You are looking for the leading edge of a very weak signal, when the signal isn’t much stronger than the noise it’s hard to be sure exactly where it started, if you got a positive noise spike just before the signal the leading edge would appear to move forward, if you got a negative noise spike the edge would move backwards. I’m sure there is some horribly complicated maths that would give you a theoretical accuracy limit but from what I’ve seen this noise seems to add around +/- 3 cm to all measurements but will average out fairly well. How much of this is due to the DW1000 not being perfect and how much is unavoidable due to the maths I don’t know.
That just leaves you with systematic errors which bias your range in one direction or the other. Those will be factors of both the DW1000 and the overall system design, both hardware and software. Things like temperature effects or the rotational effect I mentioned before. Good design and clever software tricks can minimize these but you’re never going to eliminate them all. Hence the could of cm error even when averaging.