deployment details showing how the model performs across different portfolios and over time.