52
simulated books
1,174,275
simulated rows
5
public benchmarks
11
use cases
Frequency
Poisson claim counts with an exposure offset, from motor to pets.
Real data, public678,013 rows
French motor TPL (freMTPL2)
What makes it hard. Real policies: 95% have no claim and exposure runs from days to two years, so the exposure offset carries the model.
Kaggle, public595,212 rows
Porto Seguro safe driver prediction
What makes it hard. 57 anonymized features, a claim rate near 3.6%, and missing values coded as −1 instead of left blank.
Get it on Kaggle
then upload it to Pia
Simulated24,000 rows
UK private motor — claim frequency
Known trap. Exposure is part-year for most rows — fitting counts without log(exposure) as an offset silently rebases every relativity.
Simulated19,000 rows
Ontario auto — frequency with mid-term cancellations
Known trap. exposure_days, not years. Divide by 365.25 first, or your frequency lands ~365x too low.
Simulated22,000 rows
US homeowners — per-peril claim counts
Known trap. Heavily zero-inflated per peril — a single all-peril Poisson hides that water and weather behave nothing alike.
Simulated8,000 rows
Commercial fleet — vehicle-year frequency
Known trap. Variance far exceeds the mean — Poisson standard errors are optimistic, negative binomial or quasi-Poisson is the honest fit.
Simulated26,000 rows
Telematics motor — exposure in distance
Known trap. Distance exposure and declared annual mileage disagree by design — one of them is a rating factor, the other is the offset.
Simulated15,000 rows
Pet insurance — monthly frequency with seasonality
Known trap. A strong annual cycle sits in the residuals unless month is modelled — easy to mistake for a trend.