The count against the call.
Written Tuesday 3 November at 8:45pm Eastern, with most of the country still counting. Everything on this page is partial and preliminary. Nothing on it is a race call, and nothing on it changes what we filed.
What we filed.
Locked 27 October, rerun with the final polls on 2 November. Same method, same code, same seeds. Only the polls changed. These are the numbers we will be scored against.
| Filed | Senate | Governors | Expected Senate seats |
|---|---|---|---|
| 27 October | 67% | 48% | 51.4 |
| 2 November | 76% | 53% | 51.9 |
The model starts from the polling average, adds a turnout term, and then takes 0.93 points off the Democratic margin everywhere, because polls have overstated Democrats by roughly that much in competitive midterm races. Hold on to that last step.
What the count is showing.
Four races with enough counted to be worth looking at. The model column is the margin we published; the returns column is where the count stood between 8:15 and 8:45pm Eastern.
| Race | Our model | Returns | Counted |
|---|---|---|---|
| Florida governor | R+6.7 | D 49.4, R 49.3 | 91% |
| South Carolina Senate | R+2.7 | D 52.3, R 46.0 | 50% |
| Georgia Senate | D+6.6 | D 63.7, R 36.3 | 55% |
Florida is the one to look at, because it is nearly finished. At 91% counted the governor's race was running about seven points more Democratic than we had it. By 9:11pm, with Florida about 98% counted in both races, that settled at 5.0 points, and 5.0 is the figure the live model now uses. South Carolina is running further still, around nine, but half the vote is outstanding and that number will move.
Georgia will narrow and should be read with that in mind. Georgia counts its early and mail ballots first and those lean Democratic, so a lead at 55% counted is not a lead at 100%. We are including it for completeness, not as evidence.
Returns via Associated Press, as carried on the Fox News state results pages, and South Carolina from NBC News at 8:41pm Eastern. Preliminary and subject to change.
If this pattern holds, our call was too cautious.
We reran the model with the polling miss set to the size the early count suggests. This is a sensitivity check on numbers that are still moving, not a new forecast, and it is not what we get scored on. The filed call above is.
| If Democrats beat their polls by | Senate seats | Governors | House seats |
|---|---|---|---|
| 4 points | 54 | 29 | 246 |
| 5 points, what Florida settled at | about 54 | about 30 | about 252 |
| 8 points | 55 | 32 | about 274 |
Treat the outer rows as arithmetic rather than expectation. They are what the model returns when it is handed a national miss of that size. The middle row is the one with evidence under it, and that evidence is still a single state. Ohio, Georgia, Michigan and Texas will move it.
Where the miss would be, if there is one.
The midterm correction pushed the wrong way
We took 0.93 points off the Democratic margin in every race because that is the historical average in competitive midterms. If Democrats are beating their polls tonight, that correction did not just fail to help. It moved every single race in the wrong direction, and it is the one adjustment applied everywhere.
Turnout is the likeliest culprit
Early voting ran well above the last midterm. Likely-voter screens are built from who voted before, so an electorate with a lot of new or irregular voters in it is precisely the one those screens get wrong. That is the same gap the turnout term exists to close, which makes tonight interesting rather than only disappointing.
The turnout term may have been right and too small
The term pushed Democratic in every race that carried it. If the count keeps running that way, the direction was right and the size was not. Being under-confident in the right direction is a different problem from being wrong, and it is a fixable one.
And the caution that still applies
One election cannot separate our model variants from each other. We published that before any votes were counted and it is still true at midnight. A big shared miss is measurable tonight. Which version of the turnout term is best is not.
What happens to this page.
It gets replaced by the scored version: every race we filed, against the certified result, with the misses counted and named. That page goes up whether the call looks good or bad, and the three filings stay exactly as they were published.