Yeah, its a really hard thing to do... the ground-segment (Monitor stations) isn't my speciality (I worked on an AGPS system using ephemeris predicition for a year or two) but pretty much everything you can think of affects the accuracy, weather (at multiple levels in the atmosphere), landscape, topology, season (planetary motions),etc. The guys who do the ephemeris prediction model at JPL would know more tho. But the take-away message is, yes, its really hard stuff.