On a good day, an inaccurate forecast from your weather app might mean you’re caught without your umbrella. But forecasts also impact whether cities send out snowplows, if your local grocery store puts its fans front-and-center for an upcoming heat wave, and when farmers need to protect crops from frost.
As we increasingly rely on apps for our forecasts, bad app info can cause major problems — more than just seeing two different temperatures on two different apps, weather experts say. Meteorologist Dan Satterfield isn’t a fan of many national weather apps. “On the bad days, when the weather is really serious, they can often be really badly wrong,” he said. “And the problem is most people don’t use those apps except on the really bad weather days.”
Most phones come with a built-in weather app, or you can buy ones that promise more accurate predictions with sleek and dynamic graphics. But where exactly do these forecasts come from, and how do you know if you can trust them?
Meteorology without meteorologists
Depending on which app you use, it’s either 63, 64, or 66 degrees Fahrenheit in Seattle, with a 5%, 10%, or 34% chance of precipitation at 11 p.m. Those minor discrepancies might not matter for mild weather, but when there are threats of flooding and severe thunderstorms in Charleston, South Carolina, getting accurate and complete information is crucial.
It’s not that easy to track down just where some apps get their weather data from. “What I would suggest is that, if you have an app, ask them where they get their forecasts from or look in the about page and see how they’re getting their forecasts,” said Eric Floehr, founder of ForecastWatch, which analyzes the accuracy of weather apps and sites. “Then dig in and see if they’re a provider that is consistently accurate.”
In a study of forecast accuracy between 2010 and 2017, ForecastWatch found The Weather Channel to be the most accurate. Even then, it was only accurate 77.47% of the time in 2017. In an another ForecastWatch analysis for 2015 to 2017, AccuWeather was the most accurate for precipitation and wind speed forecasts.
Sites like AccuWeather and Weather.com use the federal government’s DICast forecasting system (or other proprietary ones) to make forecasts based on a number of models. “That takes a lot of resources, money and people and time and all that, so only the biggest providers have their own pipelines,” said Floehr. Smaller providers often entirely rely on APIs and machine learning. “But you’ll find it’s not usually as good as the forecasts from the large providers,” he said. Method matters too: popular app Dark Sky prides itself on being human-free. Naturally, meteorologists don’t like this approach, and it didn’t perform as well as the larger companies in ForecastWatch’s analysis. Dark Sky didn’t respond to a request for comment on this story.
“We have implemented a new work paradigm for our forecasters called ‘over the loop,’ which enables our forecasters to be much more efficient in their jobs,” said a spokesperson for The Weather Company, which owns The Weather Channel and is part of IBM. “In fact, many national forecasting agencies around the world are beginning to adopt some form of this more efficient paradigm.”
AccuWeather has over 100 meteorologists, and The Weather Channel has more than 180. AccuWeather uses 176 models, including from Japan and China, in addition to the European model and NWS’s. It uses both governmental weather data and information from private companies’ sensors. Dr. Joel Myers, the CEO of AccuWeather, explained how meteorologists work with the company’s models and how they match up with the way the weather is currently behaving. “The snow is moving faster than the models are capturing,” he said. “And so we accelerate to start the snow an hour, an hour and a half faster than the models are indicating, and that’s how we get the jump on our competitors.”
99 (more or less) weather balloons
To forecast, government-run weather operations like the National Weather Service (NWS) need a flurry of data, much of which is shared with other agencies around the world. In addition to satellites and radar, NWS relies on weather balloons, gauges to measure river levels, ocean and lake buoys, and sensors on airplanes and ships.
The model mimics the atmosphere, with those billions of data points nudging it in specific directions for what happens next.
There are 122 Weather Forecast Offices around the country, and professionals and amateurs provide on-the-ground observations. Equipment like satellites might use altimeters to measure surface currents, in addition to beaming images back to the NWS. The buoys collect air and water temperature, wind speed and direction, and barometric pressure. All these types of information — NWS says it analyzes over 76 billion observations a year — goes into a forecast model.
That staggering amount of numbers requires massive computing power. In 2018, NOAA updated its supercomputer system. Together, its Dells, IBMs, and Crays can process 8 quadrillion calculations per second. NWS recently updated its Global Forecast System (GFS) model, which simulates how the atmosphere moves and behaves, based on the temperature, humidity, pressure, and other data fed into the system. The model mimics the atmosphere, with those billions of data points nudging it in specific directions for what happens next.
There are a few types of forecasts: short-range, which are often used for events like storms and are good for between 12 hours and two days; medium-range, which give you a view three to seven days ahead; and extended-range. Shorter forecasts are too zoomed in to take into account the storm moving in from a couple states away that’s going to arrive in a few days. Longer-range forecasts cover a larger area, and that increased scope allows them to make predictions based on weather events happening elsewhere in the world. A general rule, though, is that the farther out the forecast the less accurate it will be. Conditions change often enough that just because the prediction is 75 degrees Fahrenheit and partly cloudy seven days from now, it doesn’t mean things stay that way.
These longer-range forecasts from AccuWeather are quite specific. In New York City, for example, the forecast predicts a “morning shower” on July 17. “That is based on a combination of all the models, and so there must be pretty much agreement statistically that the highest probability of a shower on the 17th is in the morning,” said Joel Myers.
There are lots of reasons for forecasts to disagree, even if they both originate with the NWS. Maybe your app, which works just fine in Portland, Oregon, suddenly seems off when you visit Mount Rushmore. In more remote areas, there aren’t sensors and buoys every mile, so forecasts have to fill in the blanks, taking into account the nearest data points. Portland probably has better sensor coverage than parts of South Dakota, but it also has very different weather. Even different seasons will affect the accuracy of weather forecasts.
With words like “predict” and “forecast,” meteorology can seem more pseudo than science.
Thanks to the Pacific Northwest’s oceanic climate, Portland has fairly unchangeable weather in the summer. There aren’t unpredictable thunderstorms like in other areas of the United States. “Warm and sunny” is more reliably forecasted than rising, unstable air. In the summer, thunderstorms can develop quickly in certain areas of the country. Heat waves are easier to predict.
ForecastWatch tracks how accurately different companies and organizations forecast weather. If you play around with its Forecast Advisor tool, you’ll see different sites and apps vary in how well they perform in different parts of the country. And again, forecasting temperature is different from predicting precipitation and tornadoes. While AccuWeather does predict tornadoes, in the past it’s issued warnings only to its commercial customers —businesses that pay for its enterprise solutions. If you see a storm or tornado watch or warning in the free app, you’ll notice it was issued by the National Weather Service. Because sites like AccuWeather and the Weather Channel limit their more complex forecasts to paying customers, it’s difficult to determine how accurate they are for certain types of predictions, compared to the free and open NWS data.
The reason sites like Weather.com and AccuWeather beat the NWS is because of what happens after the models run. The results are compared to previous forecasts and adjusted based on earlier trends, like if humidity routinely skews too high. While the NWS does perform this type of post-processing, the Weather Channel and others use the National Center for Atmospheric Research’s (NCAR) DiCast, which operates a bit like an ensemble model. DiCast combines forecasts from the NWS other global and regional models to improve the final outcome. The result can be totally automated, though sites such as AccuWeather say they still use meteorologists to oversee the process.
“Now a lot of these sensors are connected directly to the internet at basically real-time speed.”
With words like “predict” and “forecast,” meteorology can seem more pseudo than science. Part of the mystery around weather accuracy is that statistics and probability are notoriously difficult for people to wrap their minds around. When your meteorologist says there’s a 30% chance of rain this morning, it doesn’t indicate anything about how long or hard it will rain. It just means there’s a 30% chance it will rain at your house. Thirty percent sounds low, but you might still get wet.
Despite the uncertainty, certain areas of forecasting are improving. Medium-range predictions have gotten more accurate by about one day per decade. In 2015, for example, meteorologists could predict six days out as well as they could predict the weather five days out in 2005. The amount of data feeding into the models has increased, as has the understanding of physics that support them. The data itself is also moving faster. “Back when I started, observational data would get phoned in using a modem,” said ForecastWatch’s Floehr. “Now a lot of these sensors are connected directly to the internet at basically real-time speed.” The data gets to the models more quickly, meaning they can run more often and be more accurate.
There is a point, some studies show, where forecasts just fall apart. Some sites and apps will show you the weather a month from now, but forecasts are mostly accurate up to about 10 days. That might one day reach up to about two weeks. If you see a 40-day forecast, “That’s astrology,” said meteorologist Dan Satterfield. “It’s not science.” Critics have also expressed skepticism about whether AccuWeather could accurately predict the number of tornadoes to expect this season, as the company has claimed it can do.
Having access to NWS data is how AccuWeather started over 50 years ago. Newer apps and sites rely on it as one source for their models, and it provides an important comparison for other global models, like Europe’s and the U.K.’s. It increasingly makes sense for people to rely on apps for life-saving weather information. Tornado sirens were not designed to be heard inside homes and buildings. They can be especially difficult to pick up during a cacophonous storm when you’re asleep. Sirens also don’t cover all weather emergencies, like floods or snowstorms.
If you’re not watching The Bachelorette, a broadcast-interrupting TV alert isn’t much help, either. In the past, these weren’t always relevant because they were more general. “The Weather Service used to issue warnings by county or half a county,” said Satterfield. In some cases, counties can cover thousands of miles. Now they use polygons that cover just where the area near the storm. In these cases, an alert means the storm is close by.
Weather apps can deliver very targeted alerts based on your location. “That’s the only good thing about all of these apps, that we’ve now gotten very specific warnings,” said Satterfield. “No one likes to be tracked these days but this is one case where you want to be tracked,” he added.
Even if you don’t have an app, if your phone is Wireless Emergency Alert-enabled, you’ll still get notifications for severe weather, missing children, and other serious situations. These alerts come from the NWS, state and local governments, and similar entities. Cell phone towers broadcast the alerts to any compatible phones in the area, and they should work even if networks are too jammed to deliver regular texts and calls. “There could be a gazillion people in the area who could receive the alert concurrently, and you wouldn’t be clogging up any communication lines” Mike Gerber, a meteorologist with NWS, said. He compared it to a radio broadcast of a baseball game. “It doesn’t really matter how many people are listening to it.”
Because of the way the phone vibrates and squawks during these alerts, they can be pretty jarring — especially if everyone else’s alert goes off in the same office. Frequent flooding in Kansas City, Missouri led to a spate of alerts in 2015, some very early in the morning. For rural residents, alerts can come for weather events 10 or 15 miles away, if the closest cell phone tower is designed to broadcast over large areas. If your phone isn’t getting a signal, though, the alerts also won’t reach you.
“No one likes to be tracked these days but this is one case where you want to be tracked.”
Officials want to keep people from opting out of these warnings — which you can do via your phone settings and isn’t recommended — by avoiding “overwarning.” The FCC will try to make the alerts more targeted later this year. Eventually your smart TV could pass along a weather warning when you’re watching Netflix or playing a video game, but we’re not quite there yet. The alerts could also be broadcast through a car’s infotainment system or a smart smoke detector.
Gerber says the WEA system is meant to work in concert with other warnings, whether they’re coming from an app or through the TV. “People often look for supporting information from a credible source before they will take action,” he said. Considering there have been a few high-profile false alarms, that instinct is understandable.
The Weather Channel makes the frameworks for many stations’ local weather apps. But Satterfield says he and other meteorologists will update these apps with their own forecasts based on the many models they look at, as well as their local knowledge. Still, he gets blamed every time any app his viewers’ use gets it wrong. “That happens all the time to every meteorologist in television and it drives the Weather Service folks crazy as well because people do weather by icon,” he said.
Instead of just looking at the temperature and whether the picture is showing you sun, snow, or clouds, Satterfield suggests looking for apps, especially local ones helmed by Certified Broadcast Meteorologists, that offer some written analysis. A lightning icon won’t tell you the probability of a thunderstorm or what time it’s likely to hit.
“If the weather really matters,” he said, “go to that text forecast.”
Updated 6/27/2019: Updated to include the ForecastWatch report from 2015 to 2017.
- Meteorologists not swayed by telecom companies’ plans for 5G regulations
- Dr. Cecilia Bitz tracks sea ice as it disappears at unprecedented levels
- This groundbreaking new buoy can predict earthquakes and tsunamis. Here’s how
- Awesome Tech You Can’t Buy Yet: Robot arms and keychain-sized hard drives
- Toyota partners with weather firm to use wiper blade data to predict weather