I am often asked why the likely voter polling data, especially in New York State, has been consistently off target on the margin of victory over the last decade and a half. My response is always the same: The polling data itself is not usually the problem, it is the likely voter screens that are often suspect, sometimes projecting very inaccurate spreads and margins.
A better way to predict electoral outcomes is to project the current year's turnout based upon recent trend lines, then re-weight the polling data to fit the truly "likely" electorate. When you do that, you will more often than not beat the accuracy of the public polls.