Insurance claims fraud detection datasets: what's labeled, what's legal

Two cautions travel with that convenience. Documentation depth is set by whoever uploaded the file, and the top results were last modified between 2024 and 2026, so expect anything from a full methodology card to a bare table. Contributor geography is equally mixed: US-centric claim tables sit beside synthetic African-market insurance data, so a model tuned on one upload can silently fail on the next.

Datadory scores the source 6 out of 10 on documentation, access reliability and freshness - below the 9 carried by FEMA NFIP Redacted Claims v2 (OpenFEMA) in the same industry slice. Treat the hub as a feature store to sample from after inspection, not as a benchmark to quote.

The labeled fraud tables Datadory catalogs live in a different industry entirely. Kaggle - Credit Card Fraud Detection (ULB) ships 284,807 European card transactions containing 492 frauds, a 0.172% positive rate, and Kaggle - IEEE-CIS Fraud Detection (Vesta) provides a 590,540-row transaction table with 394 features joined by TransactionID to a 144,233-row identity table. Both are excellent for studying what a fraud label looks like in practice; neither describes an insurance claim, so methods transfer while semantics do not.

For scale context, FEMA's wider open-data estate - the OpenFEMA Data Sets Hub - serves 49 datasets covering declarations, individual assistance and hazard mitigation, none of them carrying adjudication outcomes either. That gap is why this query resolves the way it does: the free labeled data comes from community uploads, and the authoritative claims history arrives unlabeled.

The same 2.72-million-record file supports targets that need no invented label: claim severity conditioned on flood zone and occupancy, payment outcomes given damage amounts, and frequency spikes aligned to named catastrophes. Because claims reach back to 1970 while policy detail starts January 1, 2009, long-horizon severity work is possible, but exposure-normalized rate models are effectively a 2009-onward exercise.

Three free sources publish the aggregates anomaly detection needs as context. The FCA general insurance value measures, surfaced through Data.gov.uk Insurance Search, give annual claims frequency, acceptance rates and payouts at firm-by-product level for calendar years 2022 onward - the closest published analogue to a company-level suspicion rate anywhere in Datadory's catalog, sitting inside a UK directory of 169 results (205 via the CKAN API).

In the US, NAIC Research and Insurance Data covers homeowners market dynamics for all 50 states plus DC and territories from 2018 through 2024 and publishes market-conduct working papers, though redistribution requires its Request to Use Content form. Triple-I's III Homeowners and Renters Insurance Statistics tabulate cause-of-loss frequency and severity for 2018-2022 plus FAIR Plan exposure for FY2014-FY2023 across roughly 16 HTML tables under a noncommercial-use allowance. For pricing-cycle covariates, FRED Insurance-Related Series (St. Louis Fed) contributes 16,924 insurance-tagged time series - among them P&C premium PPI index PCU924126924126 beginning June 1998 and insurance-carrier employment series CES5552000001 beginning January 1990 - downloadable as CSV, JSON, XML or Excel with attribution required. None of these labels a claim; together they define normal, which is half the problem.