As someone who once spent a lot of time analysing sources of GPS error, it's worth noting that the instantaneous speed indications on a commercial type GPS may well be in error. 2D RMS position errors are now pretty low, thanks to the switching off of Selective Availability (an odd term for deliberate time code jitter I've always thought....).
If seeking ultimate accuracy I'd be inclined to use time and distance between two GPS fixed waypoints rather than rely on the instantaneous indicated ground speed. Speed filtering is notably better on more modern GPS receivers, but is still inherently limited by instantaneous, and essentially random, position variations that can be around the 20m - 30m in each second region. The speed errors are related to ground speed, in that the filtering is intelligent enough to increase the speed averaging time constant at lower speeds. This serves to cause the display to read near zero velocity when stationary, which may fool the unwary into thinking the thing is accurate!
The same goes for height accuracy BTW. Because of the shape of the spherical geometry presented by the best selected Space Vehicle (SV) set for 2D accuracy (which is what commercial sets seek to optimise when selecting available SVs) the vertical errors will always be higher than the horizontal ones.
If anyone has a spare moment, try drawing a scale sketch of the pseudo range tracks from a typical 4 or 5 SV set to a receiver. This will clearly show the vertical geometry problem caused by the relatively low orbital altitude of the constellation. A low orbital altitude is obviously good from the horizontal fix accuracy perspective but is poor for vertical position determination.
Surveyors use phase measurement systems to add position accuracy, but as far as I am aware this technology only works for stationary receivers, as it takes several seconds to average the relative phase from each SV. Similarly Differential GPS (DGPS) corrects the gross errors via a second mean local error transmitting data link, but again this will be the mean position error in that region at the time, not the instantaneous error seen by the receiver, so it doesn't fix the speed error problem effectively.