Fingerprint Launches Fraud Score, Reports Nearly 50% More Fraud Detected

Fingerprint has launched Fraud Score, a predictive fraud signal that the company says detected nearly 50% more fraud than one customer's existing in-house model in a head-to-head evaluation — without increasing the number of legitimate users blocked. The figure anchors the launch of a product that combines Fingerprint's device intelligence with machine learning models trained on each customer's own fraud outcomes.

The launch targets a problem every risk team recognizes: collecting signals is easier than turning them into a reliable decision. A device associated with several accounts might indicate promotion abuse. An unfamiliar device attempting to enter an established account could warrant a different investigation. Fraud Score is built to learn those distinctions from an organization's actual users, traffic, and confirmed outcomes.

What does the reported detection gain actually show?

The headline result comes from a comparison with one customer's internal model. Fingerprint says Fraud Score detected nearly 50% more fraud at the same level of legitimate-user blocking.

That second condition matters. A fraud system can catch more suspicious activity simply by rejecting more traffic, but doing so also turns away genuine customers. Holding the false-positive rate steady while increasing detection is the harder engineering problem, and it is the condition Fingerprint reports meeting.

Two caveats frame the number. The gain measures additional fraud detected, not dollars saved. Its financial value depends on the transactions involved and on how each business uses the prediction. And the result describes one customer's in-house model, not a public benchmark across vendors.

Why one score does not fit every fraud type

Fraud Score's central design choice is specialization by fraud type. Fingerprint lists the following supported targets:

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

These categories describe different behaviors and carry different business costs. A payment-fraud model might help a merchant assess an order. A promotion-abuse model could look for repeated attempts to claim an offer through multiple accounts. An account-takeover workflow has a different objective again: detecting unauthorized access without unnecessarily interrupting the legitimate account holder.

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

That structure makes label quality decisive. A model trained on confirmed outcomes has more useful supervision than one trained on an undifferentiated list of suspicious events. If an organization's labels are incomplete or inconsistent, the resulting predictions can reflect those limitations.

How device intelligence supplies context beyond the account

Fingerprint's device intelligence platform combines persistent identification with contextual signals about web and mobile activity. Its role is to help businesses connect activity across sessions and inspect patterns that account details alone may not not reveal.

The company's Smart Signals catalog includes indicators for:

• Automation
• Browser tampering
• Virtual machines
• Proxies
• VPNs
• Unusually high device activity

These signals offer context for analysis, not verdicts. A privacy tool or an unfamiliar environment does not by itself establish fraudulent intent. Fraud Score adds a customer-trained prediction layer on top of that evidence. The distinction is between recognizing an environment and assessing how patterns in that environment relate to a particular business's fraud outcomes.

Teams can use the predictions directly or feed them into an existing decision system alongside their other information. The approach also goes beyond blocking an identifier already associated with abuse. A trained model can look for combinations of signals associated with previous cases, potentially helping a business recognize similar behavior before the next transaction is completed.

How the feedback loop works in practice

According to Fingerprint's launch materials, customers can supply fraud outcomes through three channels:

1. The Fingerprint dashboard
2. The Fingerprint API
3. S3

Customers manage labels and analyze model performance through the dashboard, while predictions arrive in Fingerprint's API responses. That architecture fits existing risk workflows: a decision engine can use the prediction alongside other checks to approve an interaction, request additional verification, or route it for review.

Continued retraining is intended to keep the service responsive to evolving abuse. The feedback cycle still depends on what the organization learns after an event, including outcomes that may only become clear later. Fingerprint's materials are explicit that monitoring model performance over time remains part of operating the system.

Why Flink wanted more than visitor-ID blocking

The announcement includes a customer account from Flink, the grocery delivery company. David Fennessey, Flink's principal software engineer, 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.

Flink's experience illustrates the collaboration behind the service. The business contributes its internal fraud analysis, and Fingerprint uses that feedback to build and maintain a model tailored to its abuse patterns. Fingerprint handles training and retraining; the fraud team retains control over how predictions inform its decisions.

Why this launch matters for fraud teams

The launch lands at a moment when device fingerprinting vendors face growing pressure to prove their signals translate into decisions, not just data. Rules built on individual signals — a VPN here, a clumped device there — flag environments rather than behaviors, and sophisticated abusers rotate environments faster than rules can be written. Fingerprint's bet is that models trained on a specific business's confirmed outcomes learn the combinations that matter for that business, and that vendors can maintain those models more cheaply than each customer can build its own.

The trade is also clear. Businesses that adopt Fraud Score hand Fingerprint their fraud labels, the most sensitive dataset in a risk operation, and accept that prediction quality will mirror label quality. For teams that already run device intelligence but struggle to operationalize it, the nearly 50% detection gain reported against one customer's internal model offers a concrete benchmark to test against their own stack. The product is available now through Fingerprint's Fraud Score page.