What is fraud detection?
Fraud detection is the practice of identifying deceptive activity, such as stolen card payments, staged insurance claims, account takeovers, or invoices from impostor suppliers, in time to stop it or recover the money. Most programs combine three layers. Rules catch known patterns, such as one card used in two countries within an hour. Supervised models learn from labeled past cases and score each new event. Anomaly detection flags activity unlike the customer’s own history or unlike similar customers, which helps with schemes nobody has labeled yet.
How fraud detection works
Each event is scored on features that describe its context. For a card payment, these include the amount compared with the cardholder’s usual spending, the number of attempts in the past hour, the merchant type, the device, and the distance from recent purchases. A payment decision is needed while the customer waits at checkout, so features are precomputed and the model runs as real-time inference within milliseconds. A threshold turns the score into an action: approve, decline, or hold for review. Graph analysis links accounts that share devices, addresses, or phone numbers, which exposes organized rings.
Fraud is rare, so training data suffers from severe class imbalance, and a model that never flags anything can still report high accuracy. Teams measure precision and recall instead and put a price on both kinds of error. A missed fraud costs the loss, and a false positive costs a declined customer who may shop elsewhere.
Fraud detection in practice
Impostors posing as a medical supply vendor email a regional hospital system’s accounts payable team and ask to change the bank account on file. The hospital scores every change to vendor master data and every payment run. A new bank account followed within days by a large payment, an invoice far outside the vendor’s usual amounts, or a near-duplicate invoice number raises the score. Flagged payments wait until a clerk confirms the change by calling the vendor at a number already on file. Confirmed fraud cases are too rare in accounts payable to train a supervised model, so the program relies on rules and anomaly detection.
What to watch for
- Labels arrive late and incomplete, because chargebacks can take weeks and undetected fraud is recorded as legitimate.
- Fraudsters adapt to whatever gets caught, so monitor for concept drift and keep a fast way to add rules between retraining cycles.
- Set thresholds together with the business owners who absorb the losses and the customer complaints.
- The EU AI Act excludes fraud detection from its high-risk credit scoring category, but an automated block on a customer can still fall under data protection rules on automated decisions.