European windstorm losses inferred from a fine-tuned foundation model

the problem

Storm Clusters Priced as Single Events

the approach

A Fine-Tuned Aurora with a Diffusion Gust Head

Aurora is a weather foundation model with 1.26 billion parameters, trained to step the global atmosphere forward in time. It does not predict wind gusts, the quantity that drives windstorm damage. The approach adapts it in two stages and leaves its dynamics untouched, so the pretrained model’s representation of cyclones is read out as a gust.

• Fine-tuning with low-rank adapters. The 1.24-billion-parameter backbone is frozen. Rank-8 LoRA adapters and the Perceiver encoder and decoder are trained, 19.2 million parameters or 1.5 per cent of the model, together with a new 10 m gust output and its own patch embedding. A spectral loss on the gust keeps the small-scale variance that a pointwise loss averages away. Against ERA5 on 30 large historical storms, this cut the loss shortfall from a factor of 11.8 (the same fine-tune with a pointwise loss) to 3.2.
• Diffusion gust head. A deterministic output predicts the average gust of all storms that look alike, and so smooths away the peaks that set the loss. The head, a conditional diffusion model with 21 million parameters, instead samples the gust field from the fine-tuned model’s latent state, eight draws per state, so each storm carries a spread of losses. The deployed head, trained on 16 ERA5 winters, brings the simulated storm-week loss to 1.22 times ERA5’s on the same start days, where the first head, trained on 4, reached under half.
• The event set. 400 winter start days, drawn by importance sampling on each day’s ERA5 loss, are each run eight days forward, and every gust draw passes through one exposure and damage chain. The 400 runs took 57.6 A100 GPU-hours, excluding training and data preparation.

Neither stage retrains the base weights of Aurora’s 1.24-billion-parameter backbone. The gust is read from the latent state of a model that is 98.5 per cent pretrained, which is what makes this transfer learning.

result

Simulated Storm Losses Compared with the 1979–2022 Record

From 1 to 43 years the event set’s loss curve (corrected below 8 USD bn) falls inside ERA5’s band, and past the record it continues to a 1-in-200 event loss of 17.3 USD bn.

• The reinsurance layer. No correction acts above 8 USD bn, so this range tests the model itself, in sample. A 10 USD bn xs 8 USD bn layer attaches in 8.2 per cent of years against 7.6 for ERA5, and its expected annual loss is 0.25 USD bn against 0.23.
• Return levels. 3.8 USD bn at 5 years against ERA5’s 3.6, 7.0 at 10 years against 6.7, and 11.6 at 43 years against 11.8. The average annual loss is 6.4 USD bn against 6.3.
• Storm-week loss. Before any correction, the event set’s 72-hour losses on the same 393 start days come to 1.22 times ERA5’s.

Each point is a loss per 72-hour event and the return period at which it is exceeded. Below 5 USD bn an exceedance-rate correction sets the losses (see Methods). The start days come from winters used in fine-tuning, and the 1-in-200 level rests on four of them. With three ERA5 storm episodes above 8 USD bn, ERA5’s band there is too wide to separate the two curves.

methods

Model, Sampling and Loss Calculation

The pretrained Aurora model runs the atmosphere forward from ERA5 on each start day. The fine-tuned copy encodes every state, and the diffusion head, an EDM-preconditioned UNet trained for 40,000 steps, draws the 10 m gust from it eight times. Every storm is put through a LitPop exposure over 17 European countries and a damage curve whose loss scale is calibrated on seven historical storms.

The 400 start days are drawn by stratified importance sampling from five stormy winters, and each is reweighted by how often a window with its largest ERA5 storm occurs over all 43 winters, with weights computed from ERA5 alone.

The raw event set produces too many moderate storms (an average annual loss of 11.0 USD bn against ERA5’s 6.3), while its storms above 8 USD bn already fall inside ERA5’s band. Each event’s loss is therefore corrected by matching exceedance rates: a loss the event set reaches a given number of times a year is replaced by the loss ERA5 reaches that often. Every event is kept and only its loss changes. The correction applies in full below 5 USD bn and fades to none at 8 USD bn, so the largest storms are untouched.

An event set is useful for pricing only if it reproduces, against the observed record:

• footprint severity in the CLIMK-WINDS convention
• serial clustering at 3 to 10 day lags, and the February clustering peak
• the rate of Lothar & Martin-class storm sequences
• tail return levels, from a generalised Pareto fit, consistent with the record
• event losses that close against published industry losses for historical storms

The tail is scored by tailspec.

supporting checks

Checks on the Raw Event Set

• Storms per window. The raw event set has 2.17 events per eight-day window and ERA5 2.11 on the same start days, so the excess at the short end comes from how severe the storms are. Part of that match is inherited from runs that follow ERA5’s storms.
• Historical storms. With one fitted loss scale, the loss module reproduces the reported losses of seven historical storms (the XWS catalogue, Roberts et al. 2014) to within a factor of 2.4 at the worst storm, and ranks Martin below Daria, Lothar and Kyrill, as the record does.
• Gust head. In pre-registered tests at four training winters, the diffusion gust head raised the storm-gust level by 2.0 times the test’s minimum detectable gain on reanalysis-driven states, and by 1.8 times on free-running Aurora rollouts.
• Reweighting. Applied to ERA5 alone, the start-day reweighting recovers ERA5’s all-winter curve to between 0.81 and 1.23 of its level at every return period up to 34 years, and its average annual loss to 6.6 USD bn against 6.3.

beyond wind

Flood, Hail and Tropical Cyclone Event Sets

The method needs a foundation model whose latent state resolves the weather behind a peril, a hazard variable the model does not predict, and enough observations of that variable to train a generative head. Three perils that reinsurers price from short records fit it.

Flood is the closest. Its loss assessment needs tens of thousands of simulated years of rainfall (Fathom, EGU 2026) at kilometre scales a 0.25° model does not resolve, and diffusion emulators already draw km-scale rainfall from coarse inputs (Addison et al. 2024). Conditioned on a fine-tuned latent state, the same head would draw it along simulated storm tracks.

Hail is far below any global model’s grid, so hail models fall back on proxies such as instability and shear, environments the foundation model does resolve. For tropical cyclones, AI weather models can rapidly produce thousands of scenarios (AXA and NVIDIA, EGU 2025), each of which needs a wind and rainfall footprint before it has a loss.