VotePredictor

Forecasters

The outfits that turn polls into win probabilities — FiveThirtyEight, Cook, the Economist, Sabato, Silver Bulletin and more — graded on the skill of their final pre-election call (2016–2024), against our own model. Most publish only a win probability — so the skill boards below are win-call skill. But the two that also publish a margin (538, the Economist) get a head-to-head on VotePredictor Elections' home turf.

Difficulty-adjusted skill, with uncertainty

Because forecasters cover different races, raw accuracy isn't comparable — rating only safe states flatters you. A Bayesian two-way model separates race difficulty (from how everyone did on each race) from forecaster skill, so skill is independent of which races they chose. The bar is each forecaster's 90% credible interval — it's wide when they've rated few races — and the board is ranked by a conservative estimate, so a strong record on a handful of races can't top a proven one. Lower (left) is better.

Governor ratings are read from Wikipedia's election-cycle tables. JHK's archive has no governor rows, so Cook, Inside Elections, Sabato and the rest enter here from the final ratings those tables record (dated columns only; a column read more than 45 days before the election is left out), converted with the same rating-to-probability table JHK uses for their House and Senate ratings: Solid 99%, Likely 90%, Lean 75%, Tilt 60%, Toss-up 50%. 538's governor rows are its own archived win probabilities (2018, 2022). Pollsters join on the combined and pollster boards.

#ForecasterRacesBrier90% range (wider = less data)
1FiveThirtyEightinactive6820.071
2VotePredictor Electionsour model7650.074
3JHK Forecasts3900.077
4Race to the WH2760.079
5Fox News2420.079
6Inside Elections5370.081
7Cook Political7470.081
8The Economist5010.082
9Sabato's Crystal Ball7630.083
10Silver Bulletin400.079
11Split Ticket1960.083
12Princeton Election Consortium1310.086
13DDHQ/Decision Desk5300.089
14Elections Daily2930.088
15CNalysis2820.090
16New York Times510.089
17RealClearPolitics3420.115

Margin accuracy vs the pros

The boards above grade win-calls. This grades the margin against the only two forecasters who publish one. We recovered 538's and the Economist's real final predicted margins (538 '18/'20/'22; Economist '20 president) and scored everyone by average margin error on the same races, using our published walk-forward record — every House district, not only the polled ones. VotePredictor Elections beats the Economist outright and runs level with 538 overall (within a tenth of a point): ahead on close presidential states, behind on close Senate and governor races, and narrowly behind on the House, where 2018 decides it. Lower is better.

VotePredictor Elections vs FiveThirtyEight

2018, 2020, 2022 · 1326 races
Margin error (pts) — lower is betterVotePredictor ElectionsFiveThirtyEightpolls
All races · 13264.714.865.73
Competitive (<10 pts) · 2904.254.134.55
Competitive races, by officeVotePredictor ElectionsFiveThirtyEightwinner
President · 162.754.41VotePredictor Elections
Senate · 284.233.93FiveThirtyEight
Governor · 243.242.31FiveThirtyEight
House · 2224.474.33FiveThirtyEight

VotePredictor Elections vs The Economist

2020 · 51 races
Margin error (pts) — lower is betterVotePredictor ElectionsThe Economistpolls
All races · 513.024.375.48
Competitive (<10 pts) · 142.394.454.07

House: every district, against 538

average miss, points · lower is better
CycleAll districts: us538Competitive: us538
2018 · 389 / 895.164.485.003.09
2020 · 402 / 745.205.634.555.83
2022 · 332 / 594.094.523.564.33
All cycles · 1123 / 2224.864.904.474.33

Our numbers are the published House model's own walk-forward record — district lean, incumbency and district polls, each cycle predicted from earlier cycles only — on every district 538 forecast; competitive means the final margin was within 10 points. Cook's rating, which the live forecast also uses for unpolled rated seats, is left out so the comparison is model against model. 538 is ahead overall, by 2018 alone: that year the whole-chamber environment ran four points more Democratic than our model expected. We are ahead in 2022, and on competitive seats in 2020 and 2022.

Win-call skill vs a prior-result baseline

Skill is how much each forecaster's win probability beats a naive prior-result baseline on the same races — so being right about obvious races earns nothing. The Competitive tab (races under 10 points) is the real test. Click a column to sort, a name for the by-office and by-cycle breakdown.

How good the win-probability call was (Brier skill vs a prior-result baseline). All races, 2016–2024.

#NameRacesWin skillBrier
1Silver Bulletin4064%0.045
2FiveThirtyEightinactive68258%0.070
3VotePredictor Elections80356%0.075
4Cook Political70452%0.079
5Inside Elections49252%0.071
6Sabato's Crystal Ball71551%0.081
7Fox News24250%0.082
8The Economist50149%0.088
9Race to the WH27649%0.072
10JHK Forecasts39047%0.080
11DDHQ/Decision Desk53045%0.093
12Princeton Election Consortium13143%0.060
13Poll average (naive)80343%0.096
14Split Ticket19643%0.089
15Elections Daily29343%0.082
16VotePredictor Elections House segmented33240%0.120
17CNalysis28240%0.084
18VotePredictor Elections House fundamentals32135%0.129
19New York Times5127%0.081
20RealClearPolitics29921%0.121

Compare on common races

The fairest test when coverage differs: score everyone on the exact same races — the ones they all forecast. Drag the threshold to drop low-coverage forecasters; the fewer you require, the more races the survivors share. It defaults to the fair board — the six forecasters who covered ≥500 races, on the 510 they all called. Lower the bar and watch the paradox: admit one that rated only a handful and the common set collapses toward zero. Click any name to fine-tune the set.

Two columns, two questions — click either header to sort. Brier rewards confident, well-calibrated probabilities; Called right is the blunt “did they name the winner” rate. They can disagree: a forecaster can have a better Brier yet name fewer winners (its misses are mostly near-tossups it called at ~49%), so it's worth checking both.

≥ 500 races · 7 forecasters · 276 common

Everyone scored on the 276 races all 7 selected cover — identical races, exactly apples-to-apples.

#NameBrier ▾Called right
1VotePredictor Elections0.06192%
2FiveThirtyEight0.07091%
3Cook Political0.07199%
4The Economist0.07189%
5Inside Elections0.07495%
6DDHQ/Decision Desk0.08189%
7Sabato's Crystal Ball0.08189%

Raw Brier (for reference)

The same forecasters on the raw Brier score of their final win probability (lower is better) — not difficulty-adjusted, so safe-race coverage flatters it. The skill board above is the fairer ranking.

392 races where the final margin was within 15 points. Lower Brier is better. Only forecasters that rated every race in the set are ranked; the rest are listed unranked, with VotePredictor's Brier on exactly the races they rated.

#ForecasterRacesBrierCalled rightVP, same races
1FiveThirtyEight3920.11386%0.118
2VotePredictor Elections3920.11883%
3Cook Political3920.12395%0.118
4Sabato's Crystal Ball3920.12983%0.118
—Race to the WH1720.11681%0.097
—Inside Elections2220.12289%0.102
—JHK Forecasts2550.12381%0.109
—Fox News1230.12393%0.110
—Split Ticket1370.12782%0.107
—The Economist3420.12880%0.117
—CNalysis1720.13679%0.097
—DDHQ/Decision Desk3570.13783%0.118
—Elections Daily1580.14180%0.098
—RealClearPolitics1330.19979%0.112