This forecast brought to you by math
How computer-driven weather prediction can help us better prepare for extreme storms.
ON A LATE WINTER AFTERNOON, IN A WHITE and glass office building 30 miles north of Boston, Jim Lidrbauch, a software engineer at the Weather Company, plugs his laptop into a giant monitor at the head of a conference table and fires up the program they call HOTL—for “human over the loop.” It’s half Google Earth, half time machine. A red slider at the bottom of the screen lets meteorologists fast-forward the weather like a movie. As Lidrbauch demonstrates, the map becomes pocked with polygons, as if a 5-year-old went nuts with a drawing tool. Each shape represents a change made by a human.
Since 2015, the Weather Company’s computerized system has run autonomously, generating weather forecasts on demand some 25 billion times a day for people around the world. The polygons represent the meteorologists’ last vestige of active involvement. One specifies “no ice,” telling the forecasting engine to output either rain or snow, but not sleet. Another blob, over Georgia, is the work of someone in Atlanta. It instructs the system to increase the cloud cover by 5 percent and decrease the temperature by 1 degree. “A tweak,” Lidrbauch says.
At the far end of the table, Lidrbauch’s boss, Peter Neilley, squints over the top of his laptop. “We have an active filter that says decrease the temperature?” he asks. “Where is that applied to? What period of time does that cover?”
“This afternoon,” Lidrbauch says. He clicks around to see who made the change: a meteorologist named Juan. Had he looked out the window and not liked the way the system agreed with the sky? Was he bored?
Neilley sighs. “Subtracting 1 degree Fahrenheit from Atlanta for the rest of the afternoon: Was that a good use of time? Humans by nature will fill their time with work.” THE WORK OF METEOROLOGY IS CHANGING. A forecasting office once needed a good view of the sky, but most of the weather you’ll see from inside this building is frozen in plastic logos above the door: the iconography of the Weather Company and three of its affiliates and divisions. There’s the Weather Channel; WSI, the branding attached to its weather “business solutions”; and Weather Underground, the Internet’s first weather site, which the Weather Channel purchased in 2012, right before swapping out “Channel” for “Company” in its name.
As when Apple dropped “Computer,” the move reflected a revolution in the firm’s business. We are all more likely than ever to get our weather reports from an app or a website rather than from a human meteorologist on television. That shift in our consumption has coincided with a quieter one in prediction. Since the 1980s, meteorologists have regularly used computer models—based on the laws of physics and run on powerful supercomputers by government agencies—that spit out details projecting the atmosphere of the future. It used to be left to the meteorologists to sift this raw data into weather forecasts. But over the past decade, the models have improved to the point that they are more than mere “guidance,” as meteorologists like to say. The data they return now is an almost user-ready forecast. The algorithms are taking over. Another logo above the door punctuates that fact: IBM, which, in 2016, purchased the Weather Company (but not the Weather Channel, which operates independently while licensing the Weather Company’s data).
But algorithms don’t write themselves. The system that generates the forecasts has been created over the past 20 years by a team led by Neilley, director of weather forecasting sciences and technologies at the Weather Company. He oversees the development of the back-end engine, doggedly making sure that improvements in the big weather models show up in the weather reports we see on our screens. That work serves not only the Weather Channel website and app, but also the billions of forecasts viewed on Google, Apple, Yahoo, Facebook, and countless other websites and television stations.
For a long time that meant bringing his human forecasters the best possible data, and leaving it to them to push it out to the world. But in July 2015, without announcement or fanfare, he activated a new iteration of the system: From then on, the Weather Company’s machine would no longer depend on human forecasters to feed its predictions directly to our widgets, apps, search results,
and digital assistants. The consequences of that change extend beyond how the weather report is served up each day. It marked the reinvention of the role of the meteorologist.
“I WAS ALWAYS INTERESTED IN HOW WE WOULD make a better forecast for tomorrow,” Neilley says, in a conference room called Tsunami. As a kid in New Jersey in the 1970s, he fell in love with meteorology because he wanted to forecast snow so he could go skiing. At graduate school at MIT in the 1980s, Neilley had a pragmatic streak. While many of his classmates pursued a more theoretical understanding of the weather, Neilley built his own PC and customized an operating system to help move the department’s research data from analog to digital.
His “flap of a butterfly moment,” as he puts it, came in 1997, while he was working as a scientist at the National Center for Atmospheric Research, in Boulder, Colorado. A team from the Weather Channel had come looking for help. They had bought the weather.com domain but realized they couldn’t “just have a bunch of TV guys with grease pencils populating the website,” he says. They needed a new system. Neilley, who was working in a lab focused on practical research applications, wasn’t invited to the meeting, but, eavesdropping from his office across the hall, grew exasperated enough to burst in and tell them they were thinking of it all wrong. The system they wanted to build would turn the methods of human forecasters from around the world into a programmable logic. Neilley saw that it wouldn’t scale. The alternative he sketched— which blended outputs of multiple models— got him the job, eventually bringing him here to Massachusetts, and started him on the two-decades-and-counting challenge of constantly improving the results.
For its first 15 years, the forecasting engine worked like a funnel. Into the wide end Neilley’s team poured a range of inputs, from real-time observations to statistically finessed weather-model data. At the narrow end was a human. A staff meteorologist would take the automatically generated forecast as a first guess, make any improvements they thought necessary, and send it out to the world. “The human always controlled the publish button,” Neilley says.
That worked fine for a decade. But by the late aughts, the narrow end of the funnel had stretched unmanageably wide. The models were getting more accurate, for more days in advance. They were getting more precise, with higher spacial resolution. And there were more of them to consider, with the addition of new ones that better accounted for the chaos of the atmosphere. The humans simply couldn’t check the computers’ work quickly enough. Meteorologists were adding less to the process. A decade ago, Neilley asked his team to stop changing the temperature.
“Look,” he told them, “when you modify the temperature forecast, you’re making it worse as likely as you’re making it better. It’s just not a good use of your time!”
Neilley was also frustrated by how much detail the human-in-the-loop system left out. The switch from desktop to mobile meant people were accessing forecasts more often and from more places. The weather models had the right data, but the human forecasters couldn’t keep up with it. If Neilley could break the logjam at the end of the process in Atlanta, where most of the Weather Channel’s forecasters sat, his system could give more geographical precision (a cooler temperature by the ocean, for example) and do it more frequently (predicting rain in hourlong increments), providing us with up-to-the-minute forecasts for our exact locations.
Neilley’s biggest concern was preserving “the wisdom of the forecaster.” Regardless of the models’ precision, there was still a gap between its metrics and our reality. The nuance of calling rain “showers” or “storms” is hard to automate. The solution was to take the forecasters out of
the end of the loop, where they were a bottleneck, and put them “over the loop”—tuning and qualifying the system’s forecasts as needed. In effect, they’d be another input.
“Before, they had to wait until the model came out, and then they did their thing and they posted it,” Neilley says. “Now the forecast is going out whether or not they touch it.”
THOUGH NEILLEY INSISTS ON CALLING THE shift to over the loop an “evolution” not a “revolution,” it’s remarkable nonetheless: Outside of extreme events, the forecasters are no longer forecasting. The Weather Company is at the forefront of this transformation, but the changes are happening across the field. While the National Weather Service has a corps of 2,500 forecasters tasked, in part, with hand-publishing the kinds of forecasts that the Weather Company manages with a staff of 13, it is now testing a more automated system.
“There’s no question we’re going to move to a role that’s more about communication than actually figuring out if, three days from now, it’s going to be 66 or 68 degrees,” says Ryan Hanrahan, an on-camera meteorologist at NBC Connecticut. Yet the recognition of this paradigm shift in meteorology is unevenly distributed. “I think some are in denial that computers can do as good a job as you can,” he says. But that doesn’t mean all the forecasters are being replaced by machines. The Weather Company keeps its humans busy, offering customized services to companies—such as airlines and energy traders—affected enough by the weather to be willing to pay for help judging what its impacts will be.
The National Weather Service is similarly shifting staff priorities. That means spending more time explaining to emergency managers and public-works officials the likelihood and severity of the event. “We’re fundamentally changing where our job actually ends,” says Louis Uccellini, the director of the National Weather Service.
Paradoxically, it’s precisely the improvements in the automated forecasts that make this new emphasis on communication important. When the forecast was wrong half the time, decisions were harder to make: Flights were canceled later, schools closed after snow had already fallen. Today’s forecasts are actionable—often several days in advance. “We can no longer ignore how the information is communicated and used,” Neilley says.
For Uccellini, it means a return to basic principles. “If you look at the Weather Service mission, the first part is to produce and deliver observations, forecasts, and warnings of weather, water, and climate,” he says. Then he references the second part, to “save lives and property and enhance the national economy.” That’s more important than ever. Whether it comes from a person or a machine, a weather forecast is only as good as the decision you make with it.
Andrew Blum’s book about the weather will be published in 2018 by Ecco/HarperCollins.