Parameters in force
the fitted numbers the live forecast uses, what each is estimated from, the few shown for completeness rather than applied, and the promotion gate thresholds that govern when the champion model changes
Generated from the committed artifacts at request time, not written down. Not every value below feeds the live forecast: the scale correction's intercept is fitted with the slope but not applied, the seats-votes relationship is an independent check rather than an input, the old-lines margin MAE is shown only for comparison, and the promotion gate thresholds are the rule for when the champion model changes, not a forecast input. Where a parameter is re-estimated per cycle from data strictly before it, the row says so, because that property is what makes the backtest mean anything. Some fitted quantities are in force but have no row here: each pollster's own house effect and excess variance, and the national model's coefficients. Either they run to many values per pollster or per race, or the file this page reads does not hold the values in force.
Correlated error across races
How much of a polling miss is shared by every race in a cycle rather than particular to one. This is the term that makes a chamber forecast three times wider than independent races would imply, and getting it wrong is the difference between a probability and a guess.
| Parameter | In force |
|---|---|
| National polling error (sigma_nat) | 2.632 pts |
| …its own standard error | 0.3799 pts |
| Race-specific residual (sigma_race) | 4.926 pts |
| Total | 5.585 pts |
fitted on 1,226 observations · using only data before 2026 · re-estimated walk-forward per cycle
Technical details: run and model identifiers
Fitted file: experiments/uncertainty_terms_pre2026.json
Scale correction (c15)
Polls understate the leader in uncompetitive races, so a forecast built on them is compressed toward the middle. The law is fitted as intercept + slope × estimate. Only the slope is applied; the intercept, a uniform shift toward one party, is not applied and is shown here because it is fitted with the slope. The scale law is refitted whenever the corpus grows.
| Parameter | In force |
|---|---|
| Slope | 1.07 |
| Intercept | -1.744 pts |
| Residual sd | 5.466 pts |
fitted on 463 observations · using only data before 2026 · re-estimated walk-forward per cycle
Technical details: run and model identifiers
Fitted file: experiments/scale_pre2026.json
Sparse-poll variance
Polls of safe, lightly polled races are wrong by far more than their sample sizes imply. The blend weights a thin race's polls by this measured error instead of by sampling error — which was the single largest calibration defect this project has found. The sparse-race error does not contradict the design effect below one in the poll dispersion group further down this page. The dispersion law pairs national generic-ballot polls taken within three days of each other by different firms, takes out each firm's house effect, and finds that they scatter a little less than their sample sizes imply. The sparse-race error is measured over the few lightly polled races of the four cycles from 2018 to 2024: how far their poll averages landed from the results, a gap that includes an error every poll of a race shares.
| Parameter | In force |
|---|---|
| Realised error sd (2024) | 9.336 pts |
| What sampling error implied | 4.334 pts |
| Ratio | 2.154 x |
fitted on 25 observations · re-estimated walk-forward per cycle
Technical details: run and model identifiers
Fitted file: experiments/sparse_poll_variance.json
Environment drift to election day
How far the national environment still moves between now and November, as a variance growing with the horizon. Fitted on non-overlapping increments of the generic-ballot path so a long series cannot be counted twice.
| Parameter | In force |
|---|---|
| Var(h) = a·h^b — a | 0.203 |
| …b | 0.4966 |
fitted on 371 observations · using only data before 2026 · re-estimated walk-forward per cycle
Technical details: run and model identifiers
Fitted file: experiments/environment_drift_pre2026.json
Poll dispersion
A poll carries noise no sample size reduces, and a firm's house effect does not average away across its own waves. This is why the national environment is weighted by measured variance rather than by n.
| Parameter | In force |
|---|---|
| Variance floor no sample size reduces | 2.368 pts2 |
| Sampling term (var = floor + slope/n) | 7,571 |
| House-effect dispersion across firms | 2.584 pts |
| Design effect | 0.7571 x |
fitted on 13,996 observations · using only data before 2026 · re-estimated walk-forward per cycle
Technical details: run and model identifiers
Fitted file: experiments/poll_dispersion_pre2026.json
District presidential baseline
For a redrawn district, its own previous margin describes an electorate that no longer exists. This relation reads the presidential vote inside the current boundaries as a House margin instead.
| Parameter | In force |
|---|---|
| Presidential margin coefficient | 0.9923 |
| Incumbency | 3.933 pts |
| Intercept | -0.1579 pts |
| Residual sd | 5.427 pts |
| MAE using this relation | 3.174 pts |
| MAE of the old-lines margin, for comparison | 13.11 pts |
fitted on 303 observations · re-estimated walk-forward per cycle
Technical details: run and model identifiers
Fitted file: experiments/baseline_relation.json
Seats-votes relationship
An independent check rather than an input: what the national House vote has historically bought in seats, with a separate level per apportionment map because redistricting moves it by roughly eighteen seats.
| Parameter | In force |
|---|---|
| Seats per point of national margin | 3.085 seats |
| Residual sd | 6.961 seats |
fitted on 11 observations · estimated once on the full record
Technical details: run and model identifiers
Fitted file: experiments/seats_votes.json
Promotion gate thresholds
The minimum improvement worth changing the champion for, per metric, set at about a fifth of the champion's own measured edge over a polling average. Not a round number somebody liked; re-measured when the champion changes.
| Parameter | In force |
|---|---|
| Champion's CRPS edge over the baseline | 0.5499 pts |
| → threshold at 20% of it | 0.11 pts |
fitted on 435 observations · estimated once on the full record
Technical details: run and model identifiers
Fitted file: experiments/crps_edge.json