Forecast Model White Xmas Analysis

Two weeks ago, I set out with a mission: To record the GFS computer Forecast Model [JessePedia] predictions for a White Christmas (snow on the ground Christmas morning) every 12 hours, then compare them to who actually had a White Christmas. The following animations of computer forecast model output for Snow Cover and Winter Precip Type on Christmas Morning 2007 are now available in the WeatherMatrix Archive. Both a "fast" and "slow" version of each animation is available, along with all the static images.

- World Snow Cover, Winter Precip Type
- North America Winter Precip Type
- U.S. Snow Cover
- Canadian Snow Cover
- Europe Snow Cover, Winter Precip Type

So you can do your own analysis if you wish, but these are my thoughts, on the GFS and North America anyway... please read the NOTES at the bottom. The NAM had a pretty good forecast but there wasn't a lot of data to go on.

Looking at the United States Snow Cover animation, the GFS generally had a larger snow cover area than it should have been from 384 hours out to 72 hours out, then after that it was more reasonable. This would seem to lend credence to the idea that this model is usually colder and snowier than it should be from Days 5-15. However, the largest snow cover area was predicted at 264 hours (11 days) out (a solution that I made fun of on my blog, and which had snow falling from Pennsylvania through Mississippi and Georgia on Christmas Morning). That said, perhaps the much-maligned "Day 15" forecast may be no worse than Days 10 or 11.

JUST AWFUL: COMPARISON OF GFS 264-HOUR SNOW COVER VS. XMAS MORNING ANALYSIS

Not surprisingly, the smallest areal coverage of snowfall was the last prediction, or the analysis of snow cover on Christmas Morning if you prefer. Then again, the 156-hour (6.5 days) prediction was decent (compared to any other GFS predictions beyond Day 4), except for having snow in the Appalachians and Carolinas. This would seem to lend credence to the idea that the GFS can give you an accurate forecast before the preferred NMM short-term model, which only forecasts out to 84 hours, but if it's only one run out of 30 (twice a day for two weeks), how will know you it's right? (The GFS 84-hour was also decent, save spots in the Deep South).

NOT BAD: COMPARISON OF GFS 156-HOUR SNOW COVER VS. XMAS MORNING ANALYSIS

For the Precip Type animation of North America, it was even worse and seemed almost random for 5 to even 10 days out. However, if you look at the 12Z-12Z and 00Z-00Z comparisons (which I have also animated) you can see a little more consistency. I'm not sure why this is, since they should be using the same data (see notes below on missing images). Maybe less images is better when looking for trends. At least with these you can start to pick out frontal systems and see how the model is tweaking their location from day to day. Notice on the North American animations how between hours 252 and 180 you can see the model "backing up" the front off the East Coast and then between 156 and 84 it does the same thing. This would seem to indicate that the GFS forecasts systems to move too fast between 15 days out and 84 hours out.

But these are all generalizations based on viewing a very limited set of surface weather data which is, in reality, controlled by the movement of upper-level systems, so there might not be any real conclusions we can draw from this. Nonetheless, I thought it was an interesting experiment.

NOTE: When I say "out" below I use the meteorological definition to mean "looking out", so "24 hours out" means the forecast for Christmas Morning as it stood on Christmas Eve Morning. These are also the numbers that appear on the map next to "hour GFS."

NOTE: The 06Z & 18Z Weird Runs [JessePedia] were specifically excluded because they tend to disagree with the 00Z/12Z runs and make animations look choppy.

NOTE: Though I tried to save off each run, the following runs were missing: 12/17 (00Z), 12/18 (00Z), 12/20 (12ZZ), 12/24 (00Z). This makes the 12Z-only animation slightly more accurate to reality since less periods were missing.

NOTE: The "ground truth" analysis is from the GFS itself (12Z 12/25 Snow Cover, 15Z for Winter Precip because the model doesn't predict precip at Hour 00). I think something may also be wrong with the 00Z 12/25 image, it looks way too much like the 12Z analysis. The missing images could introduce inaccuracy because the model is not high-res enough to detect small amounts of snow cover. I posted additional snow cover analyses on my blog on Christmas Day if you would like to look at those instead.

Report a Typo