Imagine a survey asks you one extra question: not who you will vote for, but who you think will win here, in your own seat. Most people get their own seat wrong. But take the answer that most people in a seat gave, treat that single answer as the forecast for the seat, and it does much better. The grid below has one dot per election. Across the bottom: out of every 100 people, how many named the winner in their own seat. Up the side: out of every 100 seats, how many the majority answer called right. On the dashed line the two are equal, so pooling the crowd would have added nothing.
Election 1 of 4
Your marker sits at 50 out of 100
No dot to plot. In the 2020 US election, expectations in close states Biden won and close states Trump won differed by 0.035 on a nought-to-one scale, which is not distinguishable from zero (p = 0.32). Where the race was safe, expectations were near 0 or near 1; where it was close, they sat near the middle whoever went on to win. So close races belong on the dashed line, in the rust band: the crowd adds nothing there.
The crowd is strongest about the seats that were never in doubt, and that is the opposite of what a forecaster needs.
Huber and Tucker (2023). The band is the pattern the paper reports, not a table of per-state values, which it does not publish.
And some of the gain is not wisdom: in Canadian elections, intending to vote for the winning candidate raised a person's chance of a correct district forecast by 20 to 51 percentage points (Mongrain et al., 2025), so a crowd holding more winners than losers looks wiser than it is.