Polling

Which pollsters are actually right?

22,789 polls, each matched to the result it was trying to predict. Two different things get measured here, and people constantly conflate them.

Error

5.55

How far off the average poll is, in points, regardless of direction. A high-error pollster is imprecise.

Bias

D+0.91

Which direction they miss, on average. This is the more useful number: error is noise, bias persists. Positive means the polls overstated Democrats.

Polling error by cycle

The famous misses show up exactly where you'd expect — and note that they run in both directions.

Cycle	Average error	Direction of miss	Polls
1998	8.19	R+4.79	447
2000	5.36	R+0.84	1,350
2002	5.94	D+2.72	759
2004	4.70	D+1.10	1,511
2006	5.80	R+1.16	1,240
2008	5.72	R+1.79	2,518
2010	6.18	R+0.20	1,551
2012	5.07	R+2.54	2,136
2014	5.91	D+3.34	1,217
2016	6.43	D+3.68	2,483
2018	4.98	R+0.30	1,225
2020	5.92	D+4.71	2,263
2022	4.97	D+0.87	1,143
2024	4.24	D+2.78	2,323

Most accurate

Lowest average error, 60+ polls, non-partisan.

Lean Democratic

House effects that overstate Democrats.

Lean Republican

House effects that overstate Republicans.

TI / Roper marks the 90 of 664 pollsters flagged in FiveThirtyEight's aapor_roper field, which 538 defines as a member of the AAPOR Transparency Initiative or a contributor of polls to the Roper Center archive. It is not AAPOR membership, and it is 538's last recorded status for the firm — the field is stamped identically on every one of that pollster's polls, however old, so it says nothing about any particular poll. Our 2024 polls carry no such field, so nothing after 2023 is reflected here.

Does polling closer to the election help?

Final week	5.14	D+0.39
1–2 weeks out	5.52	D+0.89
2–4 weeks out	5.60	D+1.07
1–3 months out	5.83	D+1.04
3+ months out	4.99	D+0.98

A little. Error falls as election day approaches, but not as much as you might expect.

Where polls miss most

HI	10.38	R+7.49
ND	10.22	D+8.02
ID	9.80	D+5.71
SC	9.26	D+1.56
NE	9.05	D+6.47
WV	8.56	D+3.77

Mostly small states that get polled rarely, where a handful of bad surveys move the average.

Source & caveats

1998–2023 — 20,466 polls — comes from FiveThirtyEight's raw_polls.csv, which pairs each poll with the actual result, used under CC BY 4.0. Nothing here is endorsed by FiveThirtyEight, and the analysis on this page is ours. 2024 is our own reconstruction. FiveThirtyEight was shut down in March 2025 before that file was updated, so we recovered 538's raw 2024 poll files from the Wayback Machine and matched them to results we compute ourselves — presidential margins aggregated from MEDSL's county returns, and Senate margins from MEDSL's statewide file. Those county returns are not independently reconciled against state canvasses. What survives in the archive is a snapshot, not the complete file: the largest surviving capture of the presidential polls runs through Nov 4, the day before the election, and later snapshots were emptied. That reconstruction yields 2,323 matched polls, covering presidential and Senate general elections only — 2024 House and governor polls are not yet matched, and Maine and Nebraska congressional-district polls are excluded, having no statewide result to score against. Bias is only computed where a poll is a clean Democrat-versus-Republican matchup — otherwise the sign is meaningless and the poll is excluded from bias but kept for error. Pollsters with fewer than 60 polls are left out of the rankings: with a small sample, topping an accuracy table is mostly luck.