What we don't know yet: limitations of the Canada forecast

Summary

This paper collects the known weaknesses of the Canada forecast in one place, says how much each matters, and what evidence would change the method. The forecast makes three claims: that Canadian polls have missed in a steady direction often enough for half of the average miss to be worth taking off; that uniform swing by region, moved from 2025, is the best of the simple ways to turn regional shares into riding winners; and that the simulation's errors have the size of past errors at the national, regional and riding levels. Each rests on a short record.

The main limits are these. The correction and the simulation's error sizes rest on nine elections, all of whose final polls are second-hand. The swing model has no candidates in it and moves every riding in a region by the same amount. Regional polls are few, and in some regions one pollster. The Bloc's whole vote is in one province. The simulation draws the parties' national errors independently and spreads them with today's regional pattern. The election date is fixed but not certain, and three years away. And the provincial forecasts take their polls uncorrected. The British counterpart is What we don't know yet: limitations and open questions; the method itself is in How we read Canadian polls: the average, the correction and its test.

Nine elections, and a correction that rests on them

Everything the forecast knows about Canadian polling error comes from the final polls at the nine general elections from 2000 to 2025, 82 polls in all (How far Canadian polls have missed, 1988–2025). The correction passes its leave-one-out test clearly (closer at seven of nine elections, 0.94 points per party against 1.21), but the pattern it learns is short in another sense. The Conservative understatement that drives it began in 2008. Before that, in 2004 and 2006, the polls overstated the new Conservative Party by about two points.

Quantity	2000–2025	2008–2025 only
Conservative mean miss	−0.91	−2.08
Conservative correction (half the mean)	+0.45	+1.04
Elections at which the Conservatives were understated	6 of 9	6 of 6

Table: How much of the Conservative pattern is recent. Points; from the final polls in the database.

The forecast uses the full nine elections, as the rule requires. If the post-2008 pattern is the true one, the Conservative correction is about half what it would be on that evidence alone; if 2004 and 2006 were the true pattern and 2008 to 2025 a run, it is too large. Nine elections cannot settle which.

The misses are also not independent draws. All the polls at an election share its electorate, campaign and late movement, so 82 polls carry about nine elections' worth of information. And polls far from polling day, which is where a forecast for 2029 starts, carry far more uncertainty than the final-poll record describes (Polling day: how far Canadian polls move before an election).

What would change the method. Each new general election is added to the record and the test rerun. If the correction stopped beating the polls as published, it would switch off without any change in the code. λ is fixed at ½ for every country; on the Canadian record a larger value would have done slightly better (¾ gives 0.85), and the site-wide setting is kept rather than tuned to one country.

Final polls are second-hand

All 92 final polls come from Wikipedia's tables of Canadian opinion polls (The Canada data: sources, grades and checks). No pollster's own release from before 2004 was found online, and for 2004 to 2025 the table figures have not yet been checked against the releases. A transcription error, a wrong fieldwork date (which could move a poll in or out of the ten-day window) or a figure for the wrong population would pass straight into the correction, the simulation's error sizes, the lenses and the ratings.

What would change the method. Replacing the figures for 2004 to 2025 with the pollsters' releases where they survive, and recording which could not be found. A figure that changed after checking would change the record and the test with it.

Uniform swing from 2025, without candidates

Every riding moves from its 2025 result by its region's change in points. The replays show this is the best of the three simple rules tested: 268 ridings wrong across seven pairs of elections, against 278 and 288 for the alternatives (Testing the seat engine: the plurality check, 2025 and the swing replays). But the same replays show its limits. Even with the true regional change, about one riding in nine went to a different party at the most recent pair, 2021→2025, and one in five at 2011→2015.

The model has no candidates in it. Incumbency, retirements, star candidates and members who crossed the floor are invisible; every riding starts from its 2025 vote. Since the 2025 election five members have crossed the floor to the Liberals (four Conservatives and one New Democrat), seven by-elections have been held and six seats are vacant. A riding whose 2025 member now sits for another party is projected from the 2025 vote, as if the floor-crossing had not happened: Nunavut, whose member elected for the NDP now sits as a Liberal, is projected NDP.

Uniform swing within a region also means that a region's parts move together. Toronto and its suburbs, Montreal and the rest of Quebec, the cities and the countryside of the Prairies all get the same change in points. Given the true Ontario change, the 2021→2025 replay still had ten Ontario ridings that the Conservatives won projected Liberal among its 25 closest misses.

The riding noise in the simulation stands in for all of this. It is fitted so that the replays would have been expected to miss as many ridings as they did, which calibrates the chances on average, but it is independent from riding to riding, while real swing-model errors tend to run in the same direction across similar places.

What would change the method. A swing rule using sub-regional information (the city and suburban ridings of Ontario, say, or the census profile of each riding, which is loaded but not used) would have to beat uniform regional swing in the same replays before replacing it.

Regional readings are thin

The regional step depends on what pollsters publish. On 15 October 2026 three pollsters' regional figures are in the 90-day window for Quebec and Ontario, two for Manitoba and Saskatchewan, Alberta and British Columbia, one for Atlantic Canada and none for the territories (From national polls to 343 ridings: regions, Quebec and the swing model). Léger supplies more than half of all regional readings since 2025, and its monthly report is the only one that covers all six regions in a table.

With two or three pollsters a region's reading can move a long way on one release. In British Columbia the two current readings put the Liberals 3 and 13 points above the same pollsters' national figures. Some releases name only some parties in a region, and a stand-alone regional poll has to be measured against our national average rather than its own national figure.

Measuring each regional figure against its own release's national figure removes a pollster's national house effect, but not a house effect that is regional: a pollster whose Quebec sample leans one way.

What would change the method. More pollsters publishing regional tables. The blend weight, \(n/(n+2)\), was chosen, not fitted; with a longer series of regional readings it could be tested against the 2029 result.

Quebec and the Bloc

The Bloc's national share of about 6% is all in Quebec, where it is about 27%. A national point for the Bloc is about four and a half points in Quebec, and moves about five seats. Two consequences follow.

The Bloc's national polling record is good (within about a point at every election), but a national miss of half a point is a Quebec miss of two, and the Quebec regional readings come from three pollsters. The Bloc's seat range today, 11 to 27, is wide for that reason.

On polling day the drift is worse. The polling-day simulation scales each party's drift by its national size, which gives the Bloc a national standard deviation of about 4.2 points, close to 18 points in Quebec terms. Its polling-day range runs from no seats to 43. Whether the Bloc's vote is really that uncertain three years out, or whether scaling by national size overstates it for a party concentrated in one province, has not been tested.

The simulation's structure

Independent party errors. Each party's national error is drawn independently. At six of the eight elections since 2000 where both moved, the Conservative and NDP misses went in opposite directions. A simulation that drew them together would give a wider range for the gap between the main parties. The past-miss lenses carry that pattern, but they are not probabilities (If we're wrong: today's polls with each past miss replayed).

Today's regional pattern in every run. A simulated national error is spread over the regions in proportion to each party's regional strength today. A polling miss concentrated in one region (the Conservative understatement in Ontario, say) is represented only through the separate regional noise, which was fitted on how far regions strayed from national swing when the national vote was known, not on regional polling error.

The centre. The simulation is centred on the corrected shares, with errors the full size of past misses. Since the correction has already taken off half of the steady part, the spread is, if anything, a little wide for the parties with the steadiest leans.

Calibration. The national error has the size of past final-poll misses, and the riding noise reproduces the replays' misses on average. The joint behaviour of parties and regions is not calibrated, and nine elections could not calibrate it (Simulating Canadian elections: where the chances come from).

The election date and the long horizon

The fixed date, 15 October 2029, is three years away, and the polling-day forecast is much less certain than the projection for an election held today: a Liberal majority in 67% of runs against over 99%. Three things limit it.

The date is not certain. Two of the last three elections (2021 and 2025) came before their fixed dates. An early election would shorten the horizon and narrow the ranges at once.

The pull-back towards the 2025 result passed its test by a small margin (3.78 points per party against 3.81, averaged over ten distances) and was worse than the polls as they stood between nine months and two years out. The rule takes what wins on average.

The drift curve rests on eight or nine elections at each distance, including campaigns (2011 and 2015) in which the polls a month before polling day still missed by about four points per party. Nine elections are a small basis for the size of a rare event.

Pollster ratings

The ratings rank each pollster's final polls against all 82 Canadian final polls since 2000 and shrink short records towards the median (Rating Canadian pollsters: accuracy, the lead and openness). They inherit the record's limits: second-hand final polls, records covering different elections (those that polled in 2004 and 2006 carry the post-merger misses), and openness judged from what was on each site on the day of the review. Grades compare Canadian pollsters with each other, not with those of other countries. The ratings do not weight the polls.

Provincial forecasts

The Alberta and Manitoba nowcasts take their polls as published: the provinces' own polling record has not been collected, so no correction can be tested. Their polling error in the simulations is a setting (4 points for the main parties), not an estimate, and so are their regional and district noise. Manitoba's average rests on one pollster. Both start from the 2023 results, before by-elections, and Alberta's on the 2023 districts, since the map for 2027 is not settled.

By-elections, floor-crossings and vacancies

The forecast is for a general election and starts from the 2025 result in every riding. It does not use the by-elections since 2025, which have low turnout and local causes, and it does not use today's House, in which the Liberals have 173 seats after floor-crossings and by-elections. The six vacant seats are forecast like any other riding.

Summary of limitations

Limitation	How much it matters	What would address it
Nine elections	The correction and error sizes rest on them; the Conservative pattern is six elections long	More elections; the test reruns itself
Second-hand final polls	A transcription error passes into every use of the record	Checking 2004–2025 against the releases
No candidates	Close ridings uncertain; floor-crossings ignored	Candidate data, tested in the replays
Uniform swing within regions	Sub-regional swings missed	A sub-regional rule that beats uniform swing in the replays
Thin regional readings	A region can move on one release	More regional tables from more pollsters
The Bloc	Small national errors are large in Quebec	Quebec-specific error sizes, tested
Independent party errors	Range of the main-party gap understated	Correlated draws from past error patterns
Today's regional pattern in every run	Regional polling misses only through noise	A regional polling-error record
Election date	Early election would narrow the ranges suddenly	None possible
Provincial forecasts uncorrected	Provincial polling misses unmodelled	Collecting the provinces' polling record

Table: Known limitations of the Canada forecast.

The forecast's claim is narrow: Canadian polls, corrected by a rule that has passed its test, carried through a seat engine tested on seven past elections, with uncertainty sized from the record. The next general election will be its first real test.

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

How to cite

Polling Forecast (2026). What we don't know yet: limitations of the Canada forecast. Retrieved 17 October 2026 from https://www.pollingforecast.com/research/canada-limitations. Live figures change as new polls arrive; the date above is when the text was last revised.