Risk scoring

Risk scoring assigns a numeric score to a customer, transaction, or entity that estimates its likelihood of a specific negative outcome, such as fraud, default, or money-laundering risk. It converts a mix of signals — history, behavior, demographics, transaction patterns — into a single number used to rank or triage cases.

Scores are built from rule-based point systems, logistic regression, or more complex machine-learning models trained on past labeled outcomes. A well-built score is calibrated: a score of, say, 80 should correspond consistently to roughly the same real-world probability across the population it's applied to, not just a relative ranking. Risk scoring is related to but distinct from credit scoring, which specifically estimates creditworthiness, and from propensity models, which estimate the likelihood of an action like churn rather than a risk event.

Risk scores drive automated decisions — which transactions to block, which applications to decline, which fraud detection alerts to prioritize — so their accuracy and fairness matter directly, and higher-stakes scores are typically subject to model risk management oversight. A common pitfall is that a score trained on historical decisions can inherit and perpetuate past biases in who was approved, declined, or flagged, which is why fairness is usually reviewed separately from raw predictive accuracy.