Methodology & data
Model assumptions, historical tests, included variables, and research records for the 2026 forecasts.
Senate model
Forecast data through 2026-10-19
How the forecast works
How each race is estimated
Recent general-election polls are combined into a weighted average. Newer polls and larger samples receive more influence, subject to limits on sample size and each polling firm's total weight. Repeated questions from the same survey are grouped together so they do not count as separate, independent polls.
The polling average is adjusted for patterns in past election polling misses. The Senate model uses a shared historical correction; separate pollster-specific corrections are retained as comparison cases. State partisan lean, incumbency, and available earlier candidate results also inform the baseline. When polls are sparse or old, that baseline receives more influence and the predicted range is wider.
Partisan lean and candidate history
State partisan lean is reconstructed from the two previous presidential elections, relative to the national result, with 75% weight on the more recent election and 25% on the earlier one. The same definition is used across historical elections.
Incumbency and available electoral records help distinguish candidates from the underlying partisan lean of a race. Missing history is flagged rather than treated as zero experience or zero ability. Older elections receive less weight, and presidential-election years receive half the weight of midterms when estimating the historical polling correction.
How the simulations work
The model simulates 100,000 possible elections. The default projection allows races to move together through shared national, state, and regional shifts across the House and Senate, while retaining uncertainty specific to each race. The standard option provides a comparison with separate national polling errors for the two chambers.
The 50%, 70%, and 90% ranges contain those shares of simulated margins. They account for polling errors and possible movement before Election Day; they are not the margin of error of a single poll. Changing the projection updates the map, race estimates, uncertainty ranges, and chamber outcomes together. The limited historical record has not established that the shared-error default is more accurate.
Historical candidate performance
A provisional current-election adjustment compares candidates' past results with the partisan lean and political environment of those earlier races. Prior office, incumbency, recency, and comparable district boundaries are taken into account. The resulting estimate replaces overlapping history adjustments so the same evidence is not counted twice.
Changes are capped at three margin points. Races without a supported, comparable history retain their existing treatment. This adjustment has not established a reliable historical accuracy improvement; the full explanation and race-by-race effects are available in Research & data.
Subjective scandal adjustments
Additional penalties are applied to eight Senate candidates based on judgments about the electoral effects of scandals or negative coverage. The starting values allow for some of an effect already appearing in polls. These are subjective assumptions, not findings of wrongdoing or a statistically established formula.
The controls can change individual penalties or turn them off. Changes apply to the current page; reloading restores the published defaults.
Fundraising
A provisional Senate adjustment compares total fundraising receipts reported through September 30 using the same reporting period for each candidate. It applies to races with eligible polling and a comparable financial pair, and is capped at three margin points. Missing financial coverage does not create an assumed zero fundraising advantage.
The historical improvement is small and not statistically reliable. Existing uncertainty ranges are retained, so they remain conditional on this provisional adjustment. The financial comparison, coverage, and race-by-race effects are available in Research & data.
How seats are counted
A House majority requires 218 seats. Senate control is counted as at least 51 Democratic-caucus seats or 50 Republican seats, assuming the Republican vice president breaks a tie. All elected Senate independents are assumed to caucus with Democrats. Their seats are included in Democratic-caucus totals while candidates retain their Independent labels. This is a modeling assumption, not a statement of candidate commitments.
Georgia's Senate simulations allow for a runoff and additional movement between rounds. Alaska and Maine use final-round polling where available, with extra uncertainty for ranked-choice outcomes. Unopposed and same-party House contests count toward party control; individual winners are shown only when separately modeled.
What the historical tests show
Each historical forecast is fitted using earlier elections and compared with the eventual result. Tests assess margin error, win-probability error, and how often actual results fall within the predicted ranges. An additional variable can be left out even if one score improves, because it may worsen other checks or produce too small a gain to justify its use.
Historical candidate performance and Senate total fundraising receipts are included provisionally while testing continues. The other additions listed as tested but not included receive no extra forecast weight. This does not remove the existing baseline terms for partisan lean, incumbency, or candidate history.
Limits of the evidence
Only a limited number of independent election cycles are available, and those elections have already informed model development. Historical tests help compare methods but do not guarantee future accuracy. Some older polling publication dates, candidate records, and district comparisons remain incomplete.
Grouping repeated survey questions reduces duplicate counting, but public polling totals do not identify people who responded to polls from different firms. Exact overlap between those respondents cannot be removed from the available data.
Model ingredients and test results
Included model elements are listed first, followed by current-election adjustments and variables tested but not included.
| Variable | Current use | Notes |
|---|---|---|
| Included elements | ||
| Recent race polling | Included in Model | Eligible general-election polls provide the starting estimate. Polls are matched to the candidates and election being forecast. |
| Poll recency and sample size | Included in Model | Newer polls and larger samples receive more weight. Sample-size and polling-firm weight limits prevent a large survey or frequent pollster from dominating the average. |
| Repeated survey questions | Included in Model | Questions from the same survey are grouped together rather than counted as independent polls. Likely-voter samples are preferred to registered-voter and adult samples when comparable choices are available. |
| Historical polling misses | Included in Model | A shared correction accounts for the direction of polling errors in past elections. Separate pollster-specific corrections remain comparison cases in the Senate model. |
| Historical election weighting | Included in Model | Older elections receive less weight. Presidential-election years receive half the weight of midterms when estimating the historical polling correction. |
| State partisan lean | Included in Model | State presidential results are compared with the national result, using a consistent 75% / 25% weighting of the two previous presidential elections. |
| Incumbency and open seats | Included in Model | The baseline distinguishes incumbents from challengers and open seats. This existing effect is separate from the additional career-history variables tested below. |
| National political environment | Included in Model | Other Senate races help estimate the national environment. Each race is excluded from the calculation of its own national-environment adjustment. |
| Polling and baseline blend | Included in Model | The forecast draws more on the structural baseline when polling is sparse or old. The Senate baseline contribution is limited to half the blend and a three-point margin adjustment. |
| Baseline candidate electoral record | Included in Model | Available prior Senate or governor performance and recorded election wins inform the candidate baseline. Missing records are flagged rather than treated as proof of inexperience. |
| Seven unpolled seats | Included; structural probabilities | Structural estimates replace guaranteed party holds in the seven Senate races without eligible polls. Their uncertainty ranges are broad. |
| Probability calibration | Included in Model | Historical forecast errors determine the width of the predicted ranges and the win probabilities. Senate ranges also allow for uncertainty in fitted model adjustments. |
| State / regional / joint errors | Default Projection | Races in the same state or region can move together, as can House and Senate results. This is the default projection; the available election history is too limited to establish whether it improves accuracy. |
| Independent and ranked-choice races | Included in Model | Independent candidates retain their party labels; all elected Senate independents are assumed to caucus with Democrats for control and seat totals. Final-round polls are preferred where available, and extra uncertainty allows for independent comparisons and ranked-choice transfers. |
| Georgia runoff | Included in Model | Senate simulations allow for a Georgia runoff and additional movement between the November election and the final round. |
| Current-election adjustments | ||
| Subjective Scandal Adjustments | Manual current-election adjustment | Eight Senate candidate penalties reflect judgments about possible electoral effects. The controls above can change or disable them; these are not statistically established effects or findings of wrongdoing. |
| Total fundraising receipts | Provisional current-election adjustment | Candidate fundraising totals through September 30 are compared using the same reporting period. Eligible Senate margin adjustments are capped at three points; the small historical improvement is not statistically reliable. |
| Historical candidate performance | Provisional current-election adjustment | Past results are adjusted for partisan lean, the national environment, and the circumstances of the earlier race. The new estimate replaces the existing history adjustment so the same performance is not counted twice. Changes are capped at three margin points; race-by-race effects are in Research & data. |
| Tested but not included | ||
| Generic Ballot | Tested; Not Included In Model | Adding national generic-ballot level and movement increased average prediction error in the Senate tests. No additional Senate adjustment is applied. |
| Recent movement in race polls | Tested; Not Included In Model | An extra adjustment for recent polling movement increased average prediction error in the Senate tests. |
| Demographic errors | Tested; Not Included In Model | Map-matched education and urbanization were tested, but did not reliably improve accuracy in the available elections. |
| Individual-donation shares | Tested; Not Included In Model | A candidate's share of all individual-contribution dollars raised in the race did not consistently predict results better than the dollar amount raised. Too few comparable elections were available to establish a reliable benefit. |
| Elected terms and prior office | Tested; Not Included In Model | Extra distinctions between appointed, once-elected, multiple-election, and open-seat candidates did not reliably improve accuracy beyond the existing baseline. |
| Earlier election performance | Tested; Not Included In Model | Additional House, governor, and Senate performance measures did not reliably improve accuracy beyond the existing incumbency and prior-performance terms. |
| State size and Congress approval | Tested; Not Included In Model | No material additional improvement in accuracy was found. |
| Combined candidate fundamentals | Tested; Not Included In Model | Combining the additional candidate variables into a single score did not reliably increase accuracy beyond the existing baseline. |
| Expert race ratings | Tested; Not Included In Model | Final pre-election ratings looked promising, but they can contain information published after the forecast date. Using only ratings available at that date did not improve accuracy enough to justify adding them. |
Historical track record
Each forecast is fitted using earlier elections. Lower margin and probability errors are better. Coverage is the share of actual results within each interval. The limited number of election cycles constrains what these checks establish.
| Election | Races | Margin error (pp) | Probability error | 50% coverage | 70% coverage | 90% coverage |
|---|---|---|---|---|---|---|
| 2016 | 25 | 5.38 | 0.0766 | 48% | 76% | 96% |
| 2018 | 26 | 3.42 | 0.0673 | 69% | 96% | 100% |
| 2020 | 26 | 6.04 | 0.0663 | 31% | 58% | 85% |
| 2022 | 26 | 4.82 | 0.0317 | 46% | 69% | 81% |
| 2024 | 27 | 3.28 | 0.0373 | 56% | 85% | 100% |
Research and source data
Polling cutoff: 2026-10-19 · Latest eligible poll: 2026-10-19.
Forecast data
Published shared-error race projectionsCSV ↗Standard-error race projectionsCSV ↗Methods
Senate caucus assumption and checksJSON ↗Senate methodologyMD ↗Feature and interval test resultsCSV ↗Feature decisions and validationJSON ↗Shared-error assumptionsJSON ↗Selected forecast and rationaleJSON ↗Research
October 20 model reviewHTML ↗Historical candidate performanceHTML ↗Model research overviewMD ↗Total fundraising receiptsHTML ↗Manual candidate adjustmentsHTML ↗Source records
Candidate verification coverageCSV ↗Map geography and sourcesMD ↗Polling refresh recordMD ↗September financial update and checksMD ↗House model
Forecast data through 2026-10-19
How the forecast works
How each race is estimated
Recent general-election polls are combined into a weighted average. Newer polls and larger samples receive more influence, subject to limits on sample size and each polling firm's total weight. Repeated questions from the same survey are grouped together so they do not count as separate, independent polls.
The polling average is adjusted using historical shared polling errors and limited pollster-specific corrections. A structural baseline combines district partisan lean, national generic-ballot polling, and incumbency. District polling receives more influence when it is recent and supported by several firms; unpolled districts rely on the baseline and receive wider predicted ranges.
Partisan lean and candidate history
District partisan lean compares previous presidential results with the national result on the same district boundaries. A district number alone is not sufficient to link results across different maps. Where supported, additional presidential history and prior candidate results strengthen the estimate for unpolled districts.
Incumbency and available electoral records help distinguish candidates from the underlying partisan lean of a race. Missing history is flagged rather than treated as zero experience or zero ability. Older elections receive less weight, and presidential-election years receive half the weight of midterms when estimating the historical polling correction.
How the simulations work
The model simulates 100,000 possible elections. The default projection allows races to move together through shared national, state, and regional shifts across the House and Senate, while retaining uncertainty specific to each race. The standard option provides a comparison with separate national polling errors for the two chambers.
The 50%, 70%, and 90% ranges contain those shares of simulated margins. They account for polling errors and possible movement before Election Day; they are not the margin of error of a single poll. Changing the projection updates the map, race estimates, uncertainty ranges, and chamber outcomes together. The limited historical record has not established that the shared-error default is more accurate.
Historical candidate performance
A provisional current-election adjustment compares candidates' past results with the partisan lean and political environment of those earlier races. Prior office, incumbency, recency, and comparable district boundaries are taken into account. The resulting estimate replaces overlapping history adjustments so the same evidence is not counted twice.
Changes are capped at three margin points. Races without a supported, comparable history retain their existing treatment. This adjustment has not established a reliable historical accuracy improvement; the full explanation and race-by-race effects are available in Research & data.
Subjective scandal adjustments
The scandal controls apply to Senate candidates. They can change the joint House-and-Senate outcomes, while individual House race estimates remain unchanged. The penalties are subjective assumptions about possible electoral effects, not findings of wrongdoing or statistically established effects.
The controls can change individual penalties or turn them off. Changes apply to the current page; reloading restores the published defaults.
Fundraising
The joint outcomes include the provisional Senate fundraising adjustment based on comparable receipts through September 30. Individual House race estimates receive no additional fundraising adjustment.
The historical improvement is small and not statistically reliable. Existing uncertainty ranges are retained, so they remain conditional on this provisional adjustment. The financial comparison, coverage, and race-by-race effects are available in Research & data.
How seats are counted
A House majority requires 218 seats. Senate control is counted as at least 51 Democratic-caucus seats or 50 Republican seats, assuming the Republican vice president breaks a tie. All elected Senate independents are assumed to caucus with Democrats. Their seats are included in Democratic-caucus totals while candidates retain their Independent labels. This is a modeling assumption, not a statement of candidate commitments.
Georgia's Senate simulations allow for a runoff and additional movement between rounds. Alaska and Maine use final-round polling where available, with extra uncertainty for ranked-choice outcomes. Unopposed and same-party House contests count toward party control; individual winners are shown only when separately modeled.
What the historical tests show
Each historical forecast is fitted using earlier elections and compared with the eventual result. Tests assess margin error, win-probability error, and how often actual results fall within the predicted ranges. An additional variable can be left out even if one score improves, because it may worsen other checks or produce too small a gain to justify its use.
Historical candidate performance and Senate total fundraising receipts are included provisionally while testing continues. The other additions listed as tested but not included receive no extra forecast weight. This does not remove the existing baseline terms for partisan lean, incumbency, or candidate history.
Limits of the evidence
Only a limited number of independent election cycles are available, and those elections have already informed model development. Historical tests help compare methods but do not guarantee future accuracy. Some older polling publication dates, candidate records, and district comparisons remain incomplete.
Grouping repeated survey questions reduces duplicate counting, but public polling totals do not identify people who responded to polls from different firms. Exact overlap between those respondents cannot be removed from the available data.
Model ingredients and test results
Included model elements are listed first, followed by current-election adjustments and variables tested but not included.
| Variable | Current use | Notes |
|---|---|---|
| Included elements | ||
| Recent race polling | Included in Model | Eligible general-election polls provide the starting estimate. Polls are matched to the candidates and election being forecast. |
| Poll recency and sample size | Included in Model | Newer polls and larger samples receive more weight. Sample-size and polling-firm weight limits prevent a large survey or frequent pollster from dominating the average. |
| Repeated survey questions | Included in Model | Questions from the same survey are grouped together rather than counted as independent polls. Likely-voter samples are preferred to registered-voter and adult samples when comparable choices are available. |
| Historical polling misses | Included in Model | Past polling errors inform a shared correction and limited pollster-specific adjustments. Estimates based on less historical evidence are pulled closer to zero. |
| Historical election weighting | Included in Model | Older elections receive less weight. Presidential-election years receive half the weight of midterms when estimating the historical polling correction. |
| District partisan lean | Included in Model | Previous presidential results on the same district map are compared with the national result. Results from incompatible district boundaries are not substituted. |
| Incumbency and open seats | Included in Model | The baseline distinguishes incumbents from challengers and open seats. This existing effect is separate from the additional career-history variables tested below. |
| National political environment | Included in Model | National generic-ballot polling supplies the national environment used with district lean and incumbency. |
| Polling and baseline blend | Included in Model | District polling is blended with a structural estimate. Races with fewer polls or older polling rely more on the baseline; unpolled districts receive wider uncertainty ranges. |
| Generic Ballot | Included in Model | National Democratic-versus-Republican preference is part of the House baseline. This differs from the extra Senate adjustment that was tested and left out. |
| Richer baseline in unpolled seats | Included in Model | Additional map-compatible presidential history and available prior candidate performance are used in unpolled districts. This version passed the model's deployment checks; the broader version for all districts did not. |
| Probability calibration | Included in Model | Historical forecast errors determine the width of the predicted ranges and the win probabilities. The selected interval settings are based on earlier-election tests. |
| State / regional / joint errors | Default Projection | Races in the same state or region can move together, as can House and Senate results. This is the default projection; the available election history is too limited to establish whether it improves accuracy. |
| Independent and ranked-choice races | Included in Model | House independents are tracked separately in seat totals. Final-round polls are preferred where available, and extra uncertainty allows for independent comparisons and ranked-choice transfers. |
| Unopposed and same-party contests | Included in Model | These contests count toward party seat totals. Individual winners and candidate margins are shown only when separately modeled. |
| Current-election adjustments | ||
| Subjective Scandal Adjustments | Manual current-election adjustment | The manual penalties apply to Senate candidates and affect joint House-and-Senate outcomes. They do not change individual House race estimates. |
| Senate total receipts | Provisional current-election adjustment | The provisional Senate fundraising adjustment affects the joint outcomes. House race estimates receive no additional fundraising adjustment. |
| Historical candidate performance | Provisional current-election adjustment | Past results are adjusted for partisan lean, the national environment, and the circumstances of the earlier race. The new estimate replaces the existing history adjustment so the same performance is not counted twice. Changes are capped at three margin points; race-by-race effects are in Research & data. |
| Tested but not included | ||
| Richer candidate / district baseline in all seats | Tested; Not Included In Model | Average margin error improved slightly, but probability and close-race checks did not support using this broader baseline in polled districts. The unpolled-only version is included. |
| Recent movement in race polls | Tested; Not Included In Model | An extra adjustment for recent polling movement did not improve average prediction error. |
| Demographic errors | Tested; Not Included In Model | Map-matched education and urbanization were tested, but did not reliably improve accuracy in the available elections. |
Historical track record
Each forecast is fitted using earlier elections. Lower margin and probability errors are better. Coverage is the share of actual results within each interval. The limited number of election cycles constrains what these checks establish.
| Election | Races | Margin error (pp) | Probability error | 50% coverage | 70% coverage | 90% coverage |
|---|---|---|---|---|---|---|
| 2018 | 386 | 5.48 | 0.0409 | 64% | 84% | 98% |
| 2020 | 399 | 6.19 | 0.0313 | 50% | 73% | 95% |
| 2022 | 393 | 4.21 | 0.0266 | 59% | 80% | 97% |
| 2024 | 391 | 3.21 | 0.0148 | 72% | 86% | 98% |
Research and source data
Polling cutoff: 2026-10-19 · Latest eligible poll: 2026-10-15.