Rating Canadian pollsters: accuracy, the lead and openness
Abstract. Every pollster that has polled Canada is rated out of 100: 50 points for its final polls since 2000 and 20 for its miss on the gap between the leaders, both ranked against Canada's own 82 final polls and shrunk towards a typical one, and 30 for openness, including CRIC accreditation. Léger and Mainstreet score highest (A). The ratings do not weight the average. Seven pollsters' releases, including Léger's regional crosstabs and Nanos's weekly PDF, are now read automatically; the rest are entered by hand.
Summary
We rate every pollster that has polled Canada out of 100, on the same three parts as every country on the site: 50 points for the accuracy of its final polls at general elections since 2000, 20 for its miss on the gap between the two leading parties, and 30 for how much of its working it publishes. Accuracy and the lead are scored by rank against Canada's own 82 final polls since 2000, so a Canadian grade says how a pollster compares with other Canadian pollsters, and a short record is pulled towards a typical Canadian final poll so that one good election cannot buy a top grade. Openness counts seven items, including accreditation by the Canadian Research Insights Council (CRIC).
On 15 October 2026 Léger and Mainstreet Research score highest (77, grade A), followed by Liaison Strategies (71, provisional), Nanos (68), Angus Reid and Pollara (63), Pallas (61) and Research Co. (60), all B. The rating is for readers: it does not weight the polls, which count equally in the average. The paper also describes how the site reads Canadian releases: seven pollsters' releases are now read automatically from their own sites, and the rest are entered by hand.
What the rating is for
The rating answers a narrow question: on the evidence we can check, how far a reader should trust a pollster's published figure, and how easily they can see how it was made. It does not say which pollster is right today, and it does not feed the forecast. The Canadian average is each pollster's latest poll within 21 days of the newest, one per pollster, equally weighted (How we read Canadian polls: the average, the correction and its test). We do not weight by the rating because a weighting would need to be shown to help out of sample, and with nine elections, most pollsters having fewer, that cannot be shown.
The evidence: 82 final polls since 2000
A final poll is each firm's last national poll ending within ten days of polling day (How far Canadian polls have missed, 1988–2025). The ratings use the nine general elections from 2000 to 2025, 82 final polls from 25 pollsters. The three elections before 2000 are in the record but, like every other figure on the site that measures accuracy, the ratings count only elections from 2000 on. Gallup Canada and Insight Canada, whose final polls are all from before 2000, are not rated.
For pollster \(j\) at election \(t\), with \(e_{j,t,p}\) its share for party \(p\) minus the result, the miss is the mean absolute error over the main parties it reports (those with 5% of the vote or more),
and the lead error is its error on the gap between the two parties that actually finished first and second in the vote,
The pair follows the result: the Liberals and the Conservative line in 2000, 2004, 2015 and 2025, the Conservatives and the Liberals in 2006, 2008, 2019 and 2021, and the Conservatives and the NDP in 2011. Across the 82 polls the average miss is 1.67 points per main party and the average absolute lead error 2.99 points. The closest single final poll was Nanos's in 2006 (0.12); the furthest, Innovative Research's in 2011 (3.50).
Scoring by rank
Each country's polls are scored against that country's own final polls, because a miss that is poor in one country can be good in another. Let \(\text{Rank}(x)\) be the share of Canada's 82 final-poll misses that are smaller than \(x\), interpolated between neighbouring values (0 for a miss as small as the best, 1 for one as large as the worst). The points are
where \(\tilde{m}_j\) and \(\tilde{\ell}_j\) are the rated miss and lead error defined below, and \(\text{Rank}_{\ell}\) ranks against the 82 absolute lead errors. The exponent 0.75 shapes the curve: a pollster as good as the median Canadian final poll earns about 30 of the 50 accuracy points, and one in the best tenth about 46.
The Canadian yardstick, from the current record:
| Best tenth | Median | Worst tenth | |
|---|---|---|---|
| Miss per main party | 0.88 | 1.72 | 2.51 |
| Absolute lead error | 0.44 | 2.79 | 6.16 |
Table: Quantiles of the 82 Canadian final polls since 2000, points.
Scoring by rank was introduced on 13 October 2026. The earlier scale used fixed end points from the British ratings and pulled short records towards a poor final poll (the worst quarter), and was judged too punitive: the highest grade was a B and most pollsters had a D. Ranking within the country gives every country the same spread of grades and makes them a comparison within the country, not across countries.
Shrinkage towards a typical poll
A pollster's own average is pulled towards the median Canadian final poll as if it had two more elections at that level (Efron and Morris, 1975):
with \(n_j\) the number of its final polls since 2000, \(\bar{m}_j\) and \(\bar{\ell}_j\) its averages of \(m_{j,t}\) and \(|\ell_{j,t}|\), and \(m^{\text{med}} = 1.72\), \(\ell^{\text{med}} = 2.79\). With one election the prior carries two thirds of the weight; with eight, a fifth. A pollster with fewer than two final polls is marked provisional.
Liaison Strategies shows the effect. Its one final poll, in 2025, missed by 0.52 points per main party, among the best in the record. Shrunk, its rated miss is \((0.52 + 2 \times 1.72)/3 = 1.32\), worth 38 accuracy points rather than the near-full marks its own poll alone would earn.
Openness
The 30 openness points reward publishing enough that a figure can be checked. The Canadian items:
| Item | Points | Why it counts |
|---|---|---|
| Raw or unweighted figures | 8 | Shows what respondents said before the pollster's own adjustments |
| Full tables published | 8 | Regions, sub-groups and weighting can be checked |
| Method note | 4 | How the sample was drawn and weighted |
| Sample size | 3 | The basics for judging a poll |
| Undecided share | 3 | Shows how firm the figures are |
| Fieldwork dates | 2 | When the poll was taken |
| Accredited by the industry body (CRIC) | 2 | Signs up to its standards |
Table: Canadian openness items.
Each item is Yes, Partial or No, worth the full points, a third, or nothing:
where \(C_j\) is the set of items checked for that pollster. An item not yet checked counts neither for nor against: the score is taken over the checked items and scaled to 30. A pollster whose publications have not been checked at all (every firm that has not polled since the 2025 election) is rated on its record alone, accuracy and lead scaled from 70 to 100, and marked "record only".
The review was made from each pollster's own releases, with accreditation from the CRIC member directory. In brief:
- Léger publishes a PDF for each federal poll with a crosstab by region (Atlantic, Quebec, Ontario, Manitoba and Saskatchewan, Alberta, British Columbia) and by demographic group, weighted and unweighted counts per column, a total-sample column including the undecided beside the decided-voter column, and the question wording. Full marks.
- Pollara publishes a crosstab PDF with all respondents (undecided and would-not-vote shown) and decided and leaning tables by region with weighted and unweighted counts. Full marks.
- Innovative Research publishes PDF reports with regional tables, weighted and unweighted counts and the undecided share; it is not in the CRIC directory, and several of its older PDFs are no longer online.
- Mainstreet publishes a public PDF that sometimes has regional tables with unweighted counts and the undecided share; its full crosstabs are for subscribers.
- Angus Reid publishes a PDF release with regional vote charts and links to detailed tables, but its regional shares are in stacked charts rather than text.
- Nanos publishes a weekly PDF with the national ballot, trend and method; its regional and demographic figures are on a subscriber portal.
- Abacus, Research Co., Ipsos, Pallas and EKOS publish national shares with a method note, sample and dates, and give regional figures in prose, as images or not at all.
Liaison Strategies and Pallas Data are associate members of CRIC rather than accredited agencies, so their accreditation item is left unscored.
A worked example: Nanos
Nanos Research has a final poll at eight of the nine elections since 2000. Its misses per main party were 1.95 (2004), 0.12 (2006), 1.80 (2008), 1.36 (2011), 0.49 (2015), 1.78 (2019), 0.81 (2021) and 1.00 (2025), an average of 1.16; its lead errors were −3.1, +0.3, −3.9, −3.5, +1.1, −0.4, −2.3 and +0.3, an average absolute error of 1.85. Then
which rank it among the better Canadian final polls: 38.9 accuracy points and 14.0 lead points by equation (3). On openness it scores Yes for the method note, sample size, fieldwork dates and CRIC accreditation, Partial for full tables and the undecided share, and No for raw figures: 14.7 of 30. The total, 67.6, rounds to 68, a B.
The current table
| Pollster | Score | Grade | Accuracy /50 | Lead /20 | Openness /30 | Elections since 2000 | Polls since 2025 |
|---|---|---|---|---|---|---|---|
| Léger | 77 | A | 35 | 12 | 30 | 8 | 14 |
| Mainstreet Research | 77 | A | 39 | 13 | 25 | 4 | 7 |
| Liaison Strategies | 71 | B (provisional) | 38 | 14 | 19 | 1 | 55 |
| Nanos Research | 68 | B | 39 | 14 | 15 | 8 | 76 |
| Angus Reid Institute | 63 | B | 31 | 12 | 21 | 5 | 10 |
| Pollara | 63 | B (provisional) | 19 | 14 | 30 | 1 | 1 |
| Pallas Data | 61 | B (provisional) | 35 | 13 | 14 | 1 | 5 |
| Research Co. | 60 | B | 35 | 13 | 12 | 3 | 5 |
| Ipsos Canada | 57 | C | 31 | 11 | 15 | 9 | 5 |
| Abacus Data | 56 | C | 31 | 13 | 12 | 4 | 37 |
| Innovative Research | 47 | C | 10 | 9 | 28 | 5 | 7 |
| EKOS Research | 38 | D | 18 | 6 | 14 | 9 | 11 |
Table: Pollsters polling Canada now, 15 October 2026. Components rounded; the score is rounded from the unrounded sum.
Thirteen pollsters with a final-poll record have published no poll in the last year and are listed as past pollsters, rated on their record alone: Cardinal Research (67), Harris-Decima (62), Environics, Forum Research and Counsel Public Affairs (58), Zogby and Asking Canadians (55), Campaign Research and YouGov (53), The Strategic Counsel (50), DART / Maru (49), COMPAS (42) and Focaldata (41). Two pollsters are polling now with no final poll yet, Spark Advocacy and Kolosowski Strategies; they are shown without a grade.
Three things stand out. Mainstreet has the best record of any pollster with more than one election (1.06 points per main party at four elections, including 0.28 in 2025). Léger reaches an A with a weaker record (1.41 at eight elections) because it publishes everything we look for. And EKOS, with the longest record (nine elections) and a miss of 2.07 per main party, has among the largest lead errors in the record (5.2 points on average; only COMPAS, at three elections, is larger), which is what holds its grade down.
Leans, past and present
Beside the rating the site shows how each pollster differs from the others. Its past lean on a party is the mean error of its final polls; its lean now is its latest poll in the 21-day window minus the average,
A lean now is not an error: no one knows yet who is right. But a pollster whose lean now matches its past errors gives a reader something to weigh.
The past leans repeat the industry's pattern. Of the 25 pollsters with a final poll since 2000, 23 had the NDP too high on average and 21 had the Conservatives too low. Among the four pollsters in today's average, Nanos's past lean is NDP +1.8 and Conservative −1.5; today it has the Liberals 2.5 points above the average and the NDP 1.4 below. Abacus's past lean is Conservative −2.4; today it has the Conservatives 4.4 points above the average. Liaison and EKOS have the NDP 2.5 and 2.8 points above the average today.
Reading releases automatically
The poll check reads the newest release from each pollster whose site allows it. It follows robots.txt, waits at least 1.5 seconds between requests, and uses plain requests only, with no logins, cookies or anything that gets round a block. A release it cannot read in full (no dates, a missing party) is not guessed at: the check fails with the reason, and the poll waits on the admin page for manual entry.
- Abacus Data: the RSS feed of its Canadian politics posts; the vote sentence ("46% of committed voters would vote Liberal, 34% Conservative ...") and the fieldwork dates.
- Research Co.: its RSS feed; each federal post is opened and read from its text. Provincial polls are skipped.
- Pallas Data: its RSS feed; the figures from the post's own sentence, or its title, and the dates and sample from the text.
- Léger: its "In the news" feed links each federal report's PDF; the figures come from the crosstab table, the decided-voter column for the national poll and the six regional columns for the regional readings.
- EKOS: its RSS feed; each federal post links a data-tables PDF, read for the decided vote intention including leaners.
- Liaison Strategies: the RSS feed of its press site; each weekly tracker post's text and sample. Its regional figures, quoted irregularly in the commentary, are not read.
- Nanos Research: its political-tracking feed links the weekly PDF, read for the "Ballot" paragraph and the dates.
Angus Reid, Ipsos, Innovative Research, Mainstreet, Pollara and Spark are entered by hand. Of the 238 national polls since the 2025 election, 164 come from the pollster's own release (grade A) and 74 from Wikipedia's table (grade C), 42 of them Liaison's (The Canada data: sources, grades and checks).
Limits
Nine elections. No pollster has more than nine final polls since 2000, most have far fewer, and the records cover different elections: those that polled in 2004 and 2006 carry the post-merger misses, those that started in 2019 do not.
Second-hand final polls. Every final poll comes from Wikipedia's tables. A transcription error would pass straight into a rating.
Openness is what we could see. An unchecked item is left out rather than counted as No, but a pollster's openness can still reflect what was on its site on the day of the review.
Methods change. Firms change mode, owners and names. A record built on 2000 to 2025 may describe a pollster that no longer works that way.
Grades are within Canada. A Canadian B means better than most Canadian final polls, not as good as a British or German B (Rating the British pollsters: accuracy, the Labour–Conservative lead and openness, Rating German pollsters: accuracy, the lead and openness).
The first real test of whether higher-rated Canadian pollsters are more accurate will be the next general election.
References
Efron, B. and Morris, C. (1975). Data analysis using Stein's estimator and its generalizations. Journal of the American Statistical Association, 70(350), 311–319.
Canadian Research Insights Council, member directory.
Figures in this paper are as of 15 October 2026; the pages update with each poll.
How to cite
Polling Forecast (2026). Rating Canadian pollsters: accuracy, the lead and openness. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/canada-pollsters. Live figures change as new polls arrive; the date above is when the text was last revised.



