I think part of it may be that it looks less stable if you're estimating it from the packets of work submitted. I.e., once "warmed up", the hash rate is actually pretty stable as viewed on the money, but the reported hash rate at the pool will look much less stable.