Glossary
Fraud detection
Identifying transactions, claims, or behavior that are likely fraudulent, using rules, statistics, or machine learning.
Fraud detection identifies transactions, claims, or behavior that are likely fraudulent, using techniques ranging from fixed rules — flag any transaction over a set amount from a new device — to statistical anomaly detection and supervised machine learning models trained on confirmed past fraud cases.
Systems typically produce a risk score or flag per event, reviewed by an analyst or triggering an automated action such as holding a transaction. Performance is judged on the trade-off between catching real fraud (recall) and not blocking legitimate customers (false positive rate), since tightening a threshold to catch more fraud almost always increases false declines too.
Fraud detection differs from transaction monitoring and anti-money-laundering work, which focus specifically on illicit-finance patterns under regulatory obligations, though the areas share techniques and often the same underlying data. Fraud patterns shift as fraudsters adapt to defenses, so models need regular retraining; a common pitfall is a model that performs well in testing but degrades quickly in production because tactics have moved on, or a system tuned so aggressively that ordinary customers are routinely declined.
Last reviewed September 22, 2026