The GEFS, an ensemble model, basically takes the data from the GFS (it's numerical weather model counterpart) and looks at the different uncertainties that can exist in actual weather observations. The GEFS corrects this uncertainty through its multiple ensemble forecasts during each model run (0z, 6z, 12z and 18z) and looks at the different scenarios that could play out from a single forecast. Despite trying to reduce the uncertainty, especially in the medium-long range, there still exists some spread between each run, hence any forecast beyond a certain point should be taken with caution (spread-skill relationship). As well, this spread is actually quite noticeable when you look at spaghetti plots.
This is all I know.