The Margin

Saturday 3 October 2026 · A daily bulletin from election-lab. Composed by a program from the numbers in its own database; no part of it is written by a language model, and every figure traces to a query.

Where it stands

• Democratic control of the House in November: 98%, on an expected 244 seats.
• Democratic control of the Senate in November: 39% to 47%, on an expected 50 seats. A range because seats that could go to independents whose caucus is not recorded are counted both ways.
• Democratic control of the Senate if the election were held today: withheld. coverage incomplete: 93 of 100 seats
• Democratic governors in November: 25 expected of 50.
• Democratic governors if the election were held today: 22 expected of 50.

Read those against the model's own record. Averaged over 230 scored races it has leaned +1.64 points toward the Democrats — but that is an average and not a habit: five of the ten scored cycles lean the other way, and the per-cycle figures run from −3.7 in 2006 to +6.5 in 2014.

So it is not a correction waiting to be applied. A cycle's lean does not predict the next cycle's: the correlation between consecutive cycles is −0.05, which is nothing, and subtracting the running average walk-forward leaves mean absolute error unchanged at 4.92 points while moving the bias from +2.61 to +2.33. Three pre-registered attempts to correct it were refused on that evidence. The figure is quoted so a reader can do the sensitivity themselves, and it is not subtracted from the numbers above.

What moved

Nothing moved enough to report.

What came in

0 polls recorded in the last day, against 16,154 live (55% of them checked against a primary source).

What this does not tell you

• No race moved by 3 points or more since its previous run, so nothing is reported as movement. Smaller changes exist and are not distinguishable from simulation noise by a reader.

Champion specification single_race_latent/0.8.1. Two products per race and never one: a nowcast is what today's evidence says about today, an election-day forecast is about November, and this bulletin labels which it is quoting every time. Methodology at /methodology; the model's own scored error, including where it is worst, at /model-health.