I think the question should be about how to maximize the net return on time dedicated to development, which may mean spending less time on (and investing less reliance on) estimates rather than expending unbounded effort improving the quality of estimates.