Weather forecasting has come a long way since the 1970s and is among the greatest science success stories of our time. Forecasts are generally quite accurate, even as far out as five to seven days into the future, and have improved enormously since the 1970s. They save countless lives, and their economic value is estimated to be over $30 billion annually (that estimate is from 2009; the value has probably grown considerably since).
But they’re not perfect, and that’s especially true when it comes to snow. Forecasters, like ourselves, obsess over every detail in snow forecasts and still sometimes struggle to get it right. That was clearly the case with Friday night’s forecast. We’re not wrong for lack of trying; the tools we have just aren’t quite good enough yet. Ask forecasters in Boston how they feel about predicting 7 to 13 inches of snow last Monday when only 0.1 inches was ultimately measured. There was sound logic and effort for that forecast, but the storm subtly shifted.
A similar dynamic played out here Friday night, in a less dramatic way.
Our Capital Weather Gang forecast called for about 1 to 5 inches of snow across the Washington region from south to north. Just trace amounts to a few inches fell. The forecast wasn’t wrong everywhere and it’s not the worst miss we’ve ever had, but is among the more notable ones in the last few years.
Here how much snow fell in some select locations:
- Damascus, Westminster and Elkridge: 3.0 inches
- Gaithersburg: 2.8 inches
- Lovettsville: 2.5 inches
- BWI Marshall Airport: 2.2 inches
- Rockville and Laurel: 1.8 inches
- Crofton: 1.5 inches
- Silver Spring: 1.4 inches
- Herndon: 1.2 inches
- Dulles Airport: 1.1 inchesFairfax: 1.0 inches
- Northeast Washington: 0.8 inches
- Springfield: 0.5 inches
- Reagan National Airport: 0.1 inches
Amounts missed the mark as the snow arrived later, departed earlier and fell at lighter intensity than we expected.
So why did we get it wrong? In short, the models we used as tools to make the forecast simulated too much precipitation and the heaviest amounts shifted to the north. On Friday morning, when we analyzed about half a dozen different models, they all supported the forecast we made.
But trouble started to appear in the afternoon. Ironically, about 15 minutes before the National Weather Service issued a winter storm warning for our western counties, the European computer model — which has the highest skill score of all models — suggested no location in the region would see warning-criteria snowfall, which is 5 inches or more. Its forecast suggested that if every flake stuck, amounts would range from about 1 to 4 inches from south to north.
We were reluctant to scale back our forecast based on the European model by itself. Even though it’s the most accurate model on average, it’s not the best-performing in every single storm. And its snow forecast was only slightly lower than other models. But it presaged a steep decline in amounts in the other models during the afternoon and evening.
A big warning sign was that the high-resolution NAM model — which is among the more reliable in the short-term — cut its predicted snowfall by 50 percent between its simulations released Friday morning and Friday evening.
The NAM model’s decrease motivated us to lower our predicted totals at 9:30 p.m. Friday. Importantly, another high-resolution model known as the HRRR had also consistently shown underwhelming amounts of snow since the morning — it was another clue that more aggressive snowfall forecasts — like a forecast for 4 to 6 inches from the American (GFS) model — would be wrong.
Of course, the adjustment we made at 9:30 p.m. to lower amounts was still inadequate. Actual snow totals were on the low side of our revised prediction. When a forecast is going wrong, it’s psychologically difficult to pull the plug entirely, and there was still some modeling (e.g. the American model) that supported a snowier forecast.
All of this model talk may raise the question: Why do forecasters like us rely so much on them? It’s simple — they’re the best tools we have to predict the weather and, without them, weather forecasting would be set back many decades. We get a lot of snowstorm forecasts pretty close to right and it’s largely because of models. That said, they’re guidance, not gospel. Skilled forecasters analyze the models and apply their experience and knowledge to adjust for their errors and biases. Did we not adequately do that with Friday night’s forecast?
One could make that criticism but without appreciating our thought process. Weather systems that come in from the west like Friday night’s are often moisture-deprived, especially after they pass over the Appalachians. So should we have been more skeptical of the model simulations for at least 4 to 6 inches of snow? Also, temperatures were only marginally cold enough for snowfall. Highs on Friday reached 45 to 50 degrees and were still hovering around 40 at 10 p.m. just a couple hours before precipitation began. Shouldn’t we have taken that into account?
Trust me when I say we considered both of these factors. It was for these reasons I wrote the following in our briefing for this forecast:
Amounts could end up on the low end of forecasts or in our “bust” range if precipitation intensity is a little lighter than simulated and temperatures are on the warm end of predictions. The risk of a bust is highest around downtown Washington and to the south, where above-freezing temperatures could eat into amounts. Also, if the snow moves through more quickly than forecast, that would also reduce totals. Some models show just a couple of hours of moderate to heavy snow.— CWG
Even though we knew moisture could be limited and temperatures were marginal for accumulation, there were two meteorological ingredients we thought could over come these factors. First, models were showing the potential for strong “frontogenesis” — a process in which clashing air masses result in the formation of a localized front and narrow zone of particularly heavy snow. Second, they showed frigid air about 5,000 feet above the ground. Our thinking was that these two factors would work together to cool the air and dump snowfall.
The zone of frontogenesis ended up shifting to the north — and plastered parts of east central Pennsylvania with up to 14 inches of snow. One of the hardest aspects of predicting snowstorms is figuring out where very localized bands of heavy snow will set up because models are bad at this. In the days leading up to this forecast, the models consistently showed strong potential for such a band over Washington. It just shifted.
I hope this perspective offered a helpful, under the hood, lens into our forecast process. And the point of this review is not to apologize for our forecast but to explain that errant snow forecasts like this are still simply a fact of life. The National Weather Service and television meteorologists predicted comparable amounts or even more than we did. But, once in a while, despite all of our best efforts, we’re just going to miss the mark because the tools we have aren’t quite good enough.
In such situations, the best we can do is to continue to try to provide the best possible forecast and communicate its uncertainty. I believe we did that in this case although, with the benefit of hindsight, we probably should have done so more forcefully.
Despite this missed forecast, there are a lot exciting developments in weather forecasting that should, over time, reduce the likelihood of these “busted” snow forecasts. The rapid advancement of AI and machine learning, for example, shows great promise to improve forecasts — but this technology is still in its infancy.
