Weather Models 101


Numerical weather forecasting (NWF) is the heart and soul of the locally generated forecasts you see on TV and the Internet. Very powerful computers ingest huge amounts of numerical data to simulate (model) the atmosphere on a computer screen. The data originates from the large network of official weather observations, remote sensor measurements (satellite data), and weather balloons that are launched up to two times daily from strategic National Weather Service (NWS) locations. A three dimensional rendering of the atmosphere is produced and is cut into horizontal "slices" at predetermined levels so that the output can be viewed on the web. Terms like "500mb" or "1000-500mb thickness" (to be discussed later) all originate from comparing/juxtaposing these atmospheric slices.

Using incredibly complex math equations that would make calculus professors spit out their coffee, these model-based forecasts are generated up to 3 weeks in the future. They are most accurate within 72 hours of start (initialization) time, and even six days out can give us a big picture of the potential state of the atmosphere. Because there remain many things about the atmosphere that we still don't understand, models lose their accuracy rather quickly.

There are two primary model groupings based upon the type of statistical model used:: probabilistic ("ensemble" or "stochastic") or deterministic ("operational"). These two types of statistical models are basically the opposite of each other.

Probabilistic models ("ensembles") vary the starting conditions slightly, resulting in "what if" scenarios for the atmosphere. The more members of the ensemble that are in agreement, the higher confidence one has in the forecast. Deterministic models ("operational") are the ones that run with a specific set of initial conditions and result in a single conclusion or output.

On the CWG blog you'll read many references to such geeky terms as GFS, ECMWF (Euro), NAM, UKMET, CMC (Canadian), HRRR, RUC, and SREF. These are all weather models that forecast the state of the atmosphere using slightly different equations and resolutions (ability to discern atmospheric detail). They each have various strengths and weaknesses. The important point is not to rely on any one model, but to consult all of them to build a consensus picture of the atmosphere, and monitor their trends over time. This is exactly what the NWS does with their manually created products. Consistency of model output is key. If a model advertises the same result run after run, then confidence in its forecast is relatively high. If, however, the model output flips back and forth between runs, then it is next to useless.

There are also many more specialized models for varying purposes, but those are beyond the scope of this FAQ (for now anyway).

Individual Operational (Deterministic) Models

GFS - Global Forecasting System: This is the primary U.S. weather model upon which many local forecasts rely. The GFS covers the U.S., Canada, and northern Mexico. Because it's produced by the U.S. government, access to its output is free. The GFS output is updated four times daily, at 0Z, 6Z, 12Z, and 18Z. The 0Z and 12Z versions incorporate a new, full observational data set whereas the 6Z and 18Z output does not. The GFS produces forecasts out to 384 hours!

Biases - tends to be a colder, dryer model. In winter, the GFS has a tendency to suppress east coast storms south and east of where they actually wind up. In summer, can suffer from "convective feedback" making it show much more precipitation than will really occur. Convective feedback also negatively impacts downstream features and can render the model run suspect. GFS is typically a very good short range model, especially within 72 hours. Has a tendency to "flip" suddenly as well as lose (shear) big east coast storms at 5 days out when originally seeing them 1 week out.

NAM - North American Model: Another U.S.-based model. Like the GFS, the NAM is free to access and covers the same geographic region. It has a much shorter forecast range, ending at 84 hours.

Biases - this is a warm and wet model, opposite of the GFS. Has a tendency to overdo precipitation by as much as 50% in certain situations. Can be "out to lunch" frequently, but just when you swear you'll ignore it forever, the NAM will shock you by nailing a difficult forecast that other models missed. If it continually advertises a solution different from other models and does not waffle, then it's worth paying attention to it.

ECMWF - European Centre for Medium Range Weather Forecast (aka "Euro"): The Euro is a very different animal from the GFS or NAM. It is a proprietary weather model run by a consortium of European countries. Access to much of the model's output requires a subscription; however, there are elements of the Euro that are freely accessible at sites such as Unisys and Penn State's E-Wall (warn your eyes before clicking on the E-Wall link!).

The Euro is billed as the most accurate of the global weather models. It should be because it has the highest model resolution. It can "see" in 16 km horizontal chunks, whereas other weather models like the GFS can only deal in chunks that are 28 km in size (70 km for forecasts beyond one week). That's a huge difference and has a direct impact on the ability of a weather model to "see" smaller scale processes that have a material impact on larger scale weather systems.

Biases - can underestimate heavy precipitation situations. See this great summary of GFS/Euro compare/contrast info.

Topics to be added to this page:


Frequently Asked Questions (FAQs)
about Weather

An Introductory Guide