Understanding Fraud Score

The Fraud Score section of the Dashboard is where you train and manage your Fraud Score models. Fraud Score trains machine learning models on events you label as fraud or non-fraud, and returns a prediction for each label on your future events. In the Dashboard, you can:

• Create labels for the fraud types you want to detect
• Label events by uploading a CSV file, in bulk from the events table, or one by one
• Check the health of your training data
• Choose how strict each model should be, then review and approve new model versions
• See Fraud Score predictions on your identification events

​

Availability

The Fraud Score section is available to workspaces with Fraud Score enabled. To enable Fraud Score for your workspace, contact support. What you can do depends on your role and permissions:

​

How Fraud Score works

1. You create a label for each fraud type you want to detect, such as payment fraud or fake accounts.
2. You label events as that label, or as Confirmed non-fraud.
3. Once a label has enough labeled events, Fingerprint trains a model for it, separately for each platform: Web, iOS, and Android.
4. Fingerprint reviews each new model version for quality, then you review and approve it.
5. The approved version becomes the active model. It scores your new events, and its predictions appear in the Dashboard and in the labels field of the Server API event response.

Each label needs at least 1,000 labeled events on a platform to train its first model for that platform. Label at least as many non-fraud events as fraud events. For the best results, aim for about 10 non-fraud events for every fraud event, from the same time period.

​

View Fraud Score predictions

Once a model is active, its predictions appear on your identification events:

• Event details: the Fraud Score card shows each label’s model version, prediction (Detected or Not detected), ML score from 0 to 1, and threshold. A label is detected when its ML score reaches the threshold.
• Events table: the Fraud Score column shows the labels detected for each event. The column is hidden by default. Turn it on in the column selector.

The same predictions are available in the labels field of the Server API event response.

​

Use predictions responsibly

A Fraud Score prediction is not a confirmed outcome, and models can flag legitimate users. Use predictions as one input to your fraud decisions. For decisions that significantly affect a user, such as closing an account, review the event before you act.