Well a distinction should be made between the polls and a prediction. Many of the predictions were based on more complicated models that aggregated various polling data and weighted it differently. But as you say, even an 80% chance leaves a 20% chance that it could happen, which is still pretty significant.
Event with the state polls it depends on the specific polls, but so much of this still revolves around the models they use to estimate and the samples they have access to or respond. For those based on the likely voters, they weight/exclude/include variably based on voting in previous elections and how likely they think they are to vote. That can exclude or poorly weight people who didn't vote this time or who didn't vote in the 2008 or 2012 election but were motivated to vote this time. Plus there's still some long-standing methodological issues with the way polls have been run with the rise of cell phones.
Here's a New York Times article discussing some of these biases.
This shit isn't simple, and we're not educating people well enough. Also, pollsters aren't being transparent enough with their methods to allow people to find issues. (The LATimes poll is an exception, as they did take that model and explain what some thought was going wrong.)