Brendan Nyhan says that the early state by state polling actually does have some predictive value, but that the important caveat to this is that the "toss-up" states are genuinely toss-ups. Tom Holbrook did an analysis comparing 2004 presidential election outcomes to polling in spring 2004 and found that "across all four months [March-June] the poll result called the wrong winner in 17 of the 36 cases in which Kerry's share of the two-party vote in trial-heat polls was between 47% and 53% (this excludes two cases in which the poll result was tied)."
To me, at the end of the day this essentially reenforces the idea that early polling isn't very valuable. We don't need a poll to tell us that John McCain's going to win Utah and Barack Obama's going to win Florida. We do need a poll to tell us Obama's narrowly ahead in Colorado and McCain's not a slight edge in Nevada but those polls are too close to be predictive. We also know that forecasts based on the fundamentals suggest that Obama will likely win the popular vote.