The Global Forecast System (GFS) is a global numerical weather prediction system operated by the National Centers for Environmental Prediction (NCEP), part of the United States National Weather Service. It produces forecasts of atmospheric and land-surface conditions, including temperature, wind, precipitation, soil moisture, and atmospheric composition.^[1]

The deterministic GFS forecast is run four times each day and covers out to 16 days in the future. The system also provides initial conditions and forecast information used by other National Weather Service prediction systems.^[2]

The GFS is run four times each day, with forecasts initialized at 00:00, 06:00, 12:00, and 18:00 UTC. Each run produces forecasts for up to 16 days.^[3] The forecasts are initialized using analyses produced by the Global Data Assimilation System (GDAS). GDAS combines surface observations, weather-balloon measurements, aircraft reports, buoy observations, radar data, satellite observations, and other measurements within a three-dimensional representation of the atmosphere.^[4] As with other numerical weather prediction systems, forecast skill generally decreases as the forecast period increases because uncertainties in the initial conditions and approximations in the model grow with time. At longer ranges, forecasts are therefore more useful for identifying broad weather patterns than for predicting local conditions in detail.^[5][6]

NOAA distributes GFS analysis and forecast data through NCEP data servers and the NOAA Operational Model Archive and Distribution System (NOMADS). GFS data are also made available through commercial cloud infrastructure as part of the NOAA Open Data Dissemination program.^[8][12]

Information produced by the National Weather Service and published on its government servers is generally in the public domain unless otherwise noted and may be reused without charge for lawful purposes. Users may not claim the information as their own, imply endorsement by NOAA or the National Weather Service, or present modified information as official government material.^[13] This availability allows outside organizations to incorporate GFS output into weather-forecast products and visualizations.^[14]

The accuracy of the GFS varies with forecast lead time, region, weather variable, and the method used to evaluate the forecast. NCEP publishes ongoing comparisons of the GFS with other global models using measures including anomaly correlation, bias, root-mean-square error, and tropical-cyclone statistics.^[15]

A 2026 report from the NOAA Science Advisory Board noted that commonly used large-scale measures, such as 500 hPa geopotential-height anomaly correlation, do not fully represent forecast performance for weather extremes such as heavy precipitation, high winds, and extreme temperatures.^[16] Computing capacity was one constraint on further development of the GFS. Beginning in 2013, NOAA increased the computing resources available to the National Weather Service, including an expansion from 776 teraflops to 5.78 petaflops completed in 2016 and a further increase to 8.4 petaflops in 2018.^[17][18]

The additional capacity supported model development and testing and allowed NOAA to operate higher-resolution versions of the GFS for longer forecast periods. Earlier GFS resolution increases enabled by greater computing capacity were generally followed by improved annual forecast-skill scores, although computing power was only one factor in those improvements.^[19] Forecast skill also depends on observations, data assimilation, model physics, and other elements of the prediction system.^[20][21]

Forecasts of Hurricane Sandy in 2012 became a prominent example of differing performance among global models. Approximately six to seven days before landfall, operational forecasts from the European Centre for Medium-Range Weather Forecasts generally predicted that Sandy would turn toward the United States, while GFS forecasts generally carried the storm toward the central North Atlantic. A peer-reviewed analysis found that the difference in those early forecasts was caused primarily by the models' different treatments of cumulus convection, rather than by differences in horizontal resolution or initial conditions.^[22]

Comparisons between the GFS and other models depend on the forecast period and the variables and regions being evaluated. The 2026 NOAA Science Advisory Board report stated that GFS performance improves when the model is initialized with ECMWF initial conditions and, in a few cases, performs better than the ECMWF forecast. The report cautioned, however, that differences in forecast performance cannot be attributed to a single factor: data assimilation, observation processing and availability, model physics, and interactions between model components all contribute.^[16] Current comparative results are available through NCEP's GFS verification pages.^[15]

Before adopting FV3, NOAA researchers developed and tested the Flow-following, finite-volume Icosahedral Model (FIM), an experimental global model that used an icosahedral horizontal grid and vertically flow-following coordinates. FIM was developed for medium-range and seasonal forecasting and incorporated physics derived from the GFS.^[23] In July 2016, NOAA selected the Finite-Volume Cubed-Sphere dynamical core as the foundation of its Next Generation Global Prediction System and subsequent GFS development.^[24]

• Integrated Forecast System – ECMWF's global numerical weather prediction system
• ICON - DWD's numerical weather prediction system
• Unified Forecast System
• Weather Research and Forecasting Model
• Rapid Refresh (weather prediction)
• North American Mesoscale Model

• NCEP Environmental Modeling Center: Global Forecast System