538 gave Trump about a 30% chance to win going into the home stretch. People act like 30% never happens. Actually, it happens almost one in three times. Anyone who has ever played Texas hold’em knows the feeling of getting dealt AA, getting all-in pre-flop, and then losing to a worse hand. It happens. Improbable events happen; they are not impossible.
Just like with poker, when it comes to polling, you’re working with incomplete information that points to a range of possible outcomes. You make the best prediction you can with the data you have. I think 538’s analysis was reasonable. I’m not familiar with the other poll aggregators or their methods but, at least as far as 538 goes, I’m not going to give them a hard time for assigning a 30% likelihood to an event that ends up happening.
One thing that is interesting about this is that you cannot get a legitimately random sample of the general public’s political opinion from polls. After all, there are two types of people in this world: people who respond to polls and people who do not. It’s not reasonable to assume that the sample you draw from the type of people who respond to polls represents the type of people who do not respond to polls. These two groups likely have other differences as well and, as such, this represents a bias, and conducting analysis of poll responses as though they represent random samples of the general population is likely to produce systematic errors.