GFS: Reliable Computer Model? Technical Discussion

1640 reads link

By Quincy Vagell on February 17, 2013, 1:50pm Last modified: February 18, 2013, 1:35pm

Use your ← → (arrow) keys to browse more stories.

Next Article
Model verification for day 5 forecasts of mean sea-level pressure in the Northern Hemisphere.

What exactly makes the GFS a less reliable model than, let’s say, the ECMWF?

Quality of data:
The GFS is a lower resolution model than the ECMWF. Simply put, there is less data plugged into the model. Combine less data with more frequent updates and perhaps it makes more sense why the GFS is so inconsistent and produces (sometimes) widely different forecasts from run-to-run.
When it comes to the long-range GFS, which means forecasts beyond day 7, the GFS is almost worthless to even look at. The model resolution is downscaled after 192 hours, which means the amount of data is decreased. Start with an already “poor” forecast for day 7 and one can imagine how less accurate the forecasts get for days 8 through 16. Yes, the GFS produces forecasts up to 16 days out.

GFS MOS data:
Model output statistics, otherwise known as MOS data, also produces forecasts based off of the GFS. Basically MOS data is a table or spreadsheet of numbers that give a forecast for a specific location. Although this data is helpful, the GFS MOS has some well-known biases. Take today for example. GFS MOS was still predicting highs in the mid-30’s, even as of this morning. However, other data had consistently been pointing to highs only in the mid to upper 20’s. As of midday, temperatures range from 21 to 28 degrees across the state at the official reporting stations. In the other case of snow and ice storms, the GFS MOS data is almost always too mild. There are many reasons for this. The data is often skewed towards climatology, which means in the case of winter storms, the numbers are generally higher than reality. Also, the resolution of the GFS is so poor that the model has trouble identifying local terrain-induced changes in temperature and precipitation.
To go back to the ease of use, it’s incredibly simple to look at MOS data and create a 7-day forecast. Although it’s not a practice that I would recommend, there are many meteorologists who make forecasts based ONLY upon GFS MOS data. Sure, the forecasts are okay, but on a daily basis the GFS MOS can easily be more than 5 degrees off from reality. I’ll be following up with this more at a later date, but after sampling about 30 days worth of MOS data so far this year, the average temperature forecast error for day 1 (a forecast made in the morning for the afternoon high temperature) was about +/- 4F!

Data access:
As a result of U.S. law, all GFS data and forecasts are completely available to the public. This means that is easy for virtually anyone to find weather data from the GFS. The ECMWF and other models have some restrictions, which mean that only some of the data is publically available. This explains why it’s easy to fall back on the GFS. Not to mention the GFS produces forecast data every 3 hours from 0 to 192 hours, while the ECMWF only posts data every 6 hours. On top of that, most public access sites only show ECMWF data for every 24 hours.
Instead of paying for additional model data or just taking the time to find the data, it’s easy for many to fall back on the GFS.

I touched on some of the examples of poor forecasts made by the GFS when compared to the ECMWF, here’s I’ll attach some compelling graphs. The images below show the model verification for day 5. Notice the GFS in a solid black line and the ECMWF with red triangles. The ECMWF is, most of the time, above the GFS. This means the forecast error by the GFS is greater than the ECMWF the majority of the time. (ECMWF correlation is closer to 1)


I’ve covered a lot here and I may post follow-ups in the future to cover more topics. Basically, I’m just saying that the GFS has proven time and time again that it is less reliable than the ECMWF. There are many other models to consider and use, but that will be the focus of a separate article some other time.

Use your ← → (arrow) keys to browse more stories.

Next Article


comments powered by Disqus

Quincy Vagell

Town: Marlborough, CT  

Reporting for WXedge since January 2012.

Articles: 875

Quincy Vagell's Bio

Become a WXedge become a contributor

Let Your Voice Be Heard

Have a question? A comment? A complaint? Our team is here to service your every need. Go ahead, let them have it.