Outlier claim discovery
Unsupervised scoring that surfaces claims deviating from the norm for their segment, used to find new fraud patterns worth labelling.
Discovery, not a referral queueTechnique: Anomaly Detection
Outlier claim discovery earns its place when the concern is a fraud pattern nobody has described yet, which is exactly the gap intake rules and a risk-scoring model both have.
It compares each claim against its segment, so a claim with an unusually long treatment duration, an implausible repair cost for the stated damage, or a provider billing pattern that diverges from peers gets surfaced even though no rule was written for it and no past label taught a model to expect it.
Route its output to a research queue, not to the referral queue investigators work from, since low precision here is expected and normal, not a sign the method is broken.
Its real value shows up downstream, when confirmed patterns from this queue become new intake rules or new labelled examples for the risk-scoring model.
It is the wrong tool when the goal is a defensible referral decision on a specific claim, because an anomaly score alone gives an investigator nothing to point to.
Strengths for this problem
- Finds emerging patterns, such as a new staged-accident script or an unusual cluster of injury claims from one clinic, before any rule or model has seen them.
- Needs no confirmed fraud labels, so it works even where outcome recording is immature.
- Generates candidate claims worth sending to investigators for the labelling that later trains the risk-scoring model.
Limitations for this problem
- An unusual claim is not the same as a fraudulent one, so precision in the flagged tail is typically low.
- Explaining why a claim scored as an outlier takes real analyst work, since the score itself does not name a reason the way a fired rule does.
- Sensitive to how claims are segmented; comparing a total-loss auto claim against a fender-bender population manufactures false outliers.
Data needed
- A stable, correctly segmented population of settled and open claims to define what normal looks like, by line of business and loss type.
- Claim severity, repair cost, treatment duration, and provider identifiers, consistently captured across the comparison population.
How it is measured
- Precision of the reviewed tail once investigators close out the referred cases.
- Rate of genuinely novel patterns discovered that neither existing rules nor the risk model had captured.
- Overlap between anomaly flags and claims already caught by intake rules or the risk score, to check it is adding something new.
Pitfalls
- Routing anomaly flags straight to adjusters as if they were confirmed referrals, which floods the queue with low-yield reviews and erodes trust in every other flag.
- Segmenting the comparison population too broadly, for example scoring commercial fleet and personal auto claims together, which manufactures outliers out of ordinary differences.
- Treating a persistently high anomaly rate in one segment as solved by tuning the model, when it usually means a data quality problem in that segment's intake.
Human oversight
A fraud analyst reviews outlier claims as a discovery exercise, decides which represent a real new pattern, and hands confirmed cases to investigators and to the team maintaining the risk-scoring labels. Nothing is referred to an investigator automatically from this queue.