For claims handlers and counter-fraud leads at UK insurers and MGAs. Your investigators can check only a share of claims. We help you refer the claims that matter: more fraud found per 100 referrals. The column work is keep or drop for each column.
Layout first, no claim data. You get the readiness review within 48 hours.
| Column | Check |
|---|---|
claim_closed_date | Known after the decision. Leak? |
policy_id | ID column. Drop |
claim_channel | Known at referral |
days_to_claim | Known at referral |
fraud_confirmed | Outcome. Check its rate |
Made-up column names, to show the layout. Your review uses your own columns.
Every wrong referral costs investigator time and slows an honest customer. The measure that matters is fraud found per 100 referrals. A claims fraud referral model that scores well in a notebook can still fail on this measure. Four causes are common:
A column that you know only after the claim closes. It helps the model in training. It is not there when you must decide.
A policy or claim ID, or a column with very many values. The model learns the rows, not the pattern.
If confirmed fraud is a very small share of rows, a model can look accurate and find almost nothing.
Too little history, or history from before your rules changed. The model learns a pattern that is gone.
Send the layout of one table, as 5 short answers (below). We return a review that lists:
If a test makes sense, we talk first. Then you send the data by a secure link. The data is pseudonymised and we delete it after 30 days.
Feature Refinery gives keep or drop for each column.
Five answers about one table, for example your closed-claims table. You do not need to send a single row.
Feature Refinery gives keep or drop for each column.
Trains with several algorithms on the same data and keeps the best three. You compare them by wrong decisions, F1, precision and recall, and you choose the balance between missed cases and wrong decisions.
Read more in the docs: Feature Refinery and Model Crucible.
No. Step 1 needs only the layout of one table. Data comes after a call, and only if a test makes sense.
You send it by a secure link. It is pseudonymised, and we delete it after 30 days.
Yes, at any size. With no data scientist, you get a clear first view of which columns to keep. With one, it saves their time on column work. With a full team, it speeds up the slow part. Each column is marked keep or drop.
No. A referral model is an input for your investigators. It helps them see which claims to check first. People make the decision.
Claims and counter-fraud analytics teams at UK insurers and MGAs that have a labelled history of referrals and outcomes, with no, one or many data scientists. Your team keeps its own models.
One table per team. You get the review within 48 hours. No claim data in this form, only the layout.
Prefer to talk first? Book a call.