Good evening to you from SCW!
It’s no secret; we had a big forecast bust last week. Our forecast for a warning-level snowfall for most of the state ended up as a couple of inches of snow (if that – some folks got nothing at all) for almost everyone aside from some lucky towns on the shoreline. We took a lot of heat for that bust, and rightfully so; you depend on us to get the forecast right, and in that event, we got it wrong. In this discussion, I want to show you why we got it wrong and what was going through our minds when we made the call that we did, with hopes that you’ll not only understand what happened with this storm but what goes into making a storm forecast.
6 days out – The Signal
This storm was on the models in some form from fairly far out in time. We first mentioned it in GP’s disco on 3/15(6 days out) where he outlined the model solutions and explained that they showed everything from suppression to a snowstorm to a rainstorm – the models agreed that there would be a storm in the region, but didn’t agree on much else regarding the track. Things change rapidly at that time so there’s no point in getting into details, but we had pretty good model consensus for a hit of some sort.
4-5 days out – The Fakeout
Of course, as soon as GP mentioned the storm, it disappeared, which seems to happen every time you stick your neck out on a long range threat. Models kept the system safely out to sea, with little to no impact to the state. While we never let our guard down, it started to look like this would be a non-event.
2-3 days out – The Return
Just when we were ready to write this system off, it did what just about every other system this winter has done and jumped northwest around the 72 hour mark. Over the course of the day on that Sunday and into Monday, we saw a shift of several hundred miles in the model guidance; with a direct hit now being shown on several models. That trend continued right up through forecast time on Monday, and when we sat down to make the first call map, the models looked something like this.
It’s no secret; we had a big forecast bust last week. Our forecast for a warning-level snowfall for most of the state ended up as a couple of inches of snow (if that – some folks got nothing at all) for almost everyone aside from some lucky towns on the shoreline. We took a lot of heat for that bust, and rightfully so; you depend on us to get the forecast right, and in that event, we got it wrong. In this discussion, I want to show you why we got it wrong and what was going through our minds when we made the call that we did, with hopes that you’ll not only understand what happened with this storm but what goes into making a storm forecast.
6 days out – The Signal
This storm was on the models in some form from fairly far out in time. We first mentioned it in GP’s disco on 3/15(6 days out) where he outlined the model solutions and explained that they showed everything from suppression to a snowstorm to a rainstorm – the models agreed that there would be a storm in the region, but didn’t agree on much else regarding the track. Things change rapidly at that time so there’s no point in getting into details, but we had pretty good model consensus for a hit of some sort.
4-5 days out – The Fakeout
Of course, as soon as GP mentioned the storm, it disappeared, which seems to happen every time you stick your neck out on a long range threat. Models kept the system safely out to sea, with little to no impact to the state. While we never let our guard down, it started to look like this would be a non-event.
2-3 days out – The Return
Just when we were ready to write this system off, it did what just about every other system this winter has done and jumped northwest around the 72 hour mark. Over the course of the day on that Sunday and into Monday, we saw a shift of several hundred miles in the model guidance; with a direct hit now being shown on several models. That trend continued right up through forecast time on Monday, and when we sat down to make the first call map, the models looked something like this.
As you can see, there was unanimous consensus for at least 6” of snow in almost all of the state, with only the Euro being a bit lower in the NW corner. In fact, some of the models were so strong with the storm that the SE corner ended up flipping to rain and/or dryslotting, keeping totals down there. Combine that with the fact that the models had been continuously ticking northwest for 4-5 cycles now, and Don and I were comfortable going big with our first call map, which featured 6-12” for most with the heaviest snow falling during the PM commute. We recognized concerns around warm temps and sun angle, but believed that heavy rates would offset those factors and we would still see significant accumulations.
1 day out – The Uncertainty
Forecasts are supposed to converge as you approach go time, but in this case, the exact opposite happened. We woke up Tuesday morning to find that the overnight Euro had cut totals fairly substantially, the NAM reduced its totals slightly, and the GFS increased its forecast for the state. At that point, a consensus blend of the three models still verified the low to middle end of our forecast, and so we decided to hold off on making any changes until after the midday guidance rolled in and allowed us to assess all of the data available to us.
Below are the Euro(left) and the GFS(right) from early Tuesday morning.
1 day out – The Uncertainty
Forecasts are supposed to converge as you approach go time, but in this case, the exact opposite happened. We woke up Tuesday morning to find that the overnight Euro had cut totals fairly substantially, the NAM reduced its totals slightly, and the GFS increased its forecast for the state. At that point, a consensus blend of the three models still verified the low to middle end of our forecast, and so we decided to hold off on making any changes until after the midday guidance rolled in and allowed us to assess all of the data available to us.
Below are the Euro(left) and the GFS(right) from early Tuesday morning.
As the midday models started rolling in, everything looked on track. The NAM came in first with a big area of foot plus totals for most of the state, reverting its earlier cut from overnight. Next was the GFS, which backed off a bit from its bullish overnight run but was still in line with our forecast for most areas, followed by the GGEM and UKMET which were perfect matches for our thinking and both increases from their overnight runs. The forecast became clear; 4-8” in NW areas to account for the slight shift south, with 6-12” everywhere else, most in the 6-9” range with some lucky spots with banding seeing up to 12”. We started to sketch out the template for a map as the Euro rolled in, waiting for it to confirm that the trend that we’d been seeing on all of the other guidance was real. Instead, we ended up with a solution that looked like this.
(Sorry for the terrible image – I snapped this on my phone from my computer to send to the rest of the team and it’s the only saved image I have of this run, but I think it’s important to include as it clearly shows the contrast between the Euro and the other guidance).
Oof. While all of the other guidance had a large area of 6”+ across the state, the Euro argued that we would see essentially nothing in the north with a couple of inches in the south. Given the light rates, that solution would likely verify as little to nothing on the ground at the end of the event save the south coast and whoever else was able to get into some banding. It's worth pointing out that for our area, this solution was probably the closest to verifying out of every run I saw for this storm, with the heavy accumulations on Long Island tapering to near nothing in Connecticut being correctly modeled.
While we had the best guidance in the world showing a non-event, it was literally the only model showing a scenario like that, and the Euro isn’t so much better than every other model that I would take it over the consensus of the rest of the guidance. Clearly, we couldn’t stick with the original plan, but how far down to revise totals was the question. We ended up going with roughly a 50/50 blend of the Euro and everything else, hedging our bets that we’d see the camps converge towards a consensus. That’s a common bet that forecasters make, and it turns out that they’re right more often than not, but the difference is that we’re usually making that bet in the 3-6 day timeframe, not 12 hours before an event. I’m not sure I remember another time in my five years of forecasting that we had this much model spread this close to an event.
Gameday
When I woke up Wednesday morning, looked at the GFS, and saw a hit, I felt really good about our forecast, figuring that there was no way that the models could possibly be that far apart at literally zero hours before an event. Nope. The Euro still had little to nothing for most of the state. You may recall my update from that morning essentially saying that we had no idea what was going to happen, and I well and truly meant it when I said we had no idea what was going to happen. Tracking the storm throughout the day, it became clear that we were not seeing heavy banding move north, and by the mid afternoon we largely knew that the storm was a bust, but some snow would finally move in after dinner. With that in mind, we were able to release a new map with expected amounts for the overnight hours, and we got those numbers largely correct. A small consolation prize for the massive forecast bust from the main event.
What happened to keep the main event south of us? Primarily, the culprit was a very dry airmass over us and to our north. The heavy bands of snow were there, and actually briefly made it into the state for a time in the afternoon(especially aloft), but because it was so difficult to saturate the column, the leading edge of the precipitation kept evaporating away as it tried to push north and eventually the storm was beat back out of the state. The more aggressive solutions assumed that we would saturate the column much more quickly than we did, and when that failed to materialize, it became evident that we had a bust on our hands. Furthermore, because precipitation ended up being much lighter, the March sun angle was able to keep that snow from accumulating on roads and other hard surfaces. If we’d seen this event in January, we probably would have still managed a couple of inches in most areas during the daytime, but light snow doesn’t cut it during the daytime in late March. It’s no surprise that we started seeing steadier accumulations after dark.
Conclusion
All in all, the forecast was a bust. That said, I’m not unhappy with how we handled the event. We clearly conveyed the uncertainty, showed the different scenarios, and chose the one that we thought was most likely. To me, the most important lesson from this storm is to not hide anything from the readers. I know that many of you appreciated the fact that we were willing to admit when we didn’t know entirely what was going to happen instead of bluffing and hoping we were right, and it reinforces our long-standing belief that people want to know all of the possible outcomes when looking at an event. With that said, we’d love your feedback on what we can do to provide a better and more informative forecast for you. Drop me a line at [email protected] or connect with us on our social media channels.
With that said, we’ll have a full forecast update tomorrow (don’t worry – there’s no snow in the forecast!), until then, thank you for reading SCW!
-SA
Oof. While all of the other guidance had a large area of 6”+ across the state, the Euro argued that we would see essentially nothing in the north with a couple of inches in the south. Given the light rates, that solution would likely verify as little to nothing on the ground at the end of the event save the south coast and whoever else was able to get into some banding. It's worth pointing out that for our area, this solution was probably the closest to verifying out of every run I saw for this storm, with the heavy accumulations on Long Island tapering to near nothing in Connecticut being correctly modeled.
While we had the best guidance in the world showing a non-event, it was literally the only model showing a scenario like that, and the Euro isn’t so much better than every other model that I would take it over the consensus of the rest of the guidance. Clearly, we couldn’t stick with the original plan, but how far down to revise totals was the question. We ended up going with roughly a 50/50 blend of the Euro and everything else, hedging our bets that we’d see the camps converge towards a consensus. That’s a common bet that forecasters make, and it turns out that they’re right more often than not, but the difference is that we’re usually making that bet in the 3-6 day timeframe, not 12 hours before an event. I’m not sure I remember another time in my five years of forecasting that we had this much model spread this close to an event.
Gameday
When I woke up Wednesday morning, looked at the GFS, and saw a hit, I felt really good about our forecast, figuring that there was no way that the models could possibly be that far apart at literally zero hours before an event. Nope. The Euro still had little to nothing for most of the state. You may recall my update from that morning essentially saying that we had no idea what was going to happen, and I well and truly meant it when I said we had no idea what was going to happen. Tracking the storm throughout the day, it became clear that we were not seeing heavy banding move north, and by the mid afternoon we largely knew that the storm was a bust, but some snow would finally move in after dinner. With that in mind, we were able to release a new map with expected amounts for the overnight hours, and we got those numbers largely correct. A small consolation prize for the massive forecast bust from the main event.
What happened to keep the main event south of us? Primarily, the culprit was a very dry airmass over us and to our north. The heavy bands of snow were there, and actually briefly made it into the state for a time in the afternoon(especially aloft), but because it was so difficult to saturate the column, the leading edge of the precipitation kept evaporating away as it tried to push north and eventually the storm was beat back out of the state. The more aggressive solutions assumed that we would saturate the column much more quickly than we did, and when that failed to materialize, it became evident that we had a bust on our hands. Furthermore, because precipitation ended up being much lighter, the March sun angle was able to keep that snow from accumulating on roads and other hard surfaces. If we’d seen this event in January, we probably would have still managed a couple of inches in most areas during the daytime, but light snow doesn’t cut it during the daytime in late March. It’s no surprise that we started seeing steadier accumulations after dark.
Conclusion
All in all, the forecast was a bust. That said, I’m not unhappy with how we handled the event. We clearly conveyed the uncertainty, showed the different scenarios, and chose the one that we thought was most likely. To me, the most important lesson from this storm is to not hide anything from the readers. I know that many of you appreciated the fact that we were willing to admit when we didn’t know entirely what was going to happen instead of bluffing and hoping we were right, and it reinforces our long-standing belief that people want to know all of the possible outcomes when looking at an event. With that said, we’d love your feedback on what we can do to provide a better and more informative forecast for you. Drop me a line at [email protected] or connect with us on our social media channels.
With that said, we’ll have a full forecast update tomorrow (don’t worry – there’s no snow in the forecast!), until then, thank you for reading SCW!
-SA
