Contextual Fraud Scoring in Fintech

The key point is that fraud scores only matter inside the operating reality of the company using them. A brokerage app like Robinhood and a universal bank like Bank of America see different user behavior, different transaction types, and different loss tolerances, so the same numeric score can imply very different action. That is why Unit21 is built less like a generic fraud model and more like a decision layer that learns from each company’s own past approvals, denials, and investigations.

• Robinhood is mostly handling fast moving retail trading and cash movement from self directed users, while Bank of America handles a much wider mix of cards, deposits, wires, Zelle, and branch linked banking activity. Those flows create different normal patterns, so risk has to be calibrated against each institution’s own baseline, not a shared industry number.
  1 sacra 2 bankofamerica
• Unit21’s product logic is to ingest whatever signals a customer has, identity checks, device data, transaction history, account changes, and investigator outcomes, then let risk teams build explainable rules and scores on top. In practice, that makes Unit21 a workflow system for deciding what to review, block, or escalate, not just a model that spits out a score.
  1 sacra 3 unit21
• This also explains the market split with companies like Alloy and Sardine. Alloy is strongest at identity and onboarding orchestration, Sardine contributes device and fraud signals, and Unit21 sits downstream as the customizable place where those inputs get combined with a company’s own case history to make a final decision.
  1 sacra 4 sacra 5 sacra

Fraud tooling is moving toward company specific decisioning systems trained on local context, not one size fits all scores. The winners will be the platforms that can absorb more signals, show investigators why an alert fired, and improve approval rates without raising loss rates, especially as real time money movement makes bad decisions more expensive.