Written on 22 October from the forecast as it stood at the end of 19 October.

The forecast published at 15:28 Mountain time on 19 October gave Democrats a 58% chance of winning the Senate. The next run, three hours later, gave them 63%. The House moved from 92% to 95%.

Very little of that came from new polls. It came from three changes to how the model works, made that afternoon. When a forecast changes because its authors changed it, readers should be told so, and told what changed.

1. Old polls are moved onto today's national picture

Some races have not been polled for weeks. Before this change, a poll from early September counted as a reading of the race as it stood then, even though the national environment has moved since.

Now each race poll is adjusted by how much the national generic ballot has moved between the day the poll was taken and today, one for one. If the country has moved two points towards one party since a state was last polled, the model assumes, absent other evidence, that the state has moved two points too. This is the "trendline adjustment" FiveThirtyEight documented and used for years. Recent polls are barely affected, and old ones are brought up to date.

Because the generic ballot has moved towards Democrats since the summer, this moved older polls in that direction. In a year when the national picture had moved the other way, the same adjustment would have moved them the other way.

2. The expert ratings are actually used

The model has always been designed to give some weight to the ratings published by the Cook Political Report, Inside Elections and Sabato's Crystal Ball: about a sixth of the forecast in polled congressional races, a quarter in polled governor's races, and more where there are no polls. Before the ratings are used, they are adjusted for the national environment they imply, so they inform individual races without moving the national picture.

Until 19 October, the file that should have held those ratings held none. The expert part of the blend was silently doing nothing. It is now fed from the three outlets' published ratings and refuses to run if it receives fewer than twenty.

The effect shows most in races with no polls, where the ratings carry the most weight. The Colorado governor's race moved from a Democratic margin of 11 points to 17, and the Louisiana Senate race from a Republican margin of 10 to 16. In both cases the experts rate the race as less competitive than our fundamentals alone did.

3. A seat that was being given to the wrong party

California uses a top-two primary, and in the 40th district both general-election candidates are Republicans. The model had been treating the seat as a normal two-party race and, through a quirk in how it handled the missing Democratic candidate, counting it as a likely Democratic hold. It now treats the seat as a certain Republican hold, which it is.

Also new today

The front page gained a map of the Senate seats in play. Clicking a state assumes it goes one way or the other, and the rest of the map and the overall chance update to reflect it. Because state outcomes are correlated, assuming one state changes the others: assuming a surprise in a state like Texas makes similar surprises elsewhere more likely, as it would on election night.

All three changes, and their tests, are in the public repository, and the methodology section on the front page describes the model as it now stands.