Canada General election15 Oct 2029in 1,059 days Liberal majority 99% No majority 1% Calendar Track record Research
Research · Canada
EN FR

How we read Canadian polls: the average, the correction and its test

Abstract. The core Canada methods paper. The average takes each pollster's latest national poll within 21 days of the newest, one per pollster. The correction takes off half of each party's average past miss, and is kept only because it passes a leave-one-out test: across the nine general elections from 2000 to 2025 it cut the average miss per party from 1.21 to 0.94 points and was closer at seven of nine. Today it adds about half a point to the Liberals and Conservatives and takes three quarters of a point from the NDP.

Polling Forecast · updated 9 November 2026 · 9 min read · Canada, poll average, correction, lambda, leave-one-out, backtest, Liberal, Conservative, NDP, method

Summary

This is the core methods paper for the Canada section. It sets out how the national figure at the heart of the forecast is built: the poll average, the correction for past polling misses, and the test that decides whether the correction is used. It then summarises the steps that turn that figure into seats, each of which has its own paper.

The average takes each pollster's latest national poll within 21 days of the newest, one vote per pollster. The correction is the rule used for every country on the site: take off half of each party's average past miss. It is kept only if, with each past election predicted from the others, it would have come closer to the result than the polls as published. In Canada it does: across the nine general elections from 2000 to 2025 it cut the average miss per party from 1.21 to 0.94 points and was closer at seven of the nine. So, unlike in Germany, the Canadian projection corrects the polls. Today that adds about half a point to the Liberals and the Conservatives and takes about three quarters of a point from the NDP.

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.

  • Our figures beside the polls (26 October). Our projection differs from the poll average in Canada, so both are shown on the hub, the leader cards and the polls page.

The question

Canadian pollsters publish national shares of the vote among decided and leaning voters. The forecast needs a best estimate of each party's national share on polling day, read from those polls and from what is known about how they have missed before. The British work found that final polls missed in steady directions over several elections, and that taking off part of the average past miss would have improved forecasts made before each result was known (Correcting the raw answers: the method behind our UK projection). The rule we took from that work is applied to every country, and the evidence decides whether it is switched on.

Step 1: the average

Let \(J\) be the set of polls formed by taking the latest national poll of each pollster whose fieldwork ended within 21 days of the newest poll in the database. Every poll's figures are first rescaled to 100 across the parties it names (the undecided are already left out in Canadian releases), with the Progressive Conservatives, Reform and the Alliance counted as Conservative and everything outside the six modelled parties as "others". The average for party \(p\) is

\[ y_p = \frac{1}{|J_p|}\sum_{j \in J_p} y_{j,p}, \tag{1} \]

where \(J_p\) is the subset of \(J\) that reports the party. The six parties are then rescaled together with "others" so the shares add up to 100; if no poll reports "others", it takes whatever is left of 100.

Each pollster counts once. Nanos publishes a four-week rolling track every week and Liaison a two-week rolling track every week; in a simple average of every poll they would outweigh a pollster that publishes once a month. A poll must also name both the Liberals and the Conservatives to be used.

On 15 October 2026 the newest poll ended on 13 October, so the window runs from 22 September, and four pollsters have a poll in it:

Pollster Fieldwork end Liberal Conservative NDP Bloc Green PPC Others
Abacus Data 13 Oct 46.0 34.0 9.0 7.0 3.0 1.0 –
Liaison Strategies 10 Oct 45.0 29.0 15.0 6.0 2.0 2.0 1.0
Nanos Research 9 Oct 49.1 29.7 11.2 6.4 2.1 1.6 –
EKOS Research 2 Oct 47.3 26.4 15.3 4.6 3.2 1.8 1.4
Average 46.5 29.6 12.5 6.0 2.6 1.6 1.2

Table: The polls in today's average, rescaled to 100, and the average after the "others" step.

The window is a setting (21 days by default, on the admin page). Eleven national polls from four pollsters ended in the last 30 days. The spread between them is wide for the second party: Abacus has the Conservatives at 34, EKOS at 26. Averaging one poll per pollster keeps either from dominating.

Step 2: the correction

For each party \(p\) and past election \(t\), let \(e_{p,t}\) be the average of the final polls minus the result (How far Canadian polls have missed, 1988–2025). With \(T_p\) elections measured, the correction is

\[ \hat{v}_p = y_p - \lambda\,\bar{e}_p, \qquad \bar{e}_p = \frac{1}{T_p}\sum_{t=1}^{T_p} e_{p,t}, \qquad \lambda = \tfrac{1}{2}, \tag{2} \]

followed by a rescale so the shares keep the same total:

\[ v_p = \frac{\max(0.1,\ \hat{v}_p)}{\sum_q \max(0.1,\ \hat{v}_q)} \times \sum_q y_q. \tag{3} \]

A party is corrected only if it has been measured at two or more elections; "others" are never corrected. The past elections are the nine general elections from 2000 to 2025: the model uses no election before the site-wide cutoff year (2000 by default), so 1988, 1993 and 1997 stay in the record but count for nothing here.

Today's figures:

Party Mean past miss Elections Added Polls Our projection
Liberal −1.00 9 +0.50 46.5 47.4
Conservative −0.91 9 +0.45 29.6 30.3
NDP +1.50 9 −0.75 12.5 11.9
Bloc −0.13 9 +0.07 6.0 6.1
Green +1.06 8 −0.53 2.6 2.0
PPC +1.08 3 −0.54 1.6 1.1
Others – – – 1.2 1.2

Table: The correction on 15 October 2026, points. The last column is after the rescale of equation (3), so it differs slightly from polls plus adjustment.

The rescale matters here because the corrections do not sum to zero: the NDP, Greens and PPC lose 1.82 points between them and the Liberals, Conservatives and Bloc gain 1.02, so after the rescale each party's share is raised by a little under one percent of itself. The Liberal share moves from 46.5 to 47.4, nearly a point, because the rescale adds about 0.4 to the half point of correction.

Half is a deliberate middle course. The full average miss would treat the past as a reliable guide to the next miss, and none of the evidence supports that much confidence; no correction would ignore a lean that has held for years. λ is a single site-wide setting, the same in every country, and it is not tuned to Canada. On the Canadian record the choice is conservative: the test below gives an average miss of 1.07 points with λ = ¼, 0.94 with ½, 0.85 with ¾ and 0.86 with 1, all closer than the uncorrected 1.21. We keep ½ because a value fitted to one country's nine elections would fit its noise too.

Step 3: the test

The correction is kept only if it would have improved past forecasts made without knowing the result. Each of the nine elections is predicted with the correction learned from the other eight alone (leave-one-out cross-validation; Stone, 1974), and the corrected average miss per party is compared with the miss of the polls as published. The correction is used if, averaged over the elections, it is closer by more than 0.01 points.

Election Polls' average miss Corrected Verdict
2000 0.44 0.53 further out
2004 1.78 1.58 closer
2006 1.45 1.27 closer
2008 1.30 1.09 closer
2011 0.85 0.94 further out
2015 1.04 0.60 closer
2019 1.35 0.83 closer
2021 1.45 0.96 closer
2025 1.25 0.66 closer
Mean 1.21 0.94 closer at 7 of 9

Table: Leave-one-out test, average miss per modelled party, points.

The correction passes clearly. It was closer at seven elections and further out at two, both elections where the polls happened to be close for the parties it corrects (2000) or the NDP was polled almost exactly (2011). The gain is largest at the four most recent elections, where the Conservative and NDP misses were largest.

A worked case shows what the test does. To predict 2025, the correction is learned from 2000 to 2021 only. The mean misses then are Liberal −0.98, Conservative −0.72, NDP +1.42, Bloc −0.12, Green +1.12 and PPC +1.20, so the 2025 final-poll average (Liberal 42.59, Conservative 38.88, NDP 8.47) becomes, after the rescale, Liberal 43.50, Conservative 39.62 and NDP 7.84, against a result of 43.76, 41.31 and 6.29. Every party moves towards its result, and the average miss falls from 1.25 to 0.66.

By party, across the nine tests, the correction reduced the average miss for the Liberals (1.34 to 1.11 points), the Conservatives (1.86 to 1.67), the NDP (1.50 to 1.02), the Greens (1.06 to 0.55) and the PPC (1.08 to 0.56), and left the Bloc almost unchanged (0.45 to 0.43).

The test is scored on the model's six parties. Scored instead on the main parties only (5% of the vote or more), the measure used on the site's accuracy pages, it gives 1.32 points for the polls and 1.06 corrected. The verdict is the same.

The test reruns whenever the forecast is built. If a future election broke the pattern and the correction stopped beating the polls on average, the code would switch it off with no change to the method.

What the correction assumes

A correction learned from the past works only while the causes of the misses persist. We do not claim to know them. What the record shows is their regularity: the Conservatives understated at all six elections since 2008, the NDP overstated at all nine since 2000, the Greens at all eight and the PPC at all three where they were measured. Pollsters review their methods after each election, and a correction that worked for nine elections can stop working at the tenth. The test is the safeguard: it is rerun on every build, and it would drop the correction once the evidence turned against it.

The correction is applied to the national average only. It does not correct the regional figures pollsters publish, which are read relative to their own national figures and so inherit the national correction rather than carrying their own (From national polls to 343 ridings: regions, Quebec and the swing model).

From national shares to seats

The corrected shares go through five further steps, each developed in its own paper.

Step 4: regions. Seven regions (Atlantic Canada, Quebec, Ontario, Manitoba and Saskatchewan, Alberta, British Columbia and the territories) are read from the regional figures pollsters print in their releases, each measured against the same release's national figure, and blended with what national swing alone gives. The weight on the regional polls is \(n/(n+2)\) for \(n\) pollsters. The regions are then scaled so they add back up to the national shares (From national polls to 343 ridings: regions, Quebec and the swing model).

Step 5: ridings. Each riding's 2025 result moves by its region's change since 2025, by uniform swing, and the candidate with most votes wins (Canada Elections Act, s.313). Uniform swing was chosen because it got the fewest ridings wrong when seven pairs of past elections were replayed with the true regional change (Testing the seat engine: the plurality check, 2025 and the swing replays). The Bloc Québécois stands only in Quebec, so its national share is carried by the 78 Quebec ridings alone.

Step 6: simulation. 2,000 simulated elections, each with a national error per party the size of its past final-poll miss, a regional error, and riding-level noise, give the seat ranges and the chances of a majority and of official party status (Simulating Canadian elections: where the chances come from).

Step 7: polling day. The fixed date is 15 October 2029. The polling-day forecast pulls each party part of the way back towards its 2025 result, as past Canadian polls taken this far out suggest, and widens the ranges for the time left (Polling day: how far Canadian polls move before an election).

Step 8: lenses. Today's polls are replayed with each of the nine past elections' misses and run through the same seat engine, to show what a repeat of a real miss would do (If we're wrong: today's polls with each past miss replayed).

What the correction changes today

Party Seats on the polls as published Seats on our projection 2025 result
Liberal 238 241 169
Conservative 68 67 144
Bloc 22 21 22
NDP 13 12 7
Green 2 2 1

Table: Seats under the same seat engine, with and without the correction, 15 October 2026. 172 seats is a majority of the 343.

With a seventeen-point lead in the poll average, the correction changes few seats: three more for the Liberals, one fewer each for the Conservatives, the Bloc and the NDP. The same shift would matter more in a close race, when more ridings sit near the line between the two main parties.

The correction also changes the simulation's centre but not its spread. The chance of a Liberal majority is over 99% on our projection and on the polls as published.

Limits

The method rests on nine elections, all of whose final polls come from Wikipedia's tables (How far Canadian polls have missed, 1988–2025, The Canada data: sources, grades and checks). λ is fixed, not fitted. The average gives every pollster equal weight, whatever its record; the ratings are shown for readers but do not weight the polls (Rating Canadian pollsters: accuracy, the lead and openness). And the correction is national: if the Conservative miss is concentrated in some regions, the projection spreads it across all of them in proportion to each region's vote. What we don't know yet: limitations of the Canada forecast collects these and the other known weaknesses.

References

Canada Elections Act, S.C. 2000, c. 9, s.313.

Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society: Series B, 36(2), 111–147.

Figures in this paper are as of 15 October 2026; the pages update with each poll.

How to cite

Polling Forecast (2026). How we read Canadian polls: the average, the correction and its test. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/canada-method. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like