Claim risk scoring
A model trained on confirmed investigation outcomes that ranks open claims by fraud risk for a fixed investigator queue.
Part of the recommended approachTechnique: Supervised Learning
Claim risk scoring is the workhorse once an insurer has enough confirmed fraud outcomes, with verdict and verdict date, to train on.
It replaces a flat referral list with a ranked one, sized to the investigator hours that actually exist, so the highest-value claims get looked at first instead of whichever claim happened to trip a rule or land at the top of the queue by report date.
Get the labels wrong and the whole approach is worse than doing nothing: training on adjuster referrals instead of confirmed outcomes bakes in whatever bias already shapes who gets referred, and letting investigation-only fields leak into the feature set produces a score that looks great in testing and fails the moment it meets a claim that has not yet been investigated.
This is the wrong starting point for an insurer without a meaningful history of confirmed outcomes.
In that situation, tune the intake rules and fix outcome recording first, since a model trained on too few or too biased labels will rank claims worse than the rules it is meant to improve on.
Strengths for this problem
- Combines weak signals across the claim, the policy, and the claimant's history into one ranked score, catching patterns no single intake rule covers.
- Sized correctly, it fills the investigator queue with the claims most likely to pay off rather than the claims that happened to trip a rule.
- Retrains on newly confirmed outcomes, so it keeps pace with fraud patterns that shift faster than a rule set gets updated.
Limitations for this problem
- Needs a real volume of confirmed investigation outcomes, with verdict and verdict date, not just adjuster referrals.
- If trained on adjuster referral decisions instead of confirmed outcomes, it reproduces whatever bias already exists in who gets referred.
- Score quality erodes as claim mix, repair costs, and fraud tactics shift, and the drift is often invisible until precision at the working referral volume drops.
Data needed
- Claim record fields as they existed at the scoring timestamp: amounts claimed, loss type, timeline of status changes.
- Policy and prior claim history joined to the claim as of scoring time, not as of settlement.
- Confirmed investigation outcomes with a verdict and the date the verdict was reached, used only as training labels.
How it is measured
- Precision at the referral volume investigators can actually work in a day, not precision on a theoretical top decile.
- Confirmed fraud dollar value prevented per investigator hour.
- Score drift and label delay, tracked continuously against the training population.
Pitfalls
- Feature leakage from fields that only populate once an investigation has opened, such as an investigator's own notes field or a special-handling flag, which produces excellent offline metrics and a useless production score.
- Training on adjuster referral outcomes rather than confirmed investigation verdicts, which teaches the model to imitate whichever claims adjusters already suspected rather than to detect fraud.
- Optimising for recall when investigator capacity is the binding constraint, which floods the queue with low-yield referrals and burns adjuster trust in the score.
Human oversight
An investigator reviews every referred claim and decides the outcome. The score orders the queue and surfaces the contributing factors; it never approves, delays, or declines a claim on its own.