Pretty bad. Dustin Hawkins looks back over the last few election cycles and shows that state polling is regularly off by large margins.
BTW -- I'm no stats whiz, but I have a serious question about the RCP methodology of averaging polls. I understand the notion that adding more and more respondents from multiple polls should theoretically improve the poll by increasing the sample size and decreasing the statistical margin for error. My problem is whether this should be employed when various pollsters are known to be tilting their results to fit a given turnout model.
If 3 pollsters use a turnout model that assumes Obama exceeds all historic levels of turnout and only 1 pollster uses a turnout model that assumes 2010 levels, an average of the 4 polls skews heavily to Obama -- even if all four polls have exactly the same raw data. In essence, all the average does is give priority to how many different polls make the same assumption with the same data.
The RCP averages are not really an average which boosts the size of the sample. It's simply a measure of how many pollsters chose assumption A vs. how many choose assumption B. The assumptions dominate so dramatically, that they dwarf any meaning from the actual responses in the samples.