Innovation & Tech

Fingerprint's Fraud Score Catches Nearly 50% More Fraud in Customer Test

Fingerprint's new Fraud Score detected nearly 50% more fraud than one customer's in-house model, without raising blocks on legitimate users. Here's how the ML signal works.

By Nathan Brooks

4 min read

Fingerprint Launches Fraud Score, Reporting Nearly 50% More Fraud Detected in Customer Test
Fingerprint Launches Fraud Score, Reporting Nearly 50% More Fraud Detected in Customer TestPexels / pixabay

What's News

  • Fraud Score detected nearly 50% more fraud than a customer's in-house model at the same legitimate-user blocking level
  • Fraud Score supports seven fraud types including payment fraud, account takeover, and promotion abuse
  • Flink principal software engineer David Fennessey said blocking visitor IDs alone was insufficient against sophisticated abuse
  • Customers feed fraud outcomes back through Fingerprint's dashboard, API, or S3 for continuous retraining

Fingerprint says its new Fraud Score product detected nearly 50% more fraud than a customer's existing in-house model in a head-to-head test — while blocking the same number of legitimate users. That second condition is the part risk teams will actually care about. Any fraud system can catch more abuse by rejecting more traffic; the trick is doing it without turning away genuine customers.

Fingerprint launched Fraud Score as a predictive fraud signal that combines its device intelligence with machine learning models trained on each customer's own confirmed fraud outcomes. For practitioners, that answers a familiar pain point: collecting signals is easy, turning them into a reliable decision is hard. A device tied to several accounts might suggest promotion abuse, while an unfamiliar device entering an established account warrants a different investigation. Fraud Score is built to learn those distinctions from an organization's actual users, traffic, and outcomes.

Fingerprint notes the gain measures additional fraud detected, and its financial value depends on the transactions involved and how each business uses the predictions.

What does the specialization change for fraud teams?

Fraud Score's central design choice is separate models per fraud type. The company lists supported targets including:

  • Payment fraud
  • Account takeover
  • Multi-accounting
  • Fake account creation
  • Promotion abuse
  • Platform abuse
  • Content abuse

Those categories carry different business costs. A payment-fraud model helps a merchant assess an order; a promotion-abuse model hunts repeated offer claims through multiple accounts; an account-takeover workflow aims to catch unauthorized access without interrupting the legitimate account holder.

Customer outcomes supply the training feedback. Fingerprint handles model training and maintenance rather than asking every organization to build and operate its own prediction layer, combining its device signals with labels the business provides. New outcomes feed retraining as traffic and fraud patterns shift.

Label quality matters here, and Fingerprint is upfront about it. A model trained on confirmed outcomes gets better supervision than one trained on an undifferentiated list of suspicious events. If an organization's labels are incomplete or inconsistent, the predictions will reflect those limitations.

How device intelligence feeds the model

Fingerprint's device intelligence platform pairs persistent identification with contextual signals about web and mobile activity, helping businesses connect activity across sessions and spot patterns that account details alone miss. Its Smart Signals catalog flags automation, browser tampering, virtual machines, proxies, VPNs, and unusually high device activity.

The company cautions that these signals provide context, not verdicts — a privacy tool or unfamiliar environment doesn't by itself establish fraudulent intent. Fraud Score adds the prediction layer on top of that evidence, distinguishing between recognizing an environment and assessing how patterns in it relate to a specific business's fraud outcomes. Teams can use predictions directly or pipe them into an existing decision system.

The approach also goes beyond blocking identifiers already linked to abuse. A trained model can spot combinations of signals from previous cases, potentially flagging similar behavior before the next transaction completes.

The feedback loop: dashboard, API, and S3

According to Fingerprint's launch materials, customers supply fraud outcomes through the dashboard, API, or S3, manage labels, and analyze model performance. Predictions arrive in Fingerprint's API responses, so a decision engine can use them alongside other checks to approve an interaction, request extra verification, or route it for review.

Retraining is designed to keep the service responsive to evolving abuse, but the cycle still depends on what an organization learns after an event — including outcomes that only become clear later. Monitoring model performance over time remains part of operating the system.

Flink's experience

The announcement includes an account from Flink, the delivery platform. Principal software engineer David Fennessey said blocking visitor IDs alone was insufficient against more sophisticated abuse. He described working with Fingerprint on a model informed by Flink's internal fraud analysis to help stop future abuse before an order is placed.

The takeaway for fraud teams: Fraud Score shifts the work from maintaining individual signals and rules toward learning from a business's own outcomes, with Fingerprint handling training and retraining while teams keep control over how predictions inform decisions. The product is available now through Fingerprint's Fraud Score page.

Source: unite.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.

2,232 articles

Related articles

« Previous articleNext article »