The Vugol Vector says who is ahead today. Vugol Tensor asks how much that can be trusted on November 3. It starts from the Vector's call for every race, does not tilt it toward either party, and adds what 26 years of election results say about how far polls miss, and how they miss together.
Democrats 70%Republicans 30%
Most likely about 52 Democratic-side seats; four times in five between 48 and 55.
Most of these seats rest on the Vugol Vector model, not on polls. Their extra width is an assumption we have not backtested, so treat this chamber's odds as rougher than the polled races'.
35 of 35 seats up are simulated from their race data (16 from verified polls).
Democrats 96%Republicans 4%
Most likely about 240 Democratic-side seats; four times in five between 223 and 257.
Most of these seats rest on the Vugol Vector model, not on polls. Their extra width is an assumption we have not backtested, so treat this chamber's odds as rougher than the polled races'.
435 of 435 seats up are simulated from their race data (13 from verified polls).
Democrats 90%Neither 5%Republicans 5%
Most likely about 29 Democratic-side governorships; four times in five between 25 and 32.
36 of 36 governorships up are simulated from their race data (21 from verified polls).
Each bar is one race. The dot is the polling average; the thick bar is where the result lands four times in five, the thin line nine times in ten. When the bar crosses the centre line, either party can win.
Average error of final polls in competitive races, by election year (1998–2024). In 8 of 14 cycles polls were too favorable to Democrats, including 2016, 2020 and 2024. In 2012 and 2018 they leaned the other way. Nothing so far predicts which way the next miss will go.
A poll's stated margin of error covers sampling luck only. Used as a range for the real result, it was right 42% of the time when it claimed 90%. Tensor's ranges, which add the way polls actually miss, were right about 89%.
The line marks the 90% each range promises.
The model and its numbers below were fixed on October 15, 2026, before any result. After November 3 we publish how Tensor, the plain Vugol Vector and the poll average each did, misses included, by December 15, 2026.
Let mr be the Vector's Republican-minus-Democratic margin in race r (polling average where verified polls exist, otherwise the model estimate). Tensor simulates the result as
y_r = m_r + B + e_r B ~ N(0, 2.5²) one draw per simulation, shared by every race e_r ~ N(0, σ_r²) σ_r = 4.6 (verified-poll races); 4.6 × 1.5 (model-only races) R wins race r ⇔ y_r > 0 control ⇔ held seats + seats won ≥ the chamber's majority line
The centre carries no bias term. The pooled mean of the 14 cycle-level biases is +0.75 points (t = 1.06), the between-cycle SD is 2.43, and last cycle's miss does not predict the next (correlation about 0), so the shared term has mean 0 and a sign that is unknown in advance. The 90% range for one race margin is ±8.6 points.
| Rule | MAE (points) | Result vs the plain average |
|---|---|---|
| Plain average (baseline) | 4.18 | - |
| Shrink toward state's last presidential margin | 4.07 | −0.11, 95% CI [−0.25, +0.03]; better in 10 of 12 cycles. Weak, not applied in v1. |
| Constant shift toward Republicans | ≈ 4.18 | +0.006, CI [−0.26, +0.28]; better in 6 of 12 |
| Repeat last cycle's miss; party of the White House | no better | not used |
Pollster and mode effects, regional covariance, fundamentals (generic ballot, approval, economy) as inputs to the error, and any signed late-news term. The race-level error for races with no verified poll is an assumption. Leave-one-cycle-out trains on future cycles, which flatters any learned correction; the only fair test is 2026 itself.
Data: Silver Bulletin raw polls archive (1998–2024) and the 538 raw polls file (1998–2022); file hashes, code and every negative result are published with the paper.
Read the methodology paper: Vugol Tensor v1, by Dr. Simón Dhungana →
They answer different questions and are kept apart. Full method: Methodology.
Computed 2026-10-20 05:13 UTC from the stored race data · An experiment, not a forecast guarantee · Today's Vugol Vector · Methodology
