Polling day: how far Canadian polls move before an election
Abstract. Polls at ten distances from a month to three years before the nine general elections from 2000 to 2025, compared with the result and the previous result. Pulling each party part of the way back towards its previous result beat the polls as they stood on average, narrowly, so the polling-day forecast does it: three years out, today's polls count for 43% of the gap to 2025. The typical miss grows from about 3.9 points per party a month out to 6.6 at the fixed date, 15 October 2029, and the ranges widen to match.
Summary
Canada's next general election is fixed for 15 October 2029, 1,096 days away. Polls taken that far out are not forecasts of polling day: opinion moves. This paper measures how far Canadian polls have moved before past elections and how the forecast uses that. We took the polls at ten fixed distances, from a month to three years before polling day, at the nine general elections from 2000 to 2025, and compared them with the result and with each party's previous result.
Two findings follow. First, pulling each party part of the way back towards its previous result beat the polls as they stood, on average over all distances, when each election was predicted from the others; the margin was small (3.78 points per party against 3.81) but the rule keeps whatever wins, so the polling-day forecast does it. Three years out, today's polls count for 43% of the gap between them and the 2025 result. Second, the typical miss grows with distance, from about 3.9 points per party a month out to about 6.6 at the fixed date, and the polling-day simulation widens its ranges accordingly. On today's polls that turns a near-certain Liberal majority into a 67% chance, with a Conservative majority at 18%.
The question
The projection on the Canada home page answers a hypothetical question, the result of an election held today. The polling-day forecast answers the real one. It needs two things from the past: whether polls this far out should be taken as they stand or adjusted, and how far they are likely to be from the result. The method is the one used for Britain and every other country on the site (From today to polling day: how far British polls move before a general election); only the evidence is Canadian.
The evidence
For each general election from 2000 to 2025 we took the polls at ten distances before polling day: 30, 61, 91, 182, 273, 365, 548, 730, 913 and 1,095 days. At each distance, each pollster's polls within 30 days either side were averaged, and the pollsters' averages were then averaged, so every pollster counts once. The result is a poll average for each party at each distance and election, set beside the party's result at that election and its result at the previous one. The parties are the Liberals, the Conservative line, the NDP, the Bloc, the Greens and, from 2021, the People's Party; a party with no previous result (the PPC in 2019) is left out of the test.
That gives 437 points, each one party at one distance before one election. Not every election has every distance: 2006 and 2011 each lack two of the ten, and the 2006 series has a gap in the polls from July 2004 to March 2005. Canadian elections also come at irregular intervals (the parliaments elected in 2004, 2006 and 2008 lasted between nineteen months and under three years), so at the longest distances the polls were taken before the previous election. Eight or nine elections are available at each distance.
We also hold the series for 1988, 1993 and 1997. They are not used: the model uses no election before 2000.
The model
Months before an election a party's final share has often been closer to its previous result than its polls suggested. Let \(y_p(d)\) be the poll average \(d\) days before polling day and \(v^{\text{prev}}_p\) the party's share at the previous election. The forecast for polling day is
followed by a rescale so the shares keep the same total. The weight \(w(d)\) on today's polls falls from 1 on polling day towards a floor \(f\) far out, with a time constant \(\tau\). A floor of 1 would mean polls are taken as they stand at every distance.
The curve is fitted by least squares on every party at every distance and election, searching floors from 0 to 1 in steps of 0.01 and time constants from 5 to 600 days in steps of 5:
On the nine elections the best fit is \(f = 0.43\), \(\tau = 185\) days.
| Days before polling day | 30 | 91 | 182 | 365 | 730 | 1,096 |
|---|---|---|---|---|---|---|
| Weight on today's polls, \(w(d)\) | 0.91 | 0.78 | 0.64 | 0.51 | 0.44 | 0.43 |
Table: The fitted weight at selected distances.
A month out, today's polls get 91% of the weight; six months out, 64%; beyond about two years the weight is close to its floor of 43%.
The test
A curve fitted to all nine elections would flatter itself. Each election is therefore predicted with the curve fitted on the other eight, at every distance, and the pull-back's average miss per party is compared with that of the polls as they stand. The pull-back is used only if, averaged over all distances, it is closer by more than 0.01 points.
| Days before | Elections | Polls as they stand | Previous result | Pull-back | Elections where pull-back was closer |
|---|---|---|---|---|---|
| 30 | 9 | 1.99 | 4.09 | 2.01 | 5 |
| 61 | 9 | 3.15 | 4.11 | 3.03 | 6 |
| 91 | 9 | 3.70 | 4.11 | 3.46 | 8 |
| 182 | 9 | 4.23 | 4.11 | 4.02 | 8 |
| 273 | 9 | 3.67 | 4.11 | 3.96 | 5 |
| 365 | 8 | 3.98 | 4.20 | 4.30 | 5 |
| 548 | 8 | 3.33 | 3.91 | 3.66 | 3 |
| 730 | 8 | 3.86 | 3.96 | 4.15 | 4 |
| 913 | 8 | 4.49 | 4.24 | 4.40 | 5 |
| 1,095 | 9 | 5.72 | 4.24 | 4.81 | 7 |
| Mean | 3.81 | 3.78 |
Table: Leave-one-out test, average miss per party, points. "Previous result" is the miss of simply repeating each party's last result.
Averaged over the ten distances, the pull-back missed by 3.78 points per party and the polls as they stand by 3.81, so the pull-back is used. The gain is not even. It is clearest from two to six months out (closer at eight of nine elections at 91 and 182 days) and at three years (4.81 against 5.72, closer at seven of nine). Between nine months and two years the polls as they stood were closer on average. The rule takes the curve that wins on average rather than switching method by distance, which would fit the noise of nine elections.
The test also shows how little the previous result alone is worth near polling day (4.09 points a month out, against 1.99 for the polls) and how close the two become far out: three years before polling day, the previous result missed by 4.24 points per party and the polls by 5.72.
Election by election
| Days out | 2000 | 2004 | 2006 | 2008 | 2011 | 2015 | 2019 | 2021 | 2025 |
|---|---|---|---|---|---|---|---|---|---|
| 30 | 1.26 | 1.75 | 1.80 | 0.97 | 4.06 | 3.81 | 1.19 | 1.27 | 1.77 |
| 61 | 3.00 | 1.45 | 2.99 | 1.77 | 5.75 | 5.90 | 2.07 | 2.42 | 2.99 |
| 91 | 2.79 | 1.51 | 3.47 | 2.85 | 6.52 | 5.80 | 2.00 | 2.65 | 5.70 |
| 182 | 4.51 | 4.85 | 4.15 | 2.41 | 7.26 | 3.12 | 2.22 | 2.71 | 6.85 |
| 273 | 5.14 | 3.45 | 1.76 | 2.32 | 7.66 | 2.15 | 1.77 | 2.48 | 6.32 |
| 365 | 6.01 | 4.15 | – | 2.10 | 7.19 | 2.02 | 1.79 | 2.42 | 6.16 |
| 548 | 5.14 | 3.48 | 2.18 | 2.05 | – | 2.55 | 2.25 | 2.90 | 6.10 |
| 730 | 4.01 | 4.44 | 6.95 | 1.74 | – | 2.64 | 2.16 | 2.29 | 6.69 |
| 913 | 5.01 | 7.74 | – | 0.70 | 6.24 | 3.20 | 3.67 | 2.95 | 6.41 |
| 1,095 | 4.89 | 6.16 | 7.96 | 4.50 | 8.24 | 5.43 | 5.49 | 2.70 | 6.16 |
Table: The polls' average miss per party at each distance and election, points.
Three elections stand out. In 2011 and 2015 the polls a month out still missed by about four points per party: the NDP's surge in Quebec in 2011 and the Liberals' rise from third in 2015 both came in the last weeks of the campaign. In 2025 the polls were far off for most of the year before polling day (about six points per party from three months out), because the Liberal recovery under a new leader came only in the final months. These are the campaigns that make Canadian polls far out a weak guide, and they are what the ranges try to capture.
The uncertainty
The second use of the evidence is the size of the miss. At each distance the root mean square error of the pull-back forecast, each election predicted from the others, is fitted as
for \(d\) of 30 days or more; inside 30 days it shrinks linearly to zero on polling day. The fit is to the pull-back's error because the pull-back is the method in use.
| Days before polling day | 30 | 91 | 182 | 365 | 730 | 1,096 |
|---|---|---|---|---|---|---|
| Typical miss per party, \(\delta(d)\) | 3.85 | 4.69 | 5.21 | 5.73 | 6.25 | 6.56 |
Table: The fitted drift, points per party.
\(\delta(d)\) is a typical miss across parties of all sizes. In the polling-day simulation it is scaled to each party's size, \(\delta(d)\sqrt{\max(S_p,1)/15}\), so a party on 15% gets the full value, one on 47% about 1.8 times as much and one on 6% about two thirds. That drift is added in quadrature to the party's final-poll error (Simulating Canadian elections: where the chances come from).
Today's polling-day forecast
At 1,096 days the weight on today's polls is 0.43. Each party is moved most of the way back towards its 2025 result:
| Party | 2025 result | Our projection today | Polling-day share | Seats today | Seats, polling day | 80% range, polling day |
|---|---|---|---|---|---|---|
| Liberal | 43.8 | 47.4 | 45.3 | 241 | 208 | 111–275 |
| Conservative | 41.3 | 30.3 | 36.6 | 67 | 101 | 44–191 |
| Bloc | 6.3 | 6.1 | 6.2 | 21 | 22 | 0–43 |
| NDP | 6.3 | 11.9 | 8.7 | 12 | 10 | 2–15 |
| Green | 1.2 | 2.0 | 1.6 | 2 | 2 | 0–2 |
| PPC | 0.7 | 1.1 | 0.9 | 0 | 0 | 0–0 |
Table: Polling-day forecast, 15 October 2026, from 1,000 simulated elections.
The Conservative share is a worked case: \(41.3 + 0.43 \times (30.3 - 41.3) = 36.6\), by equation (1). The Liberal lead shrinks from seventeen points to under nine, and the central seat projection from 241 to 208, still a majority.
| Outcome | Polling day |
|---|---|
| Liberal majority | 67% |
| Liberal minority | 8% |
| Conservative majority | 18% |
| Conservative minority | 6% |
| Another party largest, or a tie | under 1% |
Table: Who governs on polling day, 15 October 2026.
The ranges are very wide, as they should be three years out. The Conservatives' 80% range runs from 44 seats to 191. The Bloc's runs from none to 43, because its national drift of about 4.2 points is spread over Quebec alone. The NDP reaches the twelve seats needed for official party status in 26% of runs, the Bloc in 65%.
Before the cutoff
Until 28 September 2026 the fit also used 1988, 1993 and 1997. Those campaigns moved the polls far more than recent ones (the Progressive Conservatives' collapse in 1993, the arrival of the Bloc and Reform), and with them the best fit was \(f = 0.25\), \(\tau = 185\): today's polls would have counted for a quarter of the gap three years out. The site-wide rule now keeps them out of every calibration, and the pull-back is gentler as a result.
Limits
The date. 15 October 2029 is the fixed date under s.56.1 of the Canada Elections Act, but the Prime Minister can ask the Governor General to dissolve Parliament earlier, as happened before the 2021 and 2025 elections. An early election would shorten the distance and narrow the ranges at once; the forecast has no way to anticipate it.
Leaders and campaigns. The NDP chose a new leader in March 2026 and the Greens are choosing one. Canadian campaigns have moved polls a great deal (the NDP in 2011, the Liberals in 2015 and 2025), which is exactly what the ranges are meant to capture, but nine elections are a small basis for the size of a rare event.
Thin evidence far out. At 365 days and beyond only eight elections are available, and the per-election misses vary from under one point to over eight. The fitted curve smooths that variation.
One pull-back for all parties. The same weight is applied to every party. A party whose 2025 result was unusual (the NDP's 6.3%, its lowest since 2000) may not return towards it in the same way as a party at its usual level.
References
Canada Elections Act, S.C. 2000, c. 9, s.56.1.
Figures in this paper are as of 15 October 2026; the pages update with each poll.
How to cite
Polling Forecast (2026). Polling day: how far Canadian polls move before an election. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/canada-election-day. Live figures change as new polls arrive; the date above is when the text was last revised.



