A clarification: in the whole project, the average underestimation error against the first estimate has been measured as 15.68%.
Limiting the measurement to the last 3 months showed a value of 0.61% ==> in the first part of the project we had an underestimation error of 30.56%
==> therefore the reduction was from ~31% (in the first 7 months) to ~1% (in the last 3 months).
A few questions from KenBitskoMacLeod:
Were the estimates produced by the developers themselves improved to that accuracy? ==> Yes
Or were the developer's estimates adjusted by an error factor? ==> No: we tried it in past projects, but it really didn't work as good as exposing estimates without adjustment. Actually, this is the subject of a paper that I will present at XP2003.