After our forecast of a coating to two inches of snow in the region, most places saw no accumulation Thursday morning. Some spots didn’t even see a flake, only raindrops.
How did forecasts end up being so wrong? The flawed predictions can be traced to computer model errors and the inability of human forecasters to adequately account for them.
What the models got wrong
As of Wednesday afternoon, the majority of computer models predicted 1 to 3 inches of snow in the region, with a flip from rain to snow around 6 or 7 a.m. Some models even projected up to four or five inches.
What was particularly worrisome was that high-resolution, short-term models, which were very accurate for the past three snow events this month, showed a strong signal for heavy snowfall rates (around an inch per hour) and temperatures falling close to freezing between 7 and 10 a.m., coinciding with the morning rush.
After 7 or 8a Thursday morning, conditions will go down quick in DC area. Models showing strong signal for burst of 1"/hour snow through around 9 or 10a. Here's the one-hour snowfall forecast between 8 and 9a from hi-res NAM model. Some may melt w/ temps initially above 32. pic.twitter.com/9WSLhZgvrf— Capital Weather Gang (@capitalweather) January 19, 2022
Suffice to say, most models were several degrees too cold with their forecasts and many overdid the intensity of the precipitation.
Were there warning signs that the model forecasts could be wrong?
There were some signs that the models could be wrong, and we did consider them. While many model forecasts showed the potential for two or three inches of snow or even a bit more, we recognized some of the snow they were simulating would melt given mild temperatures ahead of the storm.
For our snowfall forecast map (see below), we essentially cut the average model snowfall simulation in half.
But we did not put enough stock in the model simulations which showed little or no snow. On Wednesday morning, the Canadian and UKMet models showed only around a coating. The afternoon run of the European model, which became available around 7 p.m., also projected only a coating would fall.
Nevertheless, the American (GFS), NAM, SREF and HRRR models, which did a great job with their simulations for the three previous January storms, were consistent in their 2 to 3-plus inch projections.
Our update at 10 p.m. Wednesday, which we also shared on social media, explained our forecast thinking given some of the contrary data:
A couple models (including the European) have backed off their forecasts for snowfall intensity which would also mean a slower decrease in temperature and less snow accumulation. If these models are right, the impact of this event would be rather modest with perhaps just a coating on a grassy areas. However, there is enough support for this burst of snow among other models that we don’t think adjustments to our accumulation forecast are necessary.— CWG
In the two hours that followed our 10 p.m. update, some of the short-term models (like the HRRR) began to slow the transition from rain to snow. By the time I went to bed Wednesday night (close to midnight) more doubt was creeping into my mind about whether meaningful snow accumulation would occur.
When I awoke at 5:45 a.m. Thursday, the writing was on the wall that a snowfall bust was becoming inevitable. Temperatures were still around 40 degrees and short-range models had pushed back the switch to snow to 9 or 10 a.m. Our 6:20 a.m. update stated snowfall amounts would be at the low-end of estimates.
Over the next two hours, the period of snow vanished from short-range model simulations amid a cold rain over the area. We dialed back our snow forecasts further, and the Weather Service discontinued its winter weather advisory.
Why were the models wrong and should we have picked up on that?
When fronts come in from the north and west, models sometimes have a bias of drawing in cold air too quickly. Some of our biggest winter storm busts over the years have occurred when we’ve accepted model forecasts indicating temperatures would rapidly cool to near freezing with rain changing to snow.
In short, we probably should have been more skeptical of the models. But we gave them considerable credence since they had done a good job with the other events this month. Notably, they nailed the Jan. 3 snow forecast the day after it was 63 degrees.
In other words, some recency bias influenced our forecast which was hard to overcome.
Could we have better communicated the bust potential?
The model data we had access to put us in a bit of a bind. It both showed the potential for a disruptive burst of snow during the heart of the morning commute, but there were those few models, which had been less reliable of late, that suggested something less. We attempted to make clear a bust was possible.
“If the majority of models are wrong and we don’t see a burst of heavier snow, we would expect just a coating or so to accumulate, mainly on grassy areas, with minimal travel impacts,” we wrote Wednesday. “This is our bust scenario.”
Also, note that in our caption accompanying the forecast snowfall map (see above), we wrote: “Totals dependent on whether burst of snow materializes.”
We could have and should have more prominently discussed why a bust was possible. In the spirit of public safety, however, we will always place more emphasis on potential hazards rather than on how the forecast could bust in both our articles and social media updates. We’d rather people be prepared for dangerous weather and for it not to materialize than the reverse.
Don’t be too hard on schools
On social media, I’ve seen schools become the target of much criticism for closing and delaying when there was no snow. I don’t think the criticism is warranted.
Schools were in a no-win situation. Had they opened and the burst of snow materialized, they would have potentially endangered students and staff. While learning is lost and parents are inconvenienced when schools needlessly close, schools shouldn’t be blamed for erring on the side of caution given a forecast even the best meteorologists can’t pin down.
Yes, you could argue schools that closed could’ve gone with a two-hour delay and then reassessed. I probably would have recommended that approach; however as the forecast was a moving target, it’s not clear that would have led to better decisions.
Weather forecasting has come a long away. We have a lot more hits than misses with winter storms compared to a decade or two ago. But the information we have available is still imperfect. For those cases when we’re right on the line between a disruptive event and a nonevent, the effect of small errors can balloon in terms of their practical effect.
We trust that the progress which has made in computer modeling will continue so that these misses become less frequent in the years ahead, but they’re still going to happen from time to time. The National Oceanic and Atmospheric Administration, private industry and international modeling centers continue to make large investments in forecasting.
Forecasters are doing their best with the information they have. Neither we nor our broadcast colleagues or the National Weather Service intentionally hype events to influence decisions or get clicks. Our goal is for people to be prepared, understand the range of possibilities, make good decisions and stay safe.
In this case, we did put too much stock in the models. There’s a saying that you live by the models, you die by the models. But the models are the best tools we have. Weather forecasting would be a fiasco without them.
Yet there’s probably a middle ground we should occupy with our use of models.
For future events, we need to both take what the models give us but also treat them with a healthy dose of skepticism so that we can better apply our knowledge of their biases to improve our forecasting and the communication of possible outcomes. That’s where we fell short Thursday.