Fraud score
A probability, computed by a model at notification or during handling, that a claim is fraudulent, which routes files to investigation without ever being proof, and whose false positives have a cost.
A fraud score is a rating assigned to a claim, or an application, estimating the probability that it is fraudulent. It combines expert rules, supervised learning models trained on past confirmed frauds, anomaly detection methods to spot unusual behavior, and sometimes network analysis, which links policyholders, witnesses, repairers and addresses to reveal organized rings. The score is not proof: it serves to route the most suspicious files to investigators, whose capacity is limited, and to let the others through quickly. Its effectiveness is measured by the share of actual frauds found among flagged files and by the amount saved, relative to investigation costs. Two difficulties are specific to the field. Known frauds are rare and represent only those that were detected, which biases learning. And false positives have a real cost: an honest policyholder whose settlement is delayed or disputed suffers harm, may go to the ombudsman and often leaves the insurer. Score governance must therefore build in fairness, to avoid it indirectly targeting certain populations.
Textbook case: an insurer automatically flags the top 3 percent of motor claims by its score. Network analysis reveals, among them, 40 files sharing the same garage, the same witnesses and accidents always at night: the investigation breaks up an organized ring and avoids 1.2 million euros of payments.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.