Analysis | We mapped weather forecast accuracy across the U.S. Look up your city. (2024)

Everyone complains about the weather forecast, but some people have more cause to grumble than others. In Miami, the temperature forecast is usually accurate a week into the future. But in Paonia, Colo., even the one-day temperature forecast is wrong by almost 6 degrees Fahrenheit on average.

That’s according to our analysis of government data shared publicly for the first time with The Washington Post. The data reveals the broad patterns and the fine contours of the accuracy of the National Weather Service’s official forecast, which triggers the country’s emergency alert systems and powers countless commercial weather products.

To make sense of the data, we made a map of the contiguous United States showing how many days into the future the Weather Service’s forecast of the daily high temperature is accurate within 3 degrees Fahrenheit.

map of weather forecast reliability

Source: National Digital Forecast Database. The analysis includes forecasts from April 2023 to March 2024.

Our map shows the accuracy of the forecast, not the skill of particular forecasters. Some places are just less predictable than others.

Coastal regions can be easier to forecast because the ocean acts like a giant thermal regulator, absorbing the sun’s warmth during the day and gradually releasing it over time. Similarly, the southwest desert is relatively predictable because its arid conditions discourage the formation of disruptive weather systems.

The vast middle of the country lacks these moderating factors. Instead, air masses frequently converge with one another: A warm, moist air mass from the Gulf of Mexico, for example, might surge northward and meet a cold, dry mass sweeping down from Canada. Such interactions can cause rapid, unpredictable temperature swings that throw off the forecast’s accuracy.

The Weather Service has for years measured the accuracy of its official forecast in larger geographic regions, but in November 2022 the organization began testing its accuracy at a much more granular scale. The Weather Service had not shared the more granular data with the public until we asked about it.

With the new dataset, we were able to calculate forecast accuracy in thousands of towns and cities, which you can search below. Because accuracy often varies depending on the time of year, we are also showing the accuracy during the warmer months (April through September) and the cooler months (October through March).

In predictable places such as Phoenix and Miami, you can choose any date at random — Jan. 20 or June 14 or wherever your dart lands on a year-long wall calendar — and the temperature will be about the same every year. But in a place like Missoula, Mont., next year’s Jan. 20 might be nothing like this year’s.

Winter in Missoula is extra hard to predict because the city sits at the convergence of five valleys, said Lewis Dortch, a meteorologist at the CBS affiliate in western Montana. During winter, cold air sinks down into the city but is too heavy to escape over the surrounding mountains. Instead, the cold gets trapped — sometimes for a day, sometimes for weeks.

During these “inversions,” computer forecasts can be several degrees warmer than the observed temperature, Dortch said. The computer thinks the frigid air has escaped the city, when in reality it’s stuck sloshing around. Then, as suddenly as it formed, the inversion will lift, and the temperature will soar up into the 40s with little warning.

“It’s just one of those things we deal with out here,” Dortch said. “The topography really makes forecasting those temperatures difficult.”

But the prize for the country’s least accurate temperature forecast goes to Paonia, a town of some 1,500 people nestled in the North Fork Valley in the western Rocky Mountains. Like Missoula, Paonia is vulnerable to inversions. On average, Paonia’s 24-hour forecast is off by 5.9 degrees.

“That’s amazing and not surprising,” said Hannah Stevens, executive director of the Western Slope Conservation Center, when told that her hometown has the least accurate one-day weather forecast in the country. “We all look at two to three different weather apps to get some idea of what might be coming. None of them are ever right.”

The Weather Service provided just one year of data, so Paonia may not have the worst forecast every year. Recall, too, that we considered the accuracy of the forecast for only daily high temperature. The Weather Service also assesses its precipitation forecast, but the patterns are a little harder to unpack — especially with just one year’s worth of data — so we stuck with temperature.

To produce its official forecast, the National Weather Service blends data from 44 computer models — produced by the U.S. government and meteorological services in Australia, Europe and Canada — using a recipe that favors the models with more accurate track records.

To process the data, the Weather Service uses “Dogwood,” a Hewlett Packard Enterprise supercomputer in Manassas, Va., that ranked as the world’s 49th-fastest computer when it was inaugurated in 2022. Its backup, “Cactus,” is located in Phoenix.

Dogwood and Cactus are just the latest weapons in the computational arsenal that has steadily improved the government’s forecasts over time. The National Weather Service does not have long-term accuracy data for its official digital forecast, which it began producing in 2002. But a unit of the Weather Service called the Weather Prediction Center has been producing temperature forecasts since the 1970s.

That data shows steady improvement over the past five decades. “When you were a kid, do you remember the forecast being as good as it is today?” said David Novak, director of the Weather Prediction Center. “It’s one of those quiet revolutions.”

Even as computer-generated forecasts have improved, the Weather Service still makes room for human expertise. At 122 offices around the country, human forecasters use computer software to edit the official forecast based on their knowledge and experience of local conditions. Those edits tend to be minor, however, and their role has diminished as the computer forecast has improved.

“There’s people considering, what should the future role of the forecaster be?” said David Ruth, the chief of digital forecast services at the Weather Service’s Meteorological Development Lab. “Is it worth our time and effort to actually try to make changes to the model’s first guess?”

As a new generation of AI-powered weather forecasts emerges, the future of the human forecaster looks increasingly dim. Just as chess grandmasters now struggle to discern the brilliant subtleties of far-stronger computer chess programs, human forecasters may soon be unable to grasp the inner workings of the most accurate computer forecasts.

“That makes some people nervous,” Ruth said. “But if the results are good and the verification statistics are good, then, yeah, why complain? We may no longer have the privilege of being able to understand the answer why.”

Dortch moved to Missoula for the scenery rather than the unpredictable weather. Yet he has come to embrace the region’s meteorological unreliability. It means human forecasters like him can still add some value to the brute power of machines such as Dogwood and Cactus.

“I use models, but you can’t always trust the models on surface value. You have to use expertise and experience,” Dortch said. “I like forecasting here, instead of somewhere you just look at the models and read whatever they say.”

Thanks to Simon Ducroquet for the illustrations and to Ian Livingston for the meteorological guidance.