The Weather Model That Runs in a Basement

Vardy’s answer, which she sent Fyfe as a map with a line drawn on it, was that the storms would form over the eastern ridge in the afternoon, as the national service said, but that the wind along the ridge would steer them north, and they would not reach the valley floor until well after dark. The festival field, she said, would stay dry until about nine. Fyfe moved the headline act from 9:30 to 7:45, told the bars to close at 8:30 and deleted the email. The first lightning reached the valley at 9:40 p.m., when the field was nearly empty. Twenty-two millimeters of rain fell on the ridge in 40 minutes; the collective’s nearest station to the festival, in a garden across the river, recorded five.

The Tamsell Valley Weather Collective is 23 volunteers, 41 homemade weather stations, 14 secondhand servers and a basement. For the valley it serves—18 kilometers long, three across at its widest, home to about 6,000 people and a great many sheep—it produces forecasts that are measurably more accurate than those of the National Forecast Service, which employs hundreds of meteorologists and runs one of the most powerful computers in the country. It would be easy to tell this as a story of amateurs beating the experts. Vardy, who is 63, would like it understood that it is nothing of the kind. She put it bluntly: “We are not smarter than the national service. We are smaller.”

Vardy spent 30 years as a hydrologist for the national water authority, and she started the collective after a summer evening in July 2014, when 66 millimeters of rain fell on the head of the valley in a little under two hours. The national forecast that morning had promised scattered showers. The river rose two meters, carried away a footbridge and flooded the campground by the old mill. Nobody died. Vardy, who had watched the storm sit on the ridge all afternoon, went home and started reading about weather models. “I was annoyed,” she said, “which is how most useful things start.”

A crease in the bedsheet

To see what she means by smaller, look at the valley the way a national forecast model does. The service divides the country into a grid of squares about nine kilometers across and calculates the weather in each one—temperature, wind, moisture, pressure—stepping forward a few minutes at a time. At that resolution the Tamsell Valley, a narrow trench with a river at the bottom and ridges 550 meters above it, barely registers. It straddles two squares, each of which averages the valley floor, the slopes and the ridgetops into a single gentle hump. “The national model sees our valley the way you’d see a crease in a bedsheet from across the room,” Vardy said. “It knows there’s something there. It doesn’t know it’s deep.”

Depth is where the valley’s weather lives. On clear, still nights, cold air slides off the slopes and pools on the valley floor like water in a bath, so the town can be five degrees colder than farms 200 meters up the hillside. On summer afternoons, the ridges shove moist air upward into thunderstorms that a coarse model smears across the whole region or misses altogether. A model that cannot see the trench cannot see either.

What the collective does about this is called downscaling, and it is less exotic than it sounds. Four times a day, the national service publishes its forecast free of charge. The collective’s servers take that forecast for the squares around the valley and use it as the edges of a much smaller, much finer model that covers only the Tamsell catchment, in squares 500 meters across. Inside that box, the real landscape is drawn in: every slope and side valley, the reservoir at the head of the river. The physics is the same as the national model’s. It simply has more detail to work with. The national service could do the same for every valley in the country, Vardy points out, if it had thousands of times its computing power and a reason to bother.

What the valley knows

The second ingredient is the volunteers. The collective’s 41 stations—most of them built from kits and mounted on garden posts, fence posts and, in one case, a retired telephone pole—report temperature, humidity, wind and rain every five minutes. Every forecast the model has made since 2016 has been checked against what those stations recorded, and the record of its mistakes is used to correct the next one. If the model has run a degree too warm on still nights at the bottom of Kiln Lane for nine years, it learns to knock off a degree. “It isn’t clever,” said Benedikt Saar, 37, a hospital systems administrator who keeps the servers alive. “It’s just paying attention for a very long time.”

The longest attention span belongs to Maud Ellison, 79, whose station stands in her garden on the north slope, 430 meters above sea level. She has read a rain gauge there at nine every morning since 1987,2 and in 2017 volunteers typed up her 30 handwritten notebooks to give the model a long memory of the upper valley. Ellison is not especially impressed by the servers. “They’re very good at telling me what I already know,” she said. “What I like is that now other people know it too.”

Local knowledge got a more skeptical hearing. In 2020, the collective gathered 31 pieces of weather lore from farmers, hikers and the valley’s older residents and tested each against 12 years of station data. Twenty-eight failed.3 Three held up well enough to be built into the model’s corrections. The most useful is a shepherd’s rule that when cloud sits on the southern ridge, known locally as the Cap, before 10 a.m., the valley floor will have rain by mid-afternoon. It turns out to be a fair description of moist air being forced up the ridge by a southwesterly wind, which the national model, unable to see the ridge, tends to miss.

Keeping score

Since 2022, a volunteer’s script has saved the national service’s published forecast for the town every evening at six, alongside the collective’s own, so that both can be scored against what actually happened.^* The clearest difference is in the overnight low, the number that matters to anyone with tender crops, young lambs or old pipes. In winter, when cold air pools on the valley floor most often, the national forecast of the next night’s minimum at the town station has been off by an average of 2.5 degrees. The collective’s has been off by 1.2.

The gap is narrower in summer, and on some measures there is barely a gap at all; the national model does well with daytime highs, which depend less on the shape of the ground. The collective has its failures, too. It forecast snow badly in the winters of 2022 and 2023, when corrections learned mostly from rain turned several centimeters of snow into a cold drizzle that never came. It has no radar, no satellite and nobody on the night shift. When the national service’s data feed went down for 11 hours in March 2024, the collective had nothing to downscale and published an apology instead of a forecast. “We’re a parasite,” Saar said cheerfully. “A useful one. But if the host stops feeding us, we stop.”

The National Forecast Service does not dispute the collective’s figures. A spokesperson said its models were built to forecast the country rather than each of its valleys, and that it welcomed “community efforts that build on our openly published data,” which is about as warm as a national institution gets.^† Two of its meteorologists told me, on the condition that I not name them, that they check the collective’s forecast whenever they have family in the valley.

Vardy now gets emails from other valleys asking how to start a collective of their own, and she answers each with the same three pieces of advice: use the free national data, find someone who can keep old servers alive, and find the people who have been writing the weather down for decades without being asked. The third is the hardest. Servers can be bought at auction. Maud Ellison’s 38 years of nine o’clock readings cannot.

Fyfe’s festival paid for six new stations last autumn, one of them on a fence post at the edge of the festival field, from where it sends its numbers every five minutes to a basement three kilometers away. Fyfe checks it more than she would like to admit. “I used to look at the sky,” she said. “Now I look at a fence post with a box on it, and then I look at the sky, and then I call Judith.” Vardy usually picks up on the second ring.