Fraud Score

Overview

A fraud score combines selected signals into a value used to prioritize review, require additional verification, limit activity, or decline a transaction. Inputs may include identity, device, behavior, velocity, network, payment, merchant, and relationship features.

Scores are model-specific and do not have universal meaning. Calibration, thresholds, missing data, population changes, and feedback quality affect performance, while attackers may deliberately manipulate observable features or shift behavior after learning decisions.

Teams should document intended use, validate discrimination and calibration, monitor drift and bias, control overrides, and measure business outcomes. High-impact actions should consider contextual evidence and provide review or recovery paths instead of treating a score as proof.

Fraud Score estimates the likelihood or severity of a specified fraud outcome, while a risk score may address broader credit, compliance, security, or operational exposure.

A fraud score is a numerical or categorical estimate of how strongly available evidence indicates that an activity may be fraudulent. A fraud score ranks uncertainty within a defined model and population; it is not proof of wrongdoing or a portable risk measure.

Operational review of Fraud Score should reconstruct a numerical or categorical estimate of how strongly available evidence indicates that an activity may be fraudulent using the identities, communications, devices, and transaction records available for the affected case. Investigators should separate confirmed facts from hypotheses about numerical, preserve the original evidence, and document why the event was cleared, escalated, or treated as a loss. Containment, recovery, and customer communication for the Fraud Score context should match the harm indicated by numerical.

Quality review for Fraud Score should compare expected and actual outcomes involving numerical, then track false positives, repeat attempts, linked losses, and unresolved remediation.