tailgust · in construction
An AI windstorm event set, scored where treaties are priced
A European windstorm catalogue generated by a fine-tuned atmospheric foundation model, held to the clustering statistics of the observed record, and terminated in the numbers a pricing actuary reads: an exceedance-probability curve, the average annual loss, the 1-in-200 occurrence loss, and the erosion of a layer under a 72-hour hours clause.
Nothing here is a result. The gates the event set has to pass are below; a number appears on this site only after they do.
the problem
Treaties Are Priced Where the Record Is Thinnest
Catastrophe models price European windstorm from event catalogues built by resampling about forty winters of reanalysis or by running expensive climate-model ensembles. Neither gives many independent samples of what sets the price: the clustered sequences in the far tail. Industry models put clustering at 5 to 10 per cent of annual aggregate European losses, rising towards 20 per cent at long return periods (Guy Carpenter 2025). Read as independent storms, Lothar and Martin in December 1999 come out near a 1-in-1,000-year loss; with clustering modelled, Moody’s RMS puts the pair at 1-in-200 to 1-in-300.
Extending the catalogue is what an event set is for, and the extension is only worth having if the synthetic seasons cluster the way real ones do. That is the quantity this project validates first, with gates written before generation.
the gates
What It Must Pass Before a Number Appears Here
- G1footprint severity against the CLIMK-WINDS conventionnot yet run
- G2duration-free serial clustering at 3 to 10 day lags, and the February clustering peaknot yet run
- G3Lothar/Martin-class sequencing ratesnot yet run
- G4tail calibration: generalised Pareto return levels against the recordnot yet run
- G5loss closure against published industry lossesnot yet run
If G2 fails, the shipped artefact becomes the exceedance-probability framework plus a modular evaluator that scores any windstorm emulator, with the negative result documented. The tail is graded by tailspec, an engine that was not built to flatter it; the problem statement in full is on tailspec’s event-set page.
get in touch