I quoted elbows from the By how much is Trump over-performing, in terms of polling? thread and DSeid because I want to emphasize that my position is in between.
It’s true that polling accuracy is probably at the lowest point in the last 50 years, a period in which experience and money overcame most of the many problems that polling inherently has. That understanding has shredded over the last few years. It’s much harder to poll over mobile phones than landlines. People are less likely to comply. They’ve less likely even to answer, since they can now see who is calling. (People tend to be polite and hate to hang up, especially when asked for their opinions.) The population is far more diverse than ever before and getting representative samples of thousands of subgroups is nearly impossible. Political opinions are far more partisan while at the same time far more people call themselves independents than ever before. After inadequate samples are weighted to an imaginary standard, fine details are lost. I’ve done door-to-door and phone bank polling. It’s an art and a mystery.
And yet polls do converge on the actual results a great majority of the time. How can this be? We need to remember that those of us who have been talking about the presidential election for a year are the minority. Most people will not really tune in until after the conventions. Until then, opinions are often based on name recognition and party labels. Moreover, only a smidge over 50% of eligible voters bother to vote. Some polls differentiate between all voters and likely voters but nobody can be sure who’s an actual voter until November 9. The polls on November 7 will be far more accurate than any today, at all levels. (Remember “Dewey Defeats Truman”? The expense of running polls was ended two weeks before a seemingly certain election, but many people made up their minds in those two weeks.)
My position has been consistent. Any individual poll this far back from the election is meaningless. It may turn out to be right, but nobody can know that. Putting together a number of polls and averaging them knocks off outliers and bad data, but doesn’t rise to the level of accuracy. Trends of averaged polls are a bit better still for the same reason. The RCP chart of daily averages shows that the polls have been consistent and yet volatile. The differences between Clinton and Trump have been utterly obvious from the beginning and Clinton has never fallen behind - yet Trump has risen to tie her on four separate and widely spaced occasions since September. Anyone who talked about their being tied at any of those times was spouting nonsense, no matter that one form of the data backed it up.
Polls do work. But. They work best for the largest elections with the greatest pool to sample from. They work best closest to the election when most people have made up their minds. They work best with two and only two choices. They are not worthless to the parties involved, because they get information on how to apply strategy and money in the future. They are meaningless to casual watchers, and especially anyone who claims that a poll has significance for the final outcome.