Read the film
- A windstorm tore across Nebraska before dawn. 9 August 2025: about 70,000 people lost power.
- Stations and spotters recorded the wind. Each dot is a high wind observed that night.
- The global weather record sees it in 31 km squares: one number for each square, each hour. A thunderstorm is only a few km across.
- 0 of 121 extreme gusts caught by the global weather record. The storm was averaged away.
- Most of the world has far fewer stations than Nebraska. There, the past is even harder to know.
- Now AI can rebuild the past at 3 km, anchored to the readings that do exist.
- Introducing RE-FOCUS: the model for what already happened. A new AI weather model from Tomorrow.io.
- RE-FOCUS rebuilt this storm from those winds: redrawn on a 3 km grid, pulled into place by the recorded winds, hour by hour, through the night.
- Of the 121 extreme gusts: 0% caught by the global record, 11% by NOAA’s HRRR, 25% by RE-FOCUS alone, 67% by RE-FOCUS with the recorded winds (scored against the winds it took in).
- Know what the wind did, where and when: settle claims on what happened, count storms the record missed, verify triggers at the site. Bring one storm. See it rebuilt.
How many hours, over how much ground, crossed the line the contract was written against?
On 9 August 2025 a storm crossed south-east Nebraska with gusts over 90 mph, hail and flooding. About 70,000 people lost power. Storm spotters and weather stations logged 121 reports of extreme gusts. The global weather record caught none of them.
What RE-FOCUS is
RE-FOCUS is an AI model that rebuilds past weather at storm scale, over periods that are already over. For US storms, it rebuilds a single event on a three-kilometre grid, hour by hour, anchored to the station readings and storm reports the event left behind. It doesn’t replace anyone’s risk model; it gives that model a sharper history of what happened. It is not a forecast, not a catastrophe model and not a climatology.
How it works
Read the steps
- Rebuild at storm scale. An AI weather model redraws the event on a 3 km grid, hour by hour, fine enough to resolve a thunderstorm. A 31 km global record can only approximate one.
- Let the evidence in. Station gust readings and storm reports are fed in and cycled through the event, pulling the storm to where the evidence says it was.
- Keep the range. It runs as 20 versions of the storm, so you see the most likely gust and the spread around it.
- Hand over the footprint. Hourly gusts across the event, cell by cell, in a standard gridded file. It enters your catastrophe model at the hazard step. One event takes minutes of compute.
Where it enters your model: at the hazard module. Vulnerability, financial terms and the model itself stay yours.
Why it matters
Severe convective storms, the hail, straight-line winds and tornadoes of a US summer, have become one of the largest sources of insured loss in the US. They are also the hardest weather to see in the historical record. A storm cell is a few kilometres across and lasts an hour or two; the global weather record is drawn in squares about 31 km wide. At that scale a storm can vanish altogether, and a storm the record can’t see is one it can’t count, price or settle on.
Claims & event response
After a storm
What did the wind do in that 3 km cell, and when? Reserve and settle on the rebuilt storm, not a regional average.
Reinsurance & cat
Checking your view
A storm the record missed is a storm your event history never counted. Check it against what actually happened.
Parametric
Verifying a trigger
A wind trigger is only as good as the record it’s settled on. Verify it on what the storm did at the site.
How we know it works
Storm study: the storm the record missed
Stations and storm spotters recorded 121 extreme gusts that night, and the global weather record caught none of them. RE-FOCUS caught a quarter on its own and two-thirds once it took in the storm reports. Drag across the map to compare.
The global record · 31 kmRE-FOCUS · 3 kmDrag to compare
Storm study
The storm the record missed
The night, report by report, and the storm rebuilt: 0 of 121 caught by the global record, 67% by RE-FOCUS.
Hurricane Milton: the footprint, hour by hour
At Milton’s landfall, the rebuilt storm reached severe strength, 58 mph or more, at 33 of the 47 places where storm reports logged severe gusts, scored against the reports the run took in.
What an insurer receives
We have delivered a rebuilt event to an insurer as a working dataset: hourly gusts on the 3 km grid for every ensemble member, with the mean, spread and extremes, in a standard NetCDF file.
Beyond single storms: ten years, and stations it never saw
Where station networks are dense, the same approach rebuilds years, not just storms. Across a decade in East Africa, scored against more than ten million station readings, RE-FOCUS came in below both records it was compared with. Lower is better.
| Ten years · average error | RE-FOCUS | Previous regional model | Global record |
|---|---|---|---|
| Temperature | 1.16 °C | 1.34 °C | 1.51 °C |
| Relative humidity | 6.03% | 9.21% | 8.85% |
Then the test a pricing actuary would design: hold stations back from the model and score it against them. Over a southern African domain, with more than a hundred stations held back for about a month, it came out clearly ahead of both: temperature error of 2.56 °C against 3.36 and 3.57, and humidity of 6.2% against 8.8% and 10.4%. In East Africa, with every station removed, it came out level on temperature. The observations are what carry it, which is why it’s strongest where they are dense.
What the tests coverThe US results are single storms, rebuilt one at a time at 3 km and scored on wind gusts against storm reports and stations. The two-thirds and Milton figures are scored against reports the run took in. They verify and settle single events; they are not a US wind climatology, so they don’t yet price a wind trigger on decades of history. RE-FOCUS works where stations and reports exist to anchor it.
SourcesUS storm reports from NOAA’s Storm Prediction Center; stations from ASOS, AWOS and METAR; the US high-resolution model is NOAA’s HRRR; the global record is ECMWF ERA5. East Africa: stations from TAHMO and METAR; the previous regional model is CBAM.
Talk to us
Bring one storm. See it rebuilt.
- You bring a storm from your loss file, the date and the area, and the record you used for it.
- We rebuild it at 3 km and compare both against the storm reports, the stations and your claims, on criteria we agree up front.
- You decide what’s next.
