Nate Cohn explains why:
There’s always the possibility that the polls could miss the outcome in a close contest. Polls have missed the result in three close Senate races in the last two cycles. But this year is particularly challenging. The rapid growth of partisan polls has contaminated the polling averages in states where surveying public opinion is already difficult. Many of these partisan polls employ dubious weighting and sampling practices. The combination will make it even harder for polls to nail the result.
So far this year, 65 percent of polls in Senate battlegrounds have been sponsored or conducted by partisan organizations, and an additional 10 percent were conducted by Rasmussen, an ostensibly nonpartisan firm that leans conservative and has a poor record.
Bernstein chimes in:
Adjust accordingly: Be wary of any conclusions about Alaska, Louisiana, Colorado, Arkansas and North Carolina. Unfortunately, the reaction of the usually astute Chuck Todd is to throw out the baby with the bathwater: He tweeted: “Reason #5,324 why averaging polls is useless, now more than ever.” Well, no. Polling averaging is sound. Sophisticated polling aggregation, which can assess how much weight (if any) to assign to less reliable polling and otherwise correct for detectable biases, is even more important, not less, when the polls may be misleading.
Sides also examines the predictive value of early Senate polling. His bottom line:
Ultimately, what we found most accurate at this point in the 2008-2012 campaigns was to weight the model roughly the same as polls for heavily polled races, and still more for lightly polled races. Then, it was best to gradually shift that weighting scheme over the coming weeks, so that our predictions heavily favored the polls by early September.