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

Testing the seat engine: the plurality check, 2025 and the swing replays

Abstract. Three tests of the Canada seat engine. The plurality rule on the official riding votes gives the declared seats at all twelve general elections from 1988 to 2025. Fed the real 2025 regional votes, the engine gives back the 2025 House exactly. And replaying seven pairs of elections from 2004→2006 to 2021→2025 with the true regional change, uniform swing got 268 of 2,281 ridings wrong, against 278 mixed and 288 proportional, so it is used. The same replays size the regional and riding noise in the simulations.

Polling Forecast · updated 9 November 2026 · 9 min read · Canada, backtest, swing, uniform swing, proportional swing, transposed results, seat check, riding noise, regional noise

Summary

This paper reports the tests behind the Canada seat engine, the part of the forecast that turns regional vote shares into 343 riding winners. There are three. First, the plurality rule applied to the official riding votes gives the declared seats at all twelve general elections from 1988 to 2025, party by party. Second, the whole engine fed the real 2025 regional votes gives back the 2025 House exactly. Third, and hardest, each later election is projected riding by riding from the earlier one, given the true change in each region, and the ridings that go to the wrong party are counted.

Across the seven pairs of elections the model uses (2004→2006 to 2021→2025, 2,281 ridings in all), uniform swing by region got 268 ridings wrong, against 278 for an even mix of uniform and proportional swing and 288 for proportional; applied nationally instead of by region, the best method got 373 wrong. So the forecast uses uniform regional swing. The same replays set the size of the regional and riding-level noise in the simulations: the riding noise is fixed so that, run with the true regional change, the replays would have been expected to miss as many ridings as they actually did.

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.

  • The map (9 November). Every federal election since 1988 can be shown on the map on the riding boundaries of its day; the seat tests are unchanged.

Three checks

The seat engine has two parts that can go wrong. The rule that turns votes into seats could be misapplied, and the swing that moves each riding from its last result could be the wrong shape. The first two checks test the rule and the arithmetic; the third tests the swing.

None of them tests the polls. In every check the vote is known, nationally or by region, and the question is only whether the engine turns a known vote into the right seats. How far the polls might be from the vote is measured separately (How far Canadian polls have missed, 1988–2025) and enters the forecast through the simulations (Simulating Canadian elections: where the chances come from).

Check 1: the plurality rule

Each riding elects the candidate with the most votes (Canada Elections Act, s.313). Applied to the official riding votes, with the candidates summed by party, our count must match the seats declared. It does at every election in the record:

Election Ridings Liberal Conservative NDP Bloc Green Others Match
2000 301 172 78 13 38 0 0 ✓
2004 308 135 99 19 54 0 1 ✓
2006 308 103 124 29 51 0 1 ✓
2008 308 77 143 37 49 0 2 ✓
2011 308 34 166 103 4 1 0 ✓
2015 338 184 99 44 10 1 0 ✓
2019 338 157 121 24 32 3 1 ✓
2021 338 160 119 25 32 2 0 ✓
2025 343 169 144 7 22 1 0 ✓

Table: Seats from the plurality rule on the official riding votes, which match the declared seats party by party. In 2000 the Conservative column is Reform/Alliance 66 and the Progressive Conservatives 12. 1988, 1993 and 1997 also match and are shown on the method page.

The check is simple, but it guards against real errors in the data: a misattributed candidate, a riding loaded twice, a by-election result mixed with a general election. The riding results to 2008 come from the Library of Parliament's record and those since 2011 from Elections Canada, so the check covers both sources (The Canada data: sources, grades and checks).

The table also shows why the forecast must work riding by riding. In 2019 and 2021 the Conservatives won the national vote (34.3% and 33.7%, against 33.1% and 32.6% for the Liberals) and the Liberals won 36 and 41 more seats.

Check 2: reproducing 2025

The second check runs the whole engine, regions and swing included, with the real 2025 regional votes as its input. Since every riding starts from its 2025 result, a region's change is zero and every riding must come back as it was:

Liberal Conservative Bloc NDP Green
Official 2025 169 144 22 7 1
Seat engine, fed the 2025 regional votes 169 144 22 7 1

Table: Reproduction check.

It must pass, but it confirms that the regional aggregation, the swing arithmetic and the rescaling change nothing when nothing has changed. It is rerun every time the forecast is built.

Check 3: the swing replays

Design

The real test of a swing model is to project an election it has not seen. For each pair of consecutive general elections we take the earlier result riding by riding, apply the true change between the two elections, nationally or in each region, and compare the projected winner with the actual one. A method that gets the swing right in shape gets the winners right wherever the true change is known.

Each pair must use one set of ridings, so that a riding's earlier and later results refer to the same place:

  • 1988→1993 on the 295 ridings of the 1987 Representation Order, and 1997→2000 on the 301 of 1996.
  • 2004→2006, 2006→2008 and 2008→2011 on the 308 ridings of 2003.
  • 2011→2015 on the 338 ridings of 2013, using Elections Canada's transposition of the 2011 votes onto them. Its notional seats are Conservative 188, NDP 109, Liberal 36, Bloc 4 and Green 1.
  • 2015→2019 and 2019→2021 on the 338 ridings of 2013, as drawn.
  • 2021→2025 on the 343 ridings of 2023, using the transposition of the 2021 votes. It reproduces the official 2021 national totals to within 9 votes per party; its notional seats are Liberal 157, Conservative 126, Bloc 34, NDP 24 and Green 2.

Before 2004 the Progressive Conservatives and Reform (the Canadian Alliance in 2000) are kept apart in the replays, because they fought each other riding by riding; merging them would hide the very contests that decided those elections.

Methods

Three swing rules are compared. Let \(b_{i,p}\) be party \(p\)'s earlier share in riding \(i\), and \(B_p\) and \(R_p\) its earlier and later shares in the riding's region (or nationally):

\[ \text{uniform: } s_{i,p} = \max(0,\ b_{i,p} + R_p - B_p), \qquad \text{proportional: } s_{i,p} = b_{i,p}\,\frac{R_p}{B_p}, \qquad \text{mixed: } \tfrac12(\text{uniform} + \text{proportional}). \tag{1} \]

In each case the riding is rescaled to 100, and a party with no earlier vote in the riding stays at zero unless it had none in the whole region, when it takes the regional share. Uniform swing adds the same number of points everywhere in the region; proportional swing multiplies by the same ratio, so a party gains or loses most where it was strongest; the mix sits between.

Which pairs count

Two pairs are shown but not counted. 1988→1993 is not a test of swing at all: the Bloc did not stand in 1988 and Reform ran in only 72 ridings, so for most of the 1993 vote there was no earlier share to move. 1997→2000 is left out under the site-wide rule that the model uses no election before 2000. The method choice and the noise fits rest on the seven pairs from 2004→2006 on, 2,281 ridings in all.

Results

Replay Ridings National uniform National proportional National mixed Regional uniform Regional proportional Regional mixed
1988→1993 (not counted) 295 98 84 82 41 42 40
1997→2000 (not counted) 301 34 40 39 37 40 37
2004→2006 308 40 48 42 31 30 31
2006→2008 308 36 32 32 36 41 37
2008→2011 308 93 91 91 34 38 34
2011→2015 338 83 86 87 63 69 69
2015→2019 338 58 55 56 39 42 42
2019→2021 338 23 27 26 25 23 24
2021→2025 343 40 45 43 40 45 41
Total counted 2,281 373 384 377 268 288 278

Table: Ridings given to the wrong party in each replay.

Replay Regional uniform Regional proportional Regional mixed
2004→2006 2.54 2.50 2.45
2006→2008 3.19 3.27 3.19
2008→2011 3.00 3.00 2.84
2011→2015 3.48 4.00 3.46
2015→2019 3.01 2.89 2.88
2019→2021 2.32 2.13 2.17
2021→2025 3.39 3.56 3.17

Table: Average share error per party per riding, points, for parties with at least 1% projected or actual.

What the replays show

Regions matter more than the shape of the swing. Moving from national to regional change cuts the misses by over a quarter (373 to 268 with uniform swing). The difference between the three shapes, applied regionally, is 20 ridings across seven elections. Regional swing helps most when one region moves on its own: in 2008→2011, when the NDP swept Quebec, national swing got 93 ridings wrong and regional swing 34; in 2015→2019, when the NDP collapsed in Quebec and the Bloc came back, 58 against 39.

Uniform swing gets the most winners right. It had the fewest misses in total, the fewest in four of the seven counted replays and was level with the mixed rule in a fifth. The mixed rule gives slightly smaller share errors in most replays, and proportional swing does best in 2019→2021, but the forecast is about winners, and the choice is made on ridings.

The hardest replay is 2011→2015. The Liberals rose from third place, with 34 seats, to a majority of 184. Even with the true regional change, uniform swing missed 63 ridings, mostly by giving the NDP and Conservatives seats the Liberals took: projected Liberal 164, Conservative 115, NDP 55, against 184, 99 and 44. A party that gains twenty points in a region does not gain them evenly.

2021→2025 overstated the Liberals. With the true regional change, uniform swing projected Liberal 187, Conservative 131, Bloc 21, NDP 4 and Green 0, against the real 169, 144, 22, 7 and 1. Of the 25 closest of its 40 misses, ten were Ontario ridings projected Liberal that the Conservatives won, among them Niagara South and Cambridge. Elsewhere the misses ran both ways: Kelowna and Peterborough were projected Conservative and went Liberal; Winnipeg Centre was projected Liberal and stayed NDP.

That last replay is the closest analogue to the forecast's task, which moves the 2025 result by today's regional change, and it is a fair warning about its accuracy. Even with the regional vote known exactly, about one riding in nine went to a different party from the projection.

Choosing the method

The code runs every replay each time the forecast is built and picks the swing method with the fewest ridings wrong across the counted regional replays, breaking a tie by the smaller summed share error. With the current record that is uniform swing, by 268 to 278. If a future election changed the order, the method would change with it, as would the noise fits below.

The choice does not depend on which pairs are counted: on the three most recent pairs alone uniform swing had 104 misses, the mixed rule 107 and proportional 110, and with 1997→2000 included the order was the same.

From the replays to the noise

The replays also say how far the engine's regional and riding projections stray from reality when the national vote is known, and the simulations use that to size two of their three layers of error (Simulating Canadian elections: where the chances come from).

Regional noise

For each counted pair and each region, we compare the region's actual later share with what proportional national swing would have given it. The regional noise coefficient is the root mean square of those differences, each scaled by the square root of the party's share:

\[ c_{\text{reg}} = \sqrt{\frac{1}{K}\sum_{t,r,p} \frac{\big(\hat{R}_{t,r,p} - R_{t,r,p}\big)^2}{\max(R_{t,r,p},\,1)}}, \tag{2} \]

where \(\hat{R}\) is the proportional-swing projection, \(R\) the actual regional share, and the sum runs over the \(K\) region-party pairs with a share in either. On the seven counted pairs \(c_{\text{reg}} = 0.99\). In a simulation each party's regional share then moves with standard deviation \(c_{\text{reg}}\sqrt{\text{share}}\), about 5.4 points for a party on 30% in a region with no regional polls, less where regional polls exist.

Riding noise

For the ridings, the question is how much noise around the projected shares makes the projection lose as many ridings as it really did. Let \(s_{i,p}\) be the projected shares in the counted regional uniform replays and \(\hat{w}_i\) the projected winner. With noise of standard deviation \(c\sqrt{s_{i,p}}\) added to each party, the expected number of ridings the projected winner loses is

\[ M(c) = \sum_i \Pr\Big(\arg\max_p \big(s_{i,p} + c\sqrt{s_{i,p}}\,z_{i,p}\big) \neq \hat{w}_i\Big), \qquad z_{i,p} \sim \mathcal{N}(0,1). \tag{3} \]

\(M(c)\) rises with \(c\). The code finds the \(c\) at which \(M(c)\) equals the 268 ridings actually missed, estimating each probability from 150 seeded draws per riding and bisecting between 0.05 and 3 for ten steps. The result is \(c = 1.01\). A party on 40% in a riding has a riding-level standard deviation of about 6.4 points; one on 10%, about 3.2.

Fitting the noise this way calibrates the simulation's riding chances against the past directly: summed over the 2,281 replayed ridings, the chance that the projected winner loses equals the number that did. It does not guarantee that the chance is right riding by riding, only on average.

What the tests do not show

The polls. Every replay is given the true change. The forecast is given a polled change, with its own error on top.

Changes in the ridings themselves. Incumbents retiring, star candidates, floor-crossings and local issues are all inside the riding-level misses, and the noise represents them only on average.

Boundaries. The forecast moves 2025 results on the 2023 ridings, so it needs no transposition. The replays across a redistribution (2011→2015 and 2021→2025) depend on Elections Canada's transpositions, which are estimates of how the earlier votes would have fallen on the new lines.

Early elections. The record has no pair as far apart as 2025→2029 would be. Swings over four years may be larger and less even than those over two.

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

How to cite

Polling Forecast (2026). Testing the seat engine: the plurality check, 2025 and the swing replays. Retrieved 21 November 2026 from https://www.pollingforecast.com/research/canada-seat-tests. Live figures change as new polls arrive; the date above is when the text was last revised.

You may also like