In the run-up to this week’s blizzard, some serious differences emerged when it came to the New York City snowfall forecast.
On the one hand, there was the National Weather Service, armed with thousands of meteorologists, a newly upgraded forecasting supercomputer, a nationwide network of weather radars and balloons, and satellite technology.* It even sent the Hurricane Hunters to fly through the storm to take additional data.
Early on, the NWS called for “historic” snowfall totals of 20 to 30 inches in New York City. It cautioned that if an intense snowfall band ended up camping out over the city (as several model forecasts suggested would happen), 3 feet wasn’t out of the question.
That obviously didn’t happen. One NWS forecaster in Philadelphia went so far as to issue a public apology via Twitter for the botched forecast. The director of the National Weather Service also scheduled a Tuesday afternoon press conference to discuss the forecast, a rare step likely aimed at restoring public trust. (The NWS forecast for Boston, by the way, is right on track. That city’s still on pace to record its biggest snowstorm in history.)
So what went wrong in New York City? In a series of Facebook posts explaining its forecast, the NWS said, “These bands are nearly impossible to predict until they develop.” True, but someone did manage to forecast it: the Weather Channel.
Throughout most the day on Monday, the Weather Channel was forecasting 12 to 18 inches for New York City, while the National Weather Service insisted a record-breaker was possible. By nightfall the Weather Channel had scaled back its forecast even further, calling for 8 to 12 inches. And that’s exactly what fell. As late as 5 p.m. Monday, the National Weather Service was still talking about top-end scenarios of up to 3 feet in the Bronx.
These days, meteorologists rely heavily on computer weather models for everything from temperature forecasts to the tracks of hurricanes to snowstorms. Usually, they’re pretty good. But the problem is, they frequently disagree—and when that happens, you need to quickly assess what information to use and what to toss. Which is where the humans come in.
As best I can piece together, the Weather Channel’s method for forecasting storms like this is not to throw out any model information, no matter how off-base it may seem at the time. And for this storm, the potential spread of model forecast placement of the most intense snow band was exceptionally large for the New York City area. This is a perfect situation where probabilistic forecasts are useful. Instead of banking on one or two specific models like the NWS did (and which turned out to be the wrong ones), the Weather Channel chose to blend the models and weight them a little more equally.
In contrast, the National Weather Service took the gutsy step of disregarding the GFS model—its own, newly improved model, by the way—and opted almost entirely for a blend of the ECMWF and NAM, two of the historically best-performing models for this sort of storm and lead time. Though that’s a more traditional style of forecasting, it’s prone to busts. Sure enough, the storm’s center ended up tracking about 120 miles farther east and about three hours faster than the ECMWF forecasted, closer to where the GFS predicted it would be, a pretty big difference.
For my forecast updates here on Slate, I sided almost exclusively with the numbers from the National Weather Service because 1) they seemed reasonable, looking at much of the same data myself that they used to construct this forecast, and 2) they’re the official forecast source. Entire governments make plans based on their forecasts, so I figured that’s good enough for Slate. I was wrong.
On Tuesday, I spoke with the Weather Channel’s winter weather expert Tom Niziol by phone in an attempt to understand how he and his colleagues got it right—and, by extension, what the National Weather Service may have missed.
“As we saw the guidance changing throughout the day, we adjusted that forecast as necessary,” Niziol told me. “I don’t want to make it sound too simple.”
This forecast was truly complicated. The Weather Channel’s forecast for a 30-mile stretch including New York City ranged from more than 18 inches on Long Island to less than 8 inches in New Jersey. It really could have gone either way. In contrast, the Boston forecast, Niziol said, was a “no-brainer.”
In recent years, there’ve been several weather-related controversies in New York state, suggesting that a storm like this was bound to become political. The twin hurricane threats of Irene in 2011 and Sandy in 2012 produced vastly different effects in New York City, despite the National Weather Service calling for a significant coastal flood in each case. Irene turned out to be nearly as catastrophic in upstate New York as Sandy was in the city, but downstate residents likely remember Irene as a busted forecast. Then New York City Mayor Michael Bloomberg took heat during the run-up to Sandy for initially downplaying its flooding threat, probably because of Irene, at least in part.
More recently, New York Gov. Andrew Cuomo was criticized for blaming the National Weather Service for a sluggish response to the Buffalo mega-lake-effect storm, which brought more than 6 feet of snow to the region in just three days.
National Weather Service forecasters (and the governor) likely had these experiences in mind this week. The NWS may have overstressed the blizzard’s worst-case scenario, and the governor may have acted too hastily, both in an effort to cover themselves in case it actually came to pass.
Niziol, who used to head up the Buffalo office of the National Weather Service, knows a lot about messaging during major weather events. He told me that emergency managers frequently want the most likely forecast and the worst-case scenario as well. In the aftermath of this storm, local leaders used that philosophy to explain the unnecessary shuttering of NYC schools and public transportation.
There’s one emerging technology that may provide a relatively quick solution: probabilistic snowfall forecasts. Several National Weather Service offices, including Washington, D.C., New York City, and Boston, have experimental pages right now, though they frequently don’t emphasize them in public messaging.
In my blizzard preview post on Sunday, I actually said “there’s very little chance of a bust.” I came to that conclusion by using the experimental probabilistic snow forecast by the NWS, which at the time showed a 67 percent chance of at least 18 inches in New York City. Although this kind of probabilistic information is exactly what Niziol was talking about, the National Weather Service (and I) got too concerned with the top-end scenarios. I should have played up the other side of that uncertainty more.
This story has a valuable lesson for the future of meteorology, and more importantly, real-time communication of extreme weather events in an era of escalating climate change. If we have the information (and, in the case of the newly upgraded GFS, it’s being produced at great expense), why not use it?
*Correction, Jan. 27, 2015: A previous version of this post incorrectly said the National Weather Service has the ability to launch its own weather satellites. Actually, the NWS parent organization, the National Oceanic and Atmospheric Administration, works with NASA and the U.S. Air Force to launch satellites, whose data the NWS then uses.