Twelve days out, the published chances that Democrats win the Senate look like this: PollsMax 38%, Poliwave 55%, Decision Desk HQ 56–57%, this site 64–65%, the Kalshi and Polymarket betting markets around 67%, and Silver Bulletin about 75%.

That spread looks alarming. It is less alarming once you see where it comes from.

The chamber is sitting on the line

Republicans hold 53 seats. Democrats need a net gain of four to reach 51, because a 50–50 chamber is decided by the Vice President, who is a Republican.

Our model expects Democrats to finish with about 51.4 seats on average. That puts the most likely outcome a fraction of a seat above the threshold, which is the worst place to be if you want forecasters to agree. When the expected result sits right on the line, a small difference in where a model centres its estimate turns into a large difference in the probability it prints. In our model, moving the expected count by one seat moves the headline chance by twelve to thirteen points.

So two models that disagree by less than one seat, a gap no one could detect on election night, can publish numbers ten points apart. Silver Bulletin's 75% corresponds to expecting roughly 52.2 seats with uncertainty similar to ours. That is the whole disagreement: about eight-tenths of a seat.

Where the fraction of a seat comes from

We compared our method with Silver Bulletin's published methodology and re-ran our model with each difference switched over, on the same day's data.

A correction to the generic ballot. Over 13 elections from 1998 to 2022, the national "which party will you vote for" question overstated Democrats by about 2.6 points on average, and we subtract that. Silver Bulletin does not describe a correction like it. We checked whether it was really about registered-voter polls, which might be fixed some other way, and it was not: in 2018–2022, the only years where archived polls record who was sampled, polls of likely voters missed by about the same amount as polls of registered voters (2.3 points against 2.6). Removing our correction would raise our Senate figure from 64% to about 68%.

How far to move old polls. When a state has not been polled for a while, both models shift its older polls to reflect how the national picture has changed since. We apply the full national change, as FiveThirtyEight documented. Silver Bulletin now applies "a fraction" of it, saying a partial correction tested better, though it does not say which fraction. Using half would move our figure down, to about 61%. So on this point copying Silver Bulletin takes us further from its number, not closer.

Everything that is not published. How much a model trusts polls compared with fundamentals such as incumbency and fundraising, how large its errors are, and how strongly it ties states together: these settle the last few points, and no forecaster publishes all of them.

The calendar does some of the work

There is one more effect, and every model shares it. With the polls held exactly where they are today, our forecast would read 60% if the election were two months away and 67% if it were tomorrow. Less time means less room for things to change, so whoever is ahead becomes more likely to stay ahead even if their lead does not grow. Part of the steady rise in every published Senate number this month is the date, not the race.

How to read the range

None of these numbers is the right one in a way the others are wrong. Each reflects defensible choices that history cannot yet choose between, because there have only been a dozen or so midterms with modern polling.

What they agree on is more useful than where they differ. Every serious forecast has the Senate close and the most likely result within a seat of 51. Texas, Kansas, Ohio, Maine, Iowa and Alaska are on every list of races that will decide it, and in several of them the average lead is smaller than the typical polling miss.

If you take one thing from the range, take this: a chamber this close to the line can be decided by an error no forecast can see in advance, in either direction.