“Distinguishing the signal from the noise requires both scientific knowledge and self-knowledge.” -Nate Silver
When you take a poll, you survey a number of people with an opinion about something in an attempt to predict the behavior of a much larger number of people. If you increase the number of people you poll, your poll uncertainty drops. This reduction in what we call a statistical error will mean your polls reflect the likely outcome better and better, given one assumption. You have to assume that data obtained from the people you’re polling are reflective of a random sample of future voters.
And that’s a big assumption! Any deviation from that, in turnout, in voter preference, in sampling bias, etc., will mean that there are additional sources of error that you have no way of accounting for. These systematic errors plague all observational and measurement sciences, and predicting an election’s outcome is no exception.