I have noticed that immediately before the election they are accurate, but I have felt for a long time that 6-8 weeks out from the election that they weren’t.
It is almost like the data radically shifts.
Most people wouldn’t measure someone’s accuracy to forecast only 1-3 days before an known event, they would do it based on several points in the past to see how well it predicted the final outcome.
I prefer to use weeks, 1, 4, 13, 26
This is despite the fact that twice as many Americans self identify as conservative compared to liberal.