Tony,

Actually, in the university spinoff company I co-founded, we employed nonlinear filtering WAG (NLFWAG).  It involved us constructing a stochastic model of the distribution of possible results (i.e. how long the actual work takes), and an observation process (i.e. our estimates).  We then estimated the conditional distribution how long things would actually take based on our estimates, and used those as our real estimates.   We did this until we realized our estimates were so bad that no amount of filtering would give us better ones :-)

Larimar:

Wow!  I am noted locally for the sophistication of my WAGs, but your technique is light-years beyond what I am familiar with.  

Actually, I estimate by using data flow diagrams to break the system down into relatively small equal size chunks (largely based on ensuring chunks have about the same number of inputs and outputs).  I then estimate for one chunk and then multiply out.   Not perfect by any means, but often leads to more than adequate results.

Tony