How far Canadian polls have missed, 1988–2025
Abstract. 92 final polls from 27 pollsters at twelve general elections, 82 of them at the nine from 2000 that the model uses, against the official result. The average missed each modelled party by 1.21 points and each main party by 1.32, and the misses have a direction: the Conservatives were understated at all six elections from 2008, the NDP overstated at all nine from 2000, the Greens and the People's Party every time. We define the error, work through 2025, split shared from pollster-specific error, and give the gap between the leaders and every pollster's record.
Summary
This paper sets out how we measure polling error in Canada: which polls count as final polls, where they come from, how the error is defined, and what the record shows. We hold 92 final polls from 27 pollsters at the twelve general elections from 1988 to 2025. The model learns from the nine elections from 2000 on (82 final polls); 1988, 1993 and 1997 are kept and shown, but not used, under the site-wide rule that no election before 2000 shapes a forecast.
At those nine elections the average of the final polls missed each of the parties we model by 1.21 points, and each main party (5% of the vote or more) by 1.32. Unlike Germany's (Measuring German polling error, 2005–2025), the Canadian misses have a direction. The Conservatives were polled below their result at all six elections from 2008 to 2025, by between 0.75 and 2.62 points, and at the last three every single final poll had them too low. The NDP was polled too high at every election since 2000, by up to 2.70 points, and so were the Greens at all eight elections at which they were measured. The People's Party has been overstated at all three of its elections.
The record has four uses on the site. Its steady part is the basis of the correction (How we read Canadian polls: the average, the correction and its test). Its size sets the national error in the simulations (Simulating Canadian elections: where the chances come from). Each election's pattern is replayed on today's polls as a lens (If we're wrong: today's polls with each past miss replayed). And its pollster-by-pollster detail feeds the ratings (Rating Canadian pollsters: accuracy, the lead and openness).
What has changed since this paper was written
Updated 9 November 2026. The rest of the paper is as first published, except where a sentence below says it was corrected.
- When the polls missed (2 November): The Wildrose lead that vanished overnight: Alberta's 2012 polling miss and The young voters who stayed home: British Columbia's 2013 polling miss.
What is measured, and against what
The quantity that decides the House of Commons is the vote in each of the 343 ridings, but no pollster can poll every riding. What Canadian pollsters publish is a national share of the vote, and that is what we measure: each party's share of the valid votes cast across Canada, from Elections Canada's official results for 2011 to 2025 and the Library of Parliament's record of federal results for earlier elections.
Canadian releases report decided and leaning voters only: the undecided are left out before the shares are printed. We rescale every poll to 100 across the parties it names, so a poll that adds up to 99 or 101 through rounding is put on the same footing as the result.
Party lines need one decision. The Conservative Party of Canada was formed in December 2003 from the Progressive Conservatives and the Canadian Alliance, itself the successor of the Reform Party. For 1988 to 2000 we measure the Progressive Conservatives and Reform (the Alliance in 2000) together as the Conservative line, in the polls and in the result. That keeps one continuous series for the party that now competes with the Liberals for government. The two were rivals in those years, so for the riding replays in Testing the seat engine: the plurality check, 2025 and the swing replays they are kept apart.
The parties measured at an election are those in the result that appear in at least half of that election's final polls. In practice that is the Liberals, the Conservative line and the NDP throughout, the Bloc Québécois from 1993 (it first stood that year), the Greens from 2004 (pollsters rarely listed them before) and the People's Party from 2019.
Definitions
Let \(v_{p,t}\) be party \(p\)'s share of the valid vote at election \(t\), and \(y_{j,p,t}\) its share in final poll \(j\), rescaled to 100. The error of one poll is
The sign is the same everywhere on the site: error = poll minus result, so a positive error means the polls overstated the party.
The industry error is the error of the simple average of the \(n_{p,t}\) final polls that report the party:
Each pollster counts once, whatever its sample size.
The average miss per party at election \(t\) is the mean absolute industry error over the parties measured there:
We quote it on two sets of parties. The model's set is the six parties it forecasts (Liberal, Conservative, NDP, Bloc, Green, People's Party) wherever measured; this is the set the correction is tested on. The site's headline set, used on the accuracy pages, the country comparison and the pollster ratings, is the main parties only: those with at least 5% of the vote, "others" excluded (Measuring British polling error in England, 2015–2024 gives the reasoning). The two differ because small parties are usually missed by less in points, which pulls the first figure down.
For each party we summarise its errors across elections by their mean (direction) and root mean square (size):
where \(s_p^2\) is the variance of the errors around their mean. The identity on the right splits a party's misses into a steady part, \(\bar{e}_p^{\,2}\), and a part that changes from one election to the next.
Finally, if \(a\) and \(b\) are the parties that finished first and second in the vote, the gap error is
A positive gap error means the polls put the leader too far ahead.
The evidence: final polls and how they were chosen
What counts as a final poll
A final poll is each firm's last national poll whose fieldwork ended within ten days of polling day. One poll per firm per election is kept. Aggregators' averages are not polls and are left out.
| Election | Final polls | Election | Final polls |
|---|---|---|---|
| 1988 | 3 | 2008 | 7 |
| 1993 | 2 | 2011 | 10 |
| 1997 | 5 | 2015 | 8 |
| 2000 | 7 | 2019 | 12 |
| 2004 | 6 | 2021 | 13 |
| 2006 | 6 | 2025 | 13 |
Table: Final polls by election. 92 in all from 27 pollsters; 82 from 2000 on.
The field has grown and changed. Ipsos and EKOS have a final poll at all nine elections from 2000; Léger and Nanos at eight; Angus Reid and Innovative Research at five; Mainstreet, Abacus, Forum and Harris-Decima at four. Ten pollsters have one final poll each since 2000, five of them (Liaison, Pallas, Cardinal, Pollara and Focaldata) from 2025 alone.
Where the figures come from
Every final poll in the record is grade C: a figure taken from Wikipedia's tables of Canadian opinion polls (CC BY-SA 4.0), with a link to the pollster's release where one survives. No pollster's own release from before 2004 was found online, and for the elections since then we have not yet replaced the table figures with the releases. The national polls since the 2025 election are mostly grade A, from the pollsters' own releases (The Canada data: sources, grades and checks). A transcription error in a final poll would pass straight into the record.
A worked example: 2025
The 2025 election, on 28 April, has thirteen final polls.
| Pollster | Fieldwork end | Liberal | Conservative | Bloc | NDP | Green | PPC | Miss per party |
|---|---|---|---|---|---|---|---|---|
| Mainstreet Research | 27 Apr | 44.0 | 41.0 | 6.0 | 6.0 | 1.0 | 1.0 | 0.27 |
| Liaison Strategies | 27 Apr | 43.0 | 41.0 | 6.0 | 7.0 | 1.0 | 1.0 | 0.43 |
| Research Co. | 27 Apr | 44.0 | 39.0 | 6.0 | 7.0 | 2.0 | 1.0 | 0.77 |
| Nanos Research | 27 Apr | 42.7 | 40.0 | 6.4 | 7.8 | 1.9 | 1.1 | 0.85 |
| Pallas Data | 26 Apr | 43.2 | 38.8 | 6.4 | 7.3 | 2.3 | 0.9 | 0.90 |
| Cardinal Research | 23 Apr | 44.0 | 39.0 | 7.0 | 7.0 | 2.0 | – | 0.95 |
| Léger | 25 Apr | 43.0 | 39.0 | 6.0 | 8.0 | 2.0 | 1.0 | 1.03 |
| Ipsos Canada | 26 Apr | 42.4 | 38.4 | 6.1 | 9.1 | 2.0 | 1.0 | 1.40 |
| EKOS | 27 Apr | 43.7 | 37.6 | 4.4 | 9.2 | 2.0 | 2.0 | 1.77 |
| Abacus Data | 27 Apr | 41.0 | 39.0 | 6.0 | 10.0 | 1.0 | 3.0 | 1.93 |
| Pollara | 24 Apr | 41.0 | 38.0 | 6.0 | 10.0 | 2.0 | 2.0 | 2.03 |
| Focaldata | 23 Apr | 40.5 | 37.5 | 6.5 | 10.7 | 2.1 | 2.4 | 2.37 |
| Innovative Research | 26 Apr | 41.0 | 37.0 | 6.0 | 11.0 | 3.0 | 2.0 | 2.53 |
| Average | 42.59 | 38.88 | 6.06 | 8.47 | 1.87 | 1.54 | ||
| Result | 43.76 | 41.31 | 6.29 | 6.29 | 1.22 | 0.70 | ||
| Error | −1.17 | −2.43 | −0.23 | +2.19 | +0.66 | +0.84 | 1.25 |
Table: Final polls and result, 2025, percentage points, each poll rescaled to 100. Rows ordered by each poll's own miss over the parties it reports. Cardinal did not report the PPC, so the PPC average is over twelve polls.
The Conservative average is 38.88 against a result of 41.31, an error of −2.43 by equation (2). Over the six modelled parties the average miss is \((1.17 + 2.43 + 0.23 + 2.19 + 0.66 + 0.84)/6 = 1.25\) points by equation (3); over the four main parties (Liberal, Conservative, Bloc, NDP) it is 1.51. The Liberals won by 2.45 points; the polls had them 3.71 ahead, a gap error of +1.26 by equation (5).
Three features recur in the record. Every one of the thirteen polls had the Conservatives too low, by between 0.3 and 4.3 points, and twelve of the thirteen had the NDP too high. That is a miss shared by the industry, not the work of one or two firms. The best individual polls beat the average: seven of the thirteen missed by less than the average's 1.25. And the final polls were close in time: the earliest ended on 23 April, five days before polling day.
The record, 2000–2025
The same calculation at every election:
| Election | Polls | Liberal | Conservative | NDP | Bloc | Green | PPC | Miss, model parties | Miss, main parties |
|---|---|---|---|---|---|---|---|---|---|
| 1988 (not used) | 3 | +1.41 | −3.78 | +1.29 | – | – | – | 2.16 | 2.16 |
| 1993 (not used) | 2 | +1.55 | +0.26 | +0.03 | −0.71 | – | – | 0.64 | 0.64 |
| 1997 (not used) | 5 | +0.78 | +1.24 | +0.02 | −1.01 | – | – | 0.76 | 0.76 |
| 2000 | 7 | +0.27 | 0.00 | +0.66 | −0.83 | – | – | 0.44 | 0.44 |
| 2004 | 6 | −3.49 | +2.30 | +2.06 | −0.31 | +0.72 | – | 1.78 | 2.04 |
| 2006 | 6 | −2.22 | +1.99 | +1.26 | +1.00 | +0.77 | – | 1.45 | 1.62 |
| 2008 | 7 | +0.29 | −2.44 | +1.71 | −0.23 | +1.83 | – | 1.30 | 1.30 |
| 2011 | 10 | +0.98 | −1.83 | +0.04 | +0.24 | +1.16 | – | 0.85 | 0.77 |
| 2015 | 8 | −1.45 | −0.75 | +1.58 | +0.17 | +1.26 | – | 1.04 | 1.26 |
| 2019 | 12 | −1.02 | −2.38 | +2.70 | −0.28 | +0.89 | +0.86 | 1.35 | 1.45 |
| 2021 | 13 | −1.20 | −2.62 | +1.35 | −0.73 | +1.24 | +1.54 | 1.45 | 1.48 |
| 2025 | 13 | −1.17 | −2.43 | +2.19 | −0.23 | +0.66 | +0.84 | 1.25 | 1.51 |
| Mean, 2000–2025 | −1.00 | −0.91 | +1.50 | −0.13 | +1.06 | +1.08 | 1.21 | 1.32 |
Table: Average final poll minus result, points. Before 2004 the Conservative line is the Progressive Conservatives and Reform or the Alliance together. The main-party column counts parties with 5% of the vote or more.
Four points stand out.
The Conservatives have been understated since 2008. At all six elections from 2008 to 2025 the average had the party too low, by 2.08 points on average, and at the last three every final poll did. Before that the line ran the other way: in 2004 and 2006, the first two elections after the merger, the polls had the new party too high by about two points, and in 2000 they had the combined Progressive Conservative and Alliance vote almost exactly. In the record that is not used, 1988 had the Progressive Conservatives 3.8 points too low.
The NDP has never been understated. It was overstated at every election from 2000, by between 0.04 (2011) and 2.70 (2019) points, and polled almost exactly in 1993 and 1997.
The Greens and the People's Party are overstated. The Greens at all eight elections where they were measured, by between 0.66 and 1.83 points; the PPC at all three, by between 0.84 and 1.54. For parties polling two or three points, these are large relative misses.
The Liberals are mixed, the Bloc is close. The Liberals were understated in 2004 and 2006 (by 3.5 and 2.2 points) and at the four elections from 2015, and overstated in 2000, 2008 and 2011. The Bloc has been polled within about a point at every election, which matters because its national share is carried by one province (From national polls to 343 ridings: regions, Quebec and the swing model).
By equation (4):
| Party | Elections | Mean error | RMS error | Share of RMS² that is steady |
|---|---|---|---|---|
| Liberal | 9 | −1.00 | 1.64 | 37% |
| Conservative | 9 | −0.91 | 2.04 | 20% |
| NDP | 9 | +1.50 | 1.69 | 80% |
| Bloc | 9 | −0.13 | 0.54 | 6% |
| Green | 8 | +1.06 | 1.13 | 89% |
| PPC | 3 | +1.08 | 1.13 | 92% |
Table: Summary of industry errors, 2000–2025, points. The last column is \(\bar{e}_p^2 / \text{RMS}_p^2\).
For the NDP, the Greens and the PPC most of the error is a steady lean. For the Conservatives only a fifth is, because the overstatements of 2004 and 2006 offset the later understatements; since 2008 alone the party's mean is −2.08. For the Bloc almost none is. This is the pattern on which the correction rests: where the lean is steady, half of it is taken off today's polls (How we read Canadian polls: the average, the correction and its test).
Shared and pollster-specific error
The average squared error of the final polls at an election splits exactly into the square of the industry error and the spread of the polls around their average:
The share of the first term is the part of the miss common to the industry, which no choice of pollster would have avoided. For the recurring misses it is high: 80% for the Conservatives in 2019, 71% in 2021 and 81% in 2025; 81% for the NDP in 2019 and 66% in 2025; 98% for the Liberals in 2004, when every final poll had them between 2.7 and 4.1 points too low. In 2011 it was low for every party (27% for the Conservatives, none for the NDP): the polls were right on average but scattered widely, from COMPAS with the Conservatives 6.4 points too high to Innovative Research with the NDP 6.5 points too low.
The average was also closer than most of its members. In 2000 and 2011 it beat every individual final poll; in 2025 it beat six of the thirteen. Averaging removes the scatter between pollsters but not the shared part, and the shared part is what the correction addresses.
The gap between the two largest parties
Under first past the post the gap in votes does not decide the government on its own: in 2019 and 2021 the Conservatives won the most votes and the Liberals the most seats. It still matters, and the polls' record on it is mixed.
| Election | Top two | Real gap | Polls' gap error | Worst poll |
|---|---|---|---|---|
| 1988 (not used) | Conservative, Liberal | 13.2 | −5.2 | −8.2 |
| 1993 (not used) | Liberal, Conservative | 6.6 | +1.3 | +2.4 |
| 1997 (not used) | Liberal, Conservative | 0.3 | −0.5 | −7.3 |
| 2000 | Liberal, Conservative | 3.2 | +0.3 | +3.9 |
| 2004 | Liberal, Conservative | 7.1 | −5.8 | −7.1 |
| 2006 | Conservative, Liberal | 6.0 | +4.2 | +8.0 |
| 2008 | Conservative, Liberal | 11.4 | −2.7 | −6.4 |
| 2011 | Conservative, NDP | 9.0 | −1.9 | +11.0 |
| 2015 | Liberal, Conservative | 7.6 | −0.7 | −3.6 |
| 2019 | Conservative, Liberal | 1.2 | −1.4 | −5.5 |
| 2021 | Conservative, Liberal | 1.1 | −1.4 | −6.4 |
| 2025 | Liberal, Conservative | 2.5 | +1.3 | +3.7 |
Table: Gap between the two parties that finished first and second in the vote, points. The polls' gap error is equation (5) on the average; the worst poll is the single final poll with the largest absolute gap error.
2004 is the largest miss on the gap since 2000: the polls had the Liberals and the new Conservatives close to level, and the Liberals won by seven points. In 2019 and 2021 the average had the Liberals slightly ahead in the vote when the Conservatives finished about a point in front; under first past the post the Liberals still won more seats, so the error did not change who formed the government.
Pollster by pollster
Each pollster's record is the average miss per main party of its own final polls at the elections since 2000. This is the figure the ratings use (Rating Canadian pollsters: accuracy, the lead and openness).
| Pollster | Elections | Miss per main party |
|---|---|---|
| Liaison Strategies | 1 | 0.52 |
| Cardinal Research | 1 | 0.99 |
| Pallas Data | 1 | 1.03 |
| Mainstreet Research | 4 | 1.06 |
| Nanos Research | 8 | 1.16 |
| Research Co. | 3 | 1.30 |
| Léger | 8 | 1.41 |
| Harris-Decima | 4 | 1.54 |
| Abacus Data | 4 | 1.55 |
| Ipsos Canada | 9 | 1.58 |
| Angus Reid Institute | 5 | 1.61 |
| Forum Research | 4 | 1.62 |
| Environics | 2 | 1.66 |
| Counsel Public Affairs | 1 | 1.79 |
| The Strategic Counsel | 2 | 1.83 |
| Asking Canadians | 1 | 1.86 |
| Zogby International | 1 | 1.93 |
| COMPAS | 3 | 1.98 |
| Campaign Research | 2 | 1.99 |
| YouGov | 1 | 2.04 |
| EKOS Research | 9 | 2.07 |
| DART / Maru | 1 | 2.26 |
| Pollara | 1 | 2.52 |
| Innovative Research | 5 | 2.67 |
| Focaldata | 1 | 2.91 |
Table: Average miss per main party of each pollster's final polls, 2000–2025, points.
Two cautions apply, as in every country. The records cover different elections: Liaison, Cardinal and Pallas have only 2025, Ipsos and EKOS carry the 2004 and 2006 elections in theirs. And the differences between pollsters are of the same order as the differences between elections, so one good or bad election moves a short record a long way. The ratings handle this by pulling short records towards a typical Canadian final poll.
The pollsters' own leans repeat the industry pattern. Of the 25 pollsters with a final poll since 2000, all but COMPAS and Zogby (whose records end in 2011 and 2000) overstated the NDP on average, and all but those two, Environics and Asking Canadians understated the Conservatives. The miss is in the industry, not in a few houses.
How the record is used
The correction. Half of each party's mean error, learned from the other elections, is taken off today's polls, and the rule is kept because it beat the plain average at seven of nine elections in a leave-one-out test (How we read Canadian polls: the average, the correction and its test).
The simulation. Each party's national error in a simulated election has a standard deviation equal to its RMS error from the summary table, scaled to its size now (Simulating Canadian elections: where the chances come from).
The lenses. Each of the nine elections' error patterns is replayed on today's polls (If we're wrong: today's polls with each past miss replayed).
The ratings. Each pollster's own final polls are scored against every Canadian final poll since 2000 (Rating Canadian pollsters: accuracy, the lead and openness).
Before 2000
The three elections before 2000 are shown on the accuracy page and in the tables above, marked "not used by the model". They are unlike the later record in two ways. They were years of realignment: the Bloc and Reform broke through in 1993 and the Progressive Conservatives fell from 169 seats to 2. And their final polls are few (two to five per election) and second-hand. In 1988 the polls had the Progressive Conservatives 3.8 points too low, the largest single-party miss in the record; in 1993 and 1997 the average miss was under a point.
The model's cutoff year is a site-wide setting, 2000 by default. Until 28 September 2026 the Canadian record started in 1988, and the correction test then ran on twelve elections. The rule that replaced it is the same in every country: elections before the cutoff are kept for the record but shape nothing.
The long view on the accuracy page goes back to 1945, with the vote by party lineage at every general election from the Library of Parliament's record; we hold no final polls before 1988.
Limits
Nine elections. Every figure here rests on nine values per party, eight for the Greens and three for the PPC. A tenth election with a different pattern would move the means.
Second-hand figures. All 92 final polls come from Wikipedia's tables. No pollster's own release has yet been checked against them for the elections since 2004.
National shares only. The record measures national misses. A miss concentrated in one region (Quebec, say, or the Prairies) can matter more for seats than its national size suggests, and the record cannot show it.
Pollsters change. Several firms in the record have stopped publishing federal polls or changed method and ownership. A pollster's record describes what it did then.
References
Elections Canada, official voting results, 41st to 45th general elections (2011–2025). Open Government Licence – Canada.
Library of Parliament, History of the Federal Electoral Ridings, results 1867–2008 (open.canada.ca). Open Government Licence – Canada.
Jennings, W. and Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour, 2(4), 276–283.
Figures in this paper are as of 15 October 2026; the pages update with each poll.
How to cite
Polling Forecast (2026). How far Canadian polls have missed, 1988–2025. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/canada-polling-error. Live figures change as new polls arrive; the date above is when the text was last revised.



