What some people (e.g.
ARG) use as a measure of a pollsters accuracy after the election is the absolute difference between the final poll's result and the final election result.
For example, if the final result turns out to be 51-47 Obama-McCain, then a pollster who predicted 54-45 would be (54-51)+(47-45)=5 off, while one whose last poll showed 48-44 would be (51-48)+(47-44)=6 off, so the 54-45 prediction is deemed "closer" than the 48-44 one.
This is absurd, but anyone who uses this sort measurement gives a pollster a strong incentive to just make their best guess at assigning the undecided voters, regardless of how accurate that guess is.