Sardine says it trained a model on card transactions to predict the next transaction, then added its embeddings as features to its production issuing fraud model.
We trained a foundation model on spending and it caught 35% more fraud ↗
What happened
- Sardine reports the model caught 24% to 35% more fraud than before at the same share of transactions flagged for high risk.
- Sardine reports a 28% improvement at a common operating point flagging the riskiest 1% of transactions.
- Sardine reports a 68% improvement in AUC-PR.
- Sardine reports the embeddings improved a held-out client's issuing model by 29%.
- Sardine states the first version performed worse than its existing model and required substantial tuning.
- Sardine says the model is being productionized and is not yet deployed in a live regulated environment.
Why it matters to Signicat
- Sardine's claim targets issuing fraud, a use case adjacent to Signicat's risk and AML screening line, but the article does not address identity verification or onboarding.
- Sardine itself says onboarding has no sequence to read and the approach would likely underperform on e-commerce, limiting direct overlap with Signicat's onboarding-focused identity products.
Evidence from the source (8 quotes)
- “That model caught 24% to 35% more fraud than before, at the same share of transactions flagged for high risk.” source ↗
- “At a common operating point, where the riskiest 1% of transactions get flagged for action, the improvement was 28%.” source ↗
- “the improvement was 68%.” source ↗
- “The embeddings improved it by 29%.” source ↗
- “However, the first version we built actually performed worse than our existing model.” source ↗
- “we’re moving forward with productionalizing it in preparation for live deployment.” source ↗
- “Onboarding and guest checkout are the far end of that. One event, no prior history at all.” source ↗
- “we believe sequential-based foundation models would probably not perform as well on such use cases.” source ↗
Fraud deepfake injectionOrchestration platformThought leadershipResearch reportGlobal
7 of 7 facts verified against the source · confidence high · claude-haiku-5-5