Senate candidate research

This Senate-only study tests whether new candidate information improves the existing forecast. None of the new feature families passes the incremental acceptance checks. New donation-share, scandal, elected-term, prior-office, prior-performance, state-size and congressional-approval interaction weights remain zero. Existing baseline incumbency and prior-performance terms remain in place.

The comparison uses 130 polled race observations across five chronological test elections: 2016, 2018, 2020, 2022 and 2024. Every fitted correction uses only earlier elections. The saved baseline residual history begins in 2010; reviewed 2006/2008 news is retained as context rather than an extra baseline forecast. The error shown is the equal-cycle average absolute D-minus-R margin error in percentage points; five cycles are five national environments, not 130 independent elections.

The existing-factor benchmark refits residual polling error against polling, reconstructed PVI, incumbency and existing prior-performance controls. It lowers development error from 4.613 to 4.498 points. That benchmark is experimental, not the active forecast. New additions must beat the same benchmark, so the benchmark gain is not credited to scandals or another new variable.

Individual-only money is itemized plus unitemized individual contribution dollars. Each candidate share divides those dollars by all modeled candidates’ individual dollars in that race, including the third Montana candidate. This is a share of dollars, not a share of unique donors and not the small-dollar/unitemized fraction of a campaign’s own receipts. A 1-million versus 99-million split is exactly 1% versus 99%; it does not imply those vote shares.

The quarter comparison is July–September. Reports and amendments must have been publicly received by the historical forecast cutoff. Cumulative committee totals are a separate sensitivity because campaign reporting windows differ. Candidate loans, PAC money, candidate personal contributions and authorized transfers are excluded from individual dollars. Missing reports stay missing; a reported and reconciled zero can be a genuine zero.

Available matched historical financial races are 2016: five; 2018: eleven quarter and ten cumulative; 2022: thirteen. No usable matched expansion for 2020/2024 was obtained before the FEC service returned HTTP 429 with an eight-hour Retry-After. Only 2022 has enough earlier matched races for the configured financial fit. This is insufficient to establish a stable general fundraising coefficient.

In the identical 2018-to-2022 cohort, quarter individual-dollar shares lower error from 3.695 to 3.612 points, while individual-dollar amounts lower it to 3.483. Including the available 2016 training data changes the share result to 3.778, worse than baseline; amounts remain about 3.49. This instability prevents adoption. Comparisons against all-source receipts and combinations with career/scandal factors are downloadable.

Current individual-dollar shares cover all 28 polled races and 57 candidates, including Montana’s third candidate. Seven unpolled races lack matched financial records in this study. Two partial-quarter campaign reports were recovered from saved cutoff-dated filings by reconciling nonoverlapping Q3 periods to every cumulative individual-contribution component; no missing June report was assumed to be zero.

Scandal discovery covers all current modeled candidates and the historical candidate identities, with alias checks. Forty-eight reviewed public-news episodes and forty-three explicit exclusions are retained. Ratings are exploratory descriptions of public news, never findings of criminal guilt. Zero means no qualifying verified episode was accepted in a basic review, not that no controversy or misconduct exists. Six current candidates have insufficiently corroborated leads, kept visible as source gaps.

The scandal window begins with the previous election for that particular seat, including special elections and runoffs. It ends at the historical forecast cutoff. Public news dates and material developments determine eligibility, not dates of alleged conduct. Old conduct newly revealed can count; mere recycling cannot. The seat class and prior election sources are downloadable.

News ratings are 1 minor, 2 moderate and 3 major. The primary ordinal model is compared with separate severity categories, a strict formal/acknowledged-conduct definition, a 180-day recency decay, and the fraction of recent poll weight fielded before precisely dated news. Broader allegations, denials, later dispositions and uncertain dates remain separate. Exact-date/corroboration and leave-one-state-out checks are retained.

Do not assign another person’s scandal to a candidate. A predecessor, spouse, staffer, donor or company is not automatically evidence of the candidate’s personal misconduct. Policy disagreement, ordinary campaign insults, health, victimization and lawful disclosure extensions are excluded. Allegations that were dismissed, uncharged or denied are labeled accordingly. Later reporting cannot supply a fact unavailable to an earlier forecast.

Only one major-news race is present in the five evaluated test cycles under this coding. The major category therefore lacks enough independent support for a calibrated automatic penalty. The strict overall news measure improves error by less than 0.004 points and slightly worsens probability scores; this does not validate a new forecast adjustment.

Wikipedia biographies cover 452 of 507 candidate-race instances and 62 of 71 current candidates. Dated office episodes are clipped before the relevant cutoff. Rejected biography suggestions are never candidate identities. Missing biographies or unparsed offices stay unknown, not no experience. A 172-race historical check agrees with the original incumbency classification.

Senate incumbents are separated into appointed with no prior Senate win, elected in one earlier cycle and elected in multiple earlier cycles. Same-day regular and special elections count as one electoral mandate. Former House, governor, state, local, federal executive and Senate service are distinct. Learned office weights and the suggested state/House/governor hierarchy are both tested, including open seats only and nonincumbents only.

Prior candidate performance is measured against comparable earlier district or statewide partisanship plus contemporaneous national environment. House comparisons require the same district map; a safe district’s raw margin is not statewide personal strength. Governor/Senate context uses reconstructed PVI plus available generic-ballot context, or a same-office/year context when unavailable. Positive candidate-centered values mean personal overperformance for either party. Raw values are capped at 15 points and decay with an eight-year half-life before transfer weights are learned.

Older district-map gaps, ranked-choice and multiparty comparisons limit prior-election coverage. These are missing comparisons rather than zero personal strength. Nine current candidates have eligible prior House comparisons. No new transferred House/governor/Senate performance family clears the incremental gate.

Smaller-state incumbency uses the log of 3 million divided by resident population, bounded at plus/minus two, with a Census vintage known before each election. Congressional job approval uses one complete September Gallup wave per cycle, with public-release dates checked. This is distinct from presidential approval and individual senator approval. State-size improvement is only 0.009 points beyond controls and disappears without Maine/West Virginia; congressional approval adds no material benefit.

Added candidate features use equal-cycle ridge fitting with stronger shrinkage than the shared controls and a total correction cap of two points. Acceptance requires at least three test cycles, a 0.05-point margin-error improvement, probability and competitive-race safeguards, and no severe worst-cycle deterioration. All new families fail the incremental check; financial fits lack enough later test elections.

The elections in this project have already been inspected during development. These comparisons are exploratory, not untouched prospective validation. News can already be reflected in polls, biographies can be incomplete, and contribution coverage can select stronger campaigns. Failure to validate a universal adjustment does not mean scandals, reputation or money never affect elections.

A combined fundamentals score has not shown a statistically reliable improvement beyond the polling forecast.

The score combines incumbency, earlier personal election performance, appointed/elected status, challenger or open-seat experience and dated scandal news. Versions add state partisan lean, the generic ballot, and individual fundraising. Each index is one expected Democrat-minus-Republican margin, learned from earlier elections without current race polling as an input.

In 130 races across 2016–2024, the candidate-plus-lean index, blended using weights chosen from earlier validation elections, changes average error from 4.613 to 4.638 points: slightly worse. It does not validate an added forecast weight.

A fixed 25% blend is more encouraging: 4.613 → 4.348 points, a 0.264-point improvement. It helps in 2020 and 2022 and worsens 2016, 2018 and 2024. The two-sided whole-election sign-flip p-value is 0.625; the exploratory 95% cycle-bootstrap interval for the improvement is -0.192 to 0.869 points.

Most of that fixed-blend gain is available from lean and incumbency alone, which reach 4.381 points. The richer candidate score adds only 0.033 points of improvement over that simpler score (p = 0.625).

Adding the generic ballot has only three later complete-case test cycles (2020/2022/2024; 79 races). The nested blend raises error from 4.726 to 4.871 points. A fixed 25% blend lowers it to 4.282, but the gain is inconsistent (p = 0.750). This subset is not directly comparable with the five-cycle table.

The full fundraising combination can be tested in only thirteen 2022 races after training on the matched 2016/2018 reports. Its fixed 25% blend raises error from 3.695 to 4.225 points. The generic-ballot version is similar. One test election cannot establish a general fundraising coefficient or a statistically significant joint effect.

This is a set of specified combinations, not every possible weighting or interaction. A linear composite packages existing information; its usefulness depends on out-of-election performance. Historical development results are exploratory because these elections have already been studied. Whole-election tests avoid pretending that races in the same national environment are independent, but five cycles remain few, and overlapping training history also limits formal inference. The five-cycle sign-flip test cannot attain a two-sided p below .0625. Bootstrap intervals are descriptive sensitivity checks, not precise confidence guarantees. More financial history and future elections are needed; no new weights have been activated.

I tested removing each factor family and all smaller combinations: 32 versions without generic-ballot or financial coverage restrictions, 64 in the generic-ballot cohort, and 128 in the matched financial cohort. Every subset within a cohort uses the same training and test races.

Across 130 races in 2016–2024, the full five-factor score is the best of all 32 fixed-25%-blend versions, with error 4.348. Removing any one factor makes it worse. Choosing the subset, shrinkage and blend using earlier validation elections gives error 4.642, versus 4.613 for the current forecast. The earlier-selected procedure does not validate a new weight.

In the generic-ballot cohort (79 races, three elections), dropping scandal news is the best fixed-blend subset but helps by only 0.000383 point. In the financial cohort (13 races, one election), the best subset drops fundraising, experience and scandal news; its error is 4.216, still above the current forecast’s 3.695. Financial significance cannot be estimated from one test election.

The best fixed subset is chosen after inspecting the evaluated elections, so it is a retrospective diagnostic. The earlier-selected procedure chooses variables before each target result, using only earlier validation elections. These are development comparisons on elections already studied; no new weights are activated. Correlated remaining inputs can retain related information, so removing a family does not isolate its causal effect. Missing source indicators are removed with their associated family.

Cook, Sabato’s Crystal Ball, and Rothenberg/Inside Elections are tested separately, as a consensus, and with candidate fundamentals. Final Wikipedia predictions for prior cycles are the user-authorized main source. The three ratings are correlated judgments that already use polling, rather than three independent new surveys.

The current model’s historical polling cutoff is sixteen days before each election. Final ratings often appear closer to Election Day and can contain later information. I therefore also retrieved the latest Wikipedia revision before each historical cutoff, and compared final versus cutoff ratings on identical available race/training cohorts. Missing archived ratings remain missing.

Across all 130 polled races in 2016–2024, the earlier-chosen blend of candidate fundamentals, lean and final expert consensus lowers average margin error from 4.613 to 4.286 points. It improves four of five cycles and worsens 2024. Its whole-election p-value versus the active forecast is 0.250. Compared with the same fundamentals model without experts, error falls from 4.638 to 4.286; that paired p-value is 0.062. These are exploratory results with only five national test elections.

The cutoff-date evidence covers 120 races across the same five cycles. Its earlier-chosen expert/fundamentals blend changes error from 4.441 to 4.427 points, an improvement of only 0.014. On exactly that available cohort, using final ratings instead reaches 4.357. This is too little cutoff-valid improvement for the configured acceptance checks; forecast weights remain unchanged.

As a direct polling correction, fitting all three expert ratings separately improves error beyond identical existing-factor controls by 0.039 points using final ratings, or 0.034 using cutoff ratings. Neither reaches the 0.05-point incremental threshold. The controls themselves are experimental; their gain over the active model is not credited to expert ratings.

Toss-up is zero; tilt, lean, likely and safe/solid are signed ordered levels 1, 2, 3 and 4. Positive means a Democratic/main-opposition advantage and negative means Republican. These are feature levels, not assumed vote margins or win probabilities. Consensus requires at least two raters; individual rater coefficients, index weights and blends are fitted only on earlier election outcomes. A toss-up is distinct from a missing rating.

All final tables for 2006–2024 and the current 2026 table are cached with source hashes. Archived tables supply the separate cutoff test; no supported 2006 cutoff table was extracted, and some older tables omit safe races. Final 2008 has no Sabato column, so it is not invented. The 2026 Inside Elections public export agrees with all 35 Wikipedia ratings. Direct Cook and Sabato pages returned HTTP 403 during this check; their Wikipedia entries are retained as the explicitly authorized source without claiming direct verification.

Source joins distinguish regular and special elections. Arizona 2020 is matched as the McSally/Kelly special election while preserving the upstream model’s flag. California 2022 shares a rating only because its source note explicitly covers both simultaneous ballots. Older transposed tables are parsed by full state names, avoiding a Virginia/West Virginia overlap; source typos and shared explicit consensus rows are documented.

The final-rating benchmark mixes a later information horizon with the frozen polling forecast. Its numerical acceptance checks alone do not authorize a present-cutoff forecast change. The cutoff test is the relevant timing check. These development elections have already been inspected; five cycles, selected historical coverage and overlapping training limit statistical claims. Rating agreement does not automatically reduce uncertainty. No new expert coefficient, probability adjustment or simulation change is activated.