American Research Group
Graded against the actual result across 205 races (from 339 polls, through 2022).
Head-to-head vs the VotePredictor model
The fair, apples-to-apples test: on the 145 races American Research Group actually polled, how its final poll's margin compared to what the VotePredictor model predicted for those same races.
| Model | Avg miss (pts) | Called right |
|---|---|---|
| American Research Group | 6.28 | 86% |
| VotePredictor | 3.22 | 94% |
VotePredictor aggregates all the pollsters, so it's expected to beat any single one on margin — that's the value of averaging. The honest comparison among forecasters is on the combined board.
Every race (203)
Each race American Research Group polled, scored on its final poll — the call right before the vote — against the actual Dem−Rep result. Click a race for its full detail.
Accuracy by time to election
Lower is better. Time to election runs right (election week) to left (~2 months out).
By the numbers
| Time to election | Polls | Avg miss | vs field | Called right |
|---|---|---|---|---|
| ≤1 wk | 35 | 4.16 | -0.12 | 66% |
| 1–3 wk | 48 | 4.27 | -0.80 | 65% |
| 3–6 wk | 76 | 6.45 | +0.71 | 68% |
| 6–9 wk | 180 | 6.71 | +0.56 | 87% |
vs field is this pollster's average miss minus all pollsters' at the same lead time — green beats the field, redtrails it. Our historical polls reach ~2 months out; earlier polling isn't in the record.
Track record by cycle — getting better?
| Year | Polls | Avg miss | Lean (house effect) |
|---|---|---|---|
| 2000 | 70 | 6.0 | D+3.5 |
| 2002 | 22 | 5.3 | R+0.5 |
| 2004 | 77 | 4.4 | D+1.4 |
| 2006 | 10 | 10.7 | R+7.2 |
| 2008 | 92 | 7.3 | R+4.9 |
| 2010 | 8 | 5.2 | D+3.1 |
| 2012 | 30 | 4.9 | R+4.5 |
| 2014 | 8 | 4.9 | D+2.8 |
| 2016 | 8 | 2.6 | D+0.0 |
| 2018 | 3 | 9.7 | D+9.7 |
| 2020 | 7 | 12.5 | D+10.9 |
Do we credit a pollster for fixing its bias? Each cycle, the model re-estimates every pollster's lean from all its earlier polls (walk-forward) and subtracts it before using the poll. We tested weighting recent cycles more — it doesn't help: a pollster's lean in one cycle barely predicts the next (correlation 0.28), so the swings above are mostly noise, and averaging over more history beats chasing the latest cycle. The all-time estimate we use came out within ~0.5% of the best option.