Salil Mehta has been writing about polling error, and while I don’t always reach the same conclusions he does, I find his work consistently worthwhile and thought provoking.
The underlying issues with polling is similar to economic data and other noisy series. Add in partisan bias and any single poll is quite possibly (even likely) wrong. Hence the approach of averaging all of them makes sense like a smoothed moving average).
Have a look at the spread here, and draw your own conclusion about he reliability of any single poll:
Source: Salil Mehta
What's been said:
Discussions found on the web: