Written on 22 October from the forecast and the coverage as they stood at the end of 20 October.

Ohio had the biggest moves of the day. In the governor's race, Amy Acton's lead over Vivek Ramaswamy narrowed from about 5 points to 2.5 as new polls arrived, including one from ActiVote showing Ramaswamy ahead. In the Senate race, Sherrod Brown's lead over Jon Husted narrowed from 4.6 points to 3.1.

Elsewhere, the Georgia Senate race tightened slightly, with Jon Ossoff's lead falling from about 12 points to 10. In the Florida governor's race, Byron Donalds' lead over David Jolly widened from 2.4 points to 4.1.

The chance of Democrats winning the Senate fell from 63% to 60% across the day, and the House from 95% to 94%.

The part that was not the polls

Some of that fall did not come from voters. One of the three inputs to the national environment is a "bottom-up" estimate: how far each polled race is running ahead of or behind its usual partisan lean, averaged across all of them. It is valuable because it is independent of the national generic-ballot polls.

To stop a few freak results from dominating it, the most extreme races at either end were being dropped. That turned out to be the wrong way to do it. A race sitting near the cut-off could fall out between one run and the next, and the whole estimate would jump. On 20 October three races dropped out and four came in, and that alone accounted for about two-thirds of the fall in the bottom-up estimate. The polls themselves accounted for the rest.

The fix went in that night. Extreme races are now capped at the cut-off rather than dropped, so every race always counts and no single one can dominate. Tests now check that walking one race past the boundary cannot make the estimate jump.

We note it here because a forecast should report its own errors as plainly as it reports the races. A reader who saw the Senate number fall that afternoon deserves to know that part of it was the model.

The money

The day's coverage was about spending. The Guardian and The Hill reported President Trump saying he would have spent $400 million supporting Republican candidates. The Guardian also reported $80 million from Michael Bloomberg for Democrats. Axios reported that House Democrats were having to make hard choices about where to spend limited money.

None of that enters the forecast directly. The model uses money each candidate raises from individual donors, not party or outside spending. Outside spending tends to follow the polls, pouring into races already known to be close, so it says more about where the parties think the races are than about where they will end up.

Ohio's two races

Acton and Brown are running on the same ballot, and both leads narrowed on the same day, which is what the model expects: races in one state share most of their polling error. Brown, a former senator, won statewide in Ohio in 2012 and 2018 while his party struggled there in presidential years. Whether that personal record still counts for something is one of the open questions in this race, and the polls are the only thing that will answer it before November.