GFS: Its Strengths and Weaknesses

[Bob Hartig]

After reading another post in Introductory Weather & Chasing, I got to thinking: How does a person go about learning how to piece together numerical models? During my baby steps in forecasting, I used only the GFS. I was aware that NAM and RUC existed, and at times I heard rumors of an ECMWF and GEM and other more esoteric models, but I didn't understand why I needed them. Wasn't one model enough?

Go ahead, smile, but remember that others here are presently in that very same place, wondering how to make sense of the model forecast maps. One of the challenges is, of course, understanding how to connect the dots within a given model. But another concern lies in determining which model has the best handle on a given synoptic situation.

So I thought I'd start a thread that focuses specifically on the GFS. If it proves helpful, other threads can be started for other models.

Here are two simple questions:
What strengths and what weaknesses have you encountered with the GFS?
What are your own "best practices" in using it?

My own knowledge in this area still has a lot of gaps to fill, so I stand plenty to learn in posing the question. However, I'll kick things off with a few of my own basic observations. I welcome correction and/or expansions on the following. And of course, please add your own insights:

STRENGTHS:

• The GFS is the only complete package of long-range model forecast maps available free to the public.
  
• It can alert you to possible synoptic systems that may evolve down the road. Look for consistency in a long-range setup, because it just may hold together until it falls within range of NAM corroboration. But don't be surprised if everything falls to pieces once it draws closer.

WEAKNESSES:

• The GFS becomes increasingly unreliable beyond around 5 days. After that, its potential for inaccuracy hugely increases. At 7.5 days out, or 180 hours--after which the GFS moves from 6-hour updates to 12 out to 384 hours--the GFS becomes more a matter of fortune-telling than forecasting.
  
  That's why pinning your hopes on a long-range forecast is pointless, let along trying to get specific about it. Three days is the threshold for starting to look more closely at the maps. Beyond that, it's a matter of looking for consistency from run to run without worrying about details. The farther out you go, the more it's just a matter of looking for a general consistency.
  
• The GFS tends to be aggressive with its timing. The GFS likes to scoot systems eastward faster than the ECMWF and NAM; then as a system moves to within the 3-day range, it often slows down and agrees with the other models. This is just a generalization, though; there are times when the GFS is accurate in its bullishness. It pays to compare the GFS with the Euro (ECMWF), which is usually more conservative in its timing, and, it's commonly conceded, more accurate overall.

 

[Greg Blumberg]

Bob,

I often don't spend a whole lot of time with the GFS and other numerical models. I'm not very well versed with the weaknesses of the individual models, but I'll say the strategies I use work when it comes to using any numerical model. If I do look at a numerical model, I tend to do a few things first.

The first thing I do is develop my own short term forecast using the forecast funnel. This forecast usually goes out 24-48 hours and can go longer depending on what the hemispheric pattern looks like. The reason I do this is that if the model gets the first few hours wrong, then I have very little confidence about the future model forecast. Models like the GFS can miss the short term forecast in bizarre ways, such as the location a front will be six hours from the initialization time. You would often think that they would always nail simple things like that.

After that, I'll compare my forecast to the numerical models. The models are guidance, and I tend to look at them as additional meteorologists on my "team." If our forecasts agree, then I have much more confidence in both their long and short term assessment. If not, then that requires me to investigate further where the differences came from. Sometimes this requires me to defend my forecast and investigate how valid their assessment is. Sometimes they are right, sometimes they are wrong. When they are wrong, I have to have both the evidence to prove that they are wrong and the confidence to say that they're wrong.

Keep in mind that the strength of a numerical model depends on what scale of the atmosphere you're forecasting for. The development and progression of large scale features are often handled well by models like the GFS, but I wouldn't completely believe derived GFS quantities like CAPE even for a 2 day forecast because so much is dependent on that calculation. If I use numerical models like the GFS, then I'm looking for large scale ingredients like troughs, surface lows, fronts, moisture, etc. It's an ingredients based forecasting approach, and I like to assess how likely those ingredients will come together based on my short term forecast. For example, how likely is it that the Gulf moisture will recover in time for the next shortwave trough?

An SPC forecaster once told me: "There are very few ways you can develop a scientifically sound forecast with deterministic models. You need to use ensembles to do that." Since then, I don't flat out look at the GFS and NAM for forecasts anymore. Ensembles help you incorporate probability and statistical quantities into your forecast. If the solutions converge on the SREF or GEFS, then you can be much more confident that those models are correct about their forecast.

Actually I do know one thing about the weaknesses of the GFS. One key thing that I've learned about models with large grid spacing...be careful of what they parameterize. This includes things like precipitation, convection, clouds, radiation, evapotranspiration, evaporation, etc. The devil is often in the weather phenomena that are sub-grid scale, or the stuff we can't explicitly model. This is often where the real important conflicts and questions come up in my forecasts.

I hope this helps!

 

[Jeff Duda]

At 7.5 days out, or 180 hours--after which the GFS moves from 6-hour updates to 12 out to 384 hours--the GFS becomes more a matter of fortune-telling than forecasting.

Actually, current GFS output from NCEP is 3-hourly up through hour 192, when it then shifts to 12-hourly. There is also a resolution change at that lead time.

The GFS has always fascinated me because it is a spectral model as opposed to a grid point model (such as the NAM and RUC). The only thing I know that means is that it is not grid point based and works on waves of different frequencies. Any more detailed than that, however, and I have no idea how it works. I've never seen a class that teaches how spectral models work either.

As Greg pointed out, the best way to forecast using models, in my opinion, is by the use of ensembles. You can use the SREF, the GEFS, time-lagged HRRR, or other poor man's ensembles from the various world meteorological organizations (i.e., GFS, ECMWF, FIM, GEM, UKMET, for example). A single deterministic model only gives you one realization/sample of the distribution of possible forecast states, whereas having ensemble data gives you a more complete distribution from which you can deduce uncertainty in the forecast. While doing this requires more time and analysis, it usually increases the value of the forecast, especially for high-impact or rare events like severe weather outbreaks.

 

[Rob H]

I personally like using the GEFS and ECMWF ensemble for 180+ hours out to see when a big pattern change is coming. The GFS sweet spot for my use seems to be around 150 hours - it stops flip-flopping and gives me enough information to know when I can expect a chase day or two without being very specific on the exact day. I made a custom STP-like parameter in IDV that I run against the GFS to help with this. Once we're within 84 hours of the event, I only look at the NAM and SREF. On the day of the event, I limit myself to RUC/HRRR.

Basically, I use the GFS to know when I need to take time off at work

 

[CGardner]

Yeah, the GFS is great at tracking significant upper level disturbances or vorticities a week out. I just ran the model and Southern& Eastern OK, SW MO, and SE KS all look very good on Sat. afternoon. Now, 3-4 days out, more than likely going to use NAM or RUC, or by then the SREF. Coincidentally, I will be in the above mentioned area but not for chasing reasons.

 

[rdale]

I just ran the model and Southern& Eastern OK, SW MO, and SE KS all look very good on Sat. afternoon.

How are you running the GFS? I've never seen that used outside of NCEP HQ...

 

[CGardner]

http://mag.ncep.noaa.gov/NCOMAGWEB/...AMER&currKey=model&returnToModel=&imageSize=M

Sorry I didn't clarify, not today. Next Sat.

 

[rdale]

You ran it? Or you looked at it? I'm curious if you're actually running the model...

 

[CGardner]

Okay, backwoods Ozarks coming out. Sorry ya'll, I looked at it.