How the seat projections work

• Polls. Every province-wide poll of British Columbia published since the 2024 election, from the polling tables on Wikipedia, as used by the site’s polling averages. Results for OneBC and CentreBC are read from the same tables’ own columns for those parties.
• The 2024 results in each of the 93 ridings, from Elections BC, as shares of valid votes. The 2024 boundaries are used again in 2026, so no transposition is needed.
• Candidates for 2026, from Wikipedia’s candidate lists checked against Elections BC, so that a party is only projected to win votes where it has a candidate.
• Regional poll results (Metro Vancouver, Vancouver Island, the rest of the province) from pollsters’ own releases, entered by hand.

The province-wide figures are the site’s polling average on the date of the newest poll: a weighted average in which a poll counts less as it ages (a Gaussian kernel with a 21-day bandwidth), more with a larger sample (square root of the sample, capped at 3,000) and less when its pollster has published several polls in the same weeks. It covers the NDP, the Conservatives, the Greens, OneBC and CentreBC; other parties and independents share the remainder.

Each riding’s 2024 result is then moved toward that average. Half of the change is applied as a uniform shift in percentage points and half in proportion to each party’s 2024 share in the riding (a party below 1% moves uniformly only). Shares are kept non-negative and summing to 100, and the ridings are then adjusted together, weighted by votes, until they reproduce the provincial average to within 0.001 points. In the backtests this blend predicted riding results about as well as either form alone.

BC United, formerly the BC Liberal Party, ran no candidates in 2024 and none in 2026. Most 2024 independents were former BC United members or candidates. Where such an independent is not running again, their vote is reallocated: 30% to CentreBC where it has a candidate, and the rest as 2020 BC Liberal voters said they would vote in Research Co.’s survey of October 2024 (Conservative 63%, NDP 25%, Green 7%, other 4%). A 2024 independent who runs again keeps that vote under their 2026 label.

OneBC and CentreBC did not exist in 2024. Each enters the 2024 baseline at its polled share, distributed in proportion to the 2024 Conservative vote, and taken from the parties its voters say they previously supported (from vote-recall questions by Angus Reid, September 2026, and Research Co., August 2026): OneBC 85% from the Conservatives, 10% from the NDP, 5% from others; CentreBC 50% from the Conservatives, 37% from the NDP, 13% from the Greens. In a test on Quebec’s Conservatives, who grew from 1.8% to 12.9% between 2018 and 2022, placing a growing party’s vote where a related party was strong located it better than spreading it evenly across the province.

A party with no candidate in a riding gets no votes there, and its projected share goes to the parties that are running. Green voters without a Green candidate are divided as 2024 Green voters who have since left the party say they will vote (NDP 60%, Conservative 21%, CentreBC 13%, OneBC 3%, other 3%, the mean of three recall tables); OneBC and CentreBC voters return to the parties they came from; any other share is divided in proportion. The projection makes no adjustment for incumbency or for individual candidates.

The exception is a sitting member running as an independent in their own riding, or an independent who won or took at least 15% there in 2024. Such a candidate keeps a share of their own 2024 vote: 39% of it if they were elected under a party label, 105% if they ran as an independent. These are the medians of every such case in Canadian provincial elections on the site (ten members who left their party and ran again, and six independents who ran again); each simulation draws the ratio at random around that median. In 2026 the rule applies in one riding, Peace River North, where Jordan Kealy, elected as a Conservative with 74.4% in 2024, runs as an independent: 29.0 points are carried forward.

A province-wide shift can misplace votes between regions. Where pollsters have published regional results within the 30 days before the newest poll, each region’s difference from that poll’s own provincial figure is averaged (weighted by regional sample size and recency) and the model’s regional result is pulled toward it. The weight given to the polls grows with their regional sample and shrinks when they disagree with each other; the shifts are balanced so the provincial totals do not change. Three regions are used: Metro Vancouver, Vancouver Island, and the rest of the province. Projections without recent regional polls use the province-wide shift alone.

The projection is simulated 10,000 times. Each simulation adds three kinds of error to every riding’s projected shares: a province-wide polling error shared by all ridings, a regional error shared within each of four regions (Metro Vancouver, the Fraser Valley, Vancouver Island, and the Interior and North), and an error for each riding. Each error moves votes between parties, so shares still sum to 100, and is larger for parties near 50% than for small parties. Their sizes come from the backtests below: 4.6, 5.25 and 6.8 percentage points respectively for a party at 50%, with heavier tails than a normal distribution (Student’s t with 4 degrees of freedom), because large riding misses happen more often than a normal curve allows.

From the simulations come each party’s median share in each riding and the range that holds the middle 90% of simulations; its chance of winning each riding; its median seat total and 90% range; and the chance that it wins a majority (47 or more of the 93 seats). The chance of no majority is the share of simulations in which no party reaches 47. A riding is Solid for a party with a 95% or higher chance of winning it, Likely at 80 to 95%, Leans at 60 to 80%, and a Toss-up when no party is above 60%.

The method was run, unchanged, on 10 provincial elections from 2020 to 2025 that had published polls and the same boundaries as the election before (766 ridings): Ontario 2022 and 2025, Alberta 2023, Manitoba 2023, Nova Scotia 2024, Quebec 2022, Prince Edward Island 2023, Newfoundland and Labrador 2021 and 2025, and British Columbia 2020. Each used only the polls published before election day. The error sizes were chosen as the most accurate setting whose 90% ranges contained at least 90% of actual riding shares and seat totals, with each election’s error sizes estimated without that election.

The projection did better than assuming every riding re-elects its party in 7 of the 10 elections; it did worse in Ontario 2022 and 2025 and Newfoundland and Labrador 2021, where incumbents held seats the provincial shift said they would lose. Given the actual provincial vote instead of the polls, the projected leader won 86.6% of ridings, so most misses come from how votes translate into seats rather than from polling error. The independent-candidate rule was estimated from cases that include these elections, so its part of the test is not independent. British Columbia is polled more often than most of the test provinces, so the error sizes probably err on the wide side.

The polls are checked several times a day. A new projection is made whenever the set of polls changes: a new poll, or a correction to one already included. Each projection is dated by the end of fieldwork of its newest poll; when several are made for the same date, the last one is kept. Every projection keeps its own page, as published. The earlier projections in the series, from before projections were first published, were computed retroactively: from the polls published up to each date, but with the model and candidate list of October 2026. Their pages say so.

No projection is made after 11:59 p.m. Pacific time on 23 October 2026. The last one before then is the final projection; none is published on or after voting day. After the election the final projection is compared with the result, seat by seat and riding by riding.

• It is only as good as the polls. If polls share an error, as they did in British Columbia in 2013, the projection shares it; the simulations allow for errors of the size seen in the backtests, not larger ones.
• Candidates are not modelled beyond who is running and the one independent rule. A well-known local candidate, a retiring member or a strong local campaign can move a riding further than the provincial shift suggests.
• Where OneBC and CentreBC voters live is assumed from where the Conservatives were strong in 2024. Their share in a given riding is therefore uncertain beyond what the simulations capture.
• The transfers for parties without a candidate and for defunct parties come from small recall questions in a few surveys.
• Regional poll results are few, entered by hand, and drawn from the pollsters’ own regional definitions, which differ from the model’s.
• Turnout is not modelled separately: the polls’ vote shares are assumed to describe those who vote.

The federal seat projection estimates how many of the 343 seats each party would win if an election were held at the time of the polls. No election has been called (the fixed date is 15 October 2029, earlier if Parliament is dissolved), so each projection describes opinion when its polls were taken; it is not a forecast of the next election. It follows the British Columbia method, with the differences set out here. A new projection is made whenever the set of national or regional polls changes, and each date keeps its own page.

Data

• National polls published since the election of 28 April 2025, as used by the site’s federal polling average. The average covers the Liberals, Conservatives, NDP, Bloc Québécois, Greens and People’s Party; others and independents share the remainder.
• The 2025 results in each of the 343 ridings (2023 Representation Order), from Elections Canada, as shares of valid votes; for a riding where a by-election has changed the party since, the by-election result (below).
• Regional figures: polls of a single province, and the regional results that national pollsters publish in their own releases. A release’s regional figures are used only if, weighted together, they reproduce the same poll’s national figures within 2 points, and each region’s shares sum to 100 within 3 points. Each figure is traced to the release it came from.
• Members of Parliament: the 2025 winners, replaced by by-election winners, with the members who have changed party since. Until retirements are announced, each member is presumed to run again in the same riding.
• French mother tongue by riding, from the 2021 census, for Quebec.

The regional model

Support does not move evenly across the country, so the projection works in seven regions: Atlantic Canada, Quebec, Ontario, Manitoba and Saskatchewan, Alberta, British Columbia, and the territories. It estimates each party’s share in each region in three steps.

1. National average. The polling average on the date of the newest poll gives each party’s change since 2025.
2. Regional rule. Each region starts from its 2025 result, moved by the national change: half as a uniform change in points and half in proportion to the party’s 2025 share in the region. The rule was chosen on the regional results of the six pairs of consecutive federal elections from 2006 to 2025, each pair predicted by a rule fitted on the other five. The root mean square error of regional shares was 4.48 points for this blend, 4.52 for a uniform change and 4.71 for a proportional one.
3. Correction by regional polls. Each regional figure is compared with what the rule gave for its region on the poll’s own date, so an older poll contributes its difference from the national trend rather than its level. The differences for each area are averaged with weights like those of the national average (a 90-day bandwidth; figures more than a year old are not used), and combined with the rule by generalised least squares: a region’s estimate moves toward its polls in proportion to their precision, allowing for sampling error, disagreement between polls, error common to all polls, and age. A poll of part of a region (Manitoba alone) or of several regions together (the Prairies) counts for the regions it covers, weighted by population. Regions without recent polls keep the rule.

The regional estimates are then adjusted together, more where they are less certain, until, weighted by votes, they reproduce the national average. Within each region every riding moves from its 2025 result as in British Columbia (half uniform, half proportional), and the ridings are adjusted so the region reproduces its estimate. The Bloc Québécois runs only in Quebec: its Quebec share is the national figure divided by Quebec’s share of the 2025 national vote, unless a Quebec poll at most 60 days old exists, in which case the Quebec polls set it.

Incumbents, language and floor-crossers

Incumbency. The party holding a riding gains 4.5 points when its member is running for a second term (the previous result does not yet include that member’s personal vote), and loses 4.0 points when the seat is open; other members running again get no adjustment. The points are added before the riding is rescaled to 100 and the region recalibrated, so the net effect is smaller: about 2.5 and 2.2 points for a party at 45%. Both sizes were estimated on the backtest elections; estimated with each election left out they ranged from 3.5 to 5.5 points and from 3.5 to 5.0, and both effects had the same sign in all six elections. Members who left their party get no effect. While no retirements are announced, there are no open seats, so the projection gives incumbent parties these points in every riding where a first-term member sits.

Quebec language term. In Quebec, each party’s change in a riding is increased by 0.6 × (the riding’s share of French mother tongue minus the Quebec average) × (the party’s change across Quebec). Mostly francophone ridings therefore move further in the direction of the province-wide change, as they did in the large Quebec swings of 2011, 2015 and 2019. The coefficient was estimated on the backtests (0.5 to 0.7 with each election left out). It reduced the error of Quebec riding shares in the backtests and made little difference to the overall scores.

Floor-crossers. Five members elected in 2025 have crossed to the Liberals and are presumed to run for them. In each of their ridings a fraction of the member’s 2025 share moves from the old party to the Liberals: a median of 10%, drawn anew in each simulation (90% range 2 to 44%). The fraction was estimated from the 16 federal members since 2000 who crossed the floor and ran again for their new party; individual cases ranged from no gain to about 80%.

By-elections. Where a by-election changed the riding’s party, its result, moved back to the regional levels of 2025, replaces the 2025 result. Riding errors are 20% larger in every riding that has held a by-election.

Uncertainty

Each of the 10,000 simulations adds a national polling error; an error for each region, the size of the uncertainty of its estimate (largest in Quebec and the territories); an error for each of 26 sub-regions (the provinces of Atlantic Canada and of Manitoba and Saskatchewan, and groups of neighbouring ridings in Ontario, Quebec, British Columbia and Alberta); and an error for each riding. All have Student’s t tails with 4 degrees of freedom. For a party at 50%, the national error is about 5.6 points, the sub-regional error 3.7 and the riding error 6.4; the regional rule’s error is 6.6 points before it is scaled for each region and reduced where regional polls exist. The sizes were estimated from the backtests and then enlarged (by 1.75 for the national error and 1.2 for the others): of the settings whose 90% ranges contained at least 90% of riding shares and party seat totals in the other elections, the one with the best Brier score. A majority is 172 of the 343 seats; the categories (Solid, Likely, Leans, Toss-up) are as above.

Backtests

The method was tested on the 6 federal general elections from 2008 to 2025, 1,973 riding contests: 2008 from 2006, 2011 from 2008, 2015 from Elections Canada’s transposition of the 2011 votes onto the 2013 boundaries, 2019 from 2015, 2021 from 2019, and 2025 from the transposition of 2021 onto the 2023 boundaries. Each election was projected from the polls published before it, with every error size, adjustment and enlargement fitted on the other five elections, so the scores are held out. Regional figures were available for every election except 2008. The Brier score is the mean, over ridings, of the squared differences between each party’s chance of winning and the outcome (1 or 0), summed over parties; lower is better.

The weakest results are 2011 and 2015, when the late surges of the NDP and the Liberals were larger than the final polls showed. In 2011 half of the party seat totals, and in 2015 one in five, fell outside their 90% ranges, so seat totals were covered 87.9% of the time rather than 90%. Given the actual regional results instead of the polls, the projected leader won 88.1% of ridings (Brier score 0.181), so most misses come from the polls and the regional estimates rather than from the translation of regional votes into seats. Pollster house effects, a correction for the average past error of the polls, a population-density term and separate language terms for each party were also tested and not adopted, because they did not improve the held-out scores.

Limitations

• With no election called, the projection estimates opinion at the time of the polls. A campaign can move support well beyond the ranges shown, and the projection does not attempt to anticipate one.
• Errors that the polls share are not corrected. In each of the seven elections from 2006 to 2025 the final polling average understated the Conservative share, by 0.3 to 3.2 points; a correction for this was tested and made the backtests worse.
• Regional figures are uneven: some regions, such as Atlantic Canada and the territories, are rarely polled on their own and follow the rule alone, and pollsters’ regions do not always match the model’s. Quebec’s sub-regions (Montreal, Quebec City, the rest) are not estimated separately, because no source publishes them regularly.
• Candidates are not modelled beyond incumbency, by-elections and floor-crossers. Until retirements and nominations are known, every member is presumed to run again, which favours the parties that hold seats. The floor-crosser fraction rests on 16 cases.
• Six elections are a small sample for sizing errors, and two of them (2011 and 2015) dominate the misses. Turnout is not modelled separately.

The Ontario seat projection estimates how many of the 124 seats each party would win if an election were held at the time of the polls. No election has been called, and Ontario no longer has a fixed election date: the Plan to Protect Ontario Act (Budget Measures), 2025 (No. 2) repealed it, and a legislature now lasts at most five years from its polling day, so the one elected on 27 February 2025 must be dissolved by 27 February 2030. Each projection describes opinion when its polls were taken. The projection starts from the 2025 results in the 124 ridings, whose boundaries have not changed since 2018, and differs from the British Columbia method in its swing and in allowing for incumbents.

• Why. Tested on the 2022 and 2025 Ontario elections, the British Columbia method’s projected leader won 82.3% of ridings each time, while keeping every riding with the party that won it last time won 88.7% and 92.7%. Given the actual province-wide result, it still did worse: the miss lay in how the change was shared out among ridings. Members running again held seats that a province-wide change said they would lose.
• Swing. Each party’s change since 2025 is applied as the same number of points in every riding (a uniform swing), not half in proportion to its 2025 share.
• Incumbents. Before the ridings are adjusted to reproduce the polling average, a party gains 6 points where its member runs again and 8 more where that member was first elected in 2025 or at a by-election since (their personal vote is not yet in the 2025 result), and loses 2 where its member is not running. The points are added before each riding is brought back to 100%, so a party at 40% gains a little over half of them, and the adjustment to the average then takes them from the party’s other ridings. Every member is presumed to run again until a retirement is announced. The two independents, Bobbi Ann Brady (elected as an independent in 2025) and Chris Scott (elected for the Progressive Conservatives), keep part of their own 2025 vote, as independents do in the British Columbia method.
• Tests. The swing and the three sizes were fitted on ten provincial elections from 2020 to 2025 (Ontario, Alberta, Manitoba, Nova Scotia, Quebec, Prince Edward Island, Newfoundland and Labrador twice, and British Columbia). Scoring each Ontario election with sizes fitted on the other nine, the projected leader won 88.7% of ridings in 2022 and 90.3% in 2025, against 88.7% and 92.7% for the last winner; the Brier score was 0.211 and 0.150, against 0.226 and 0.145 (lower is better). The same rules made the Manitoba, Nova Scotia and Quebec tests worse, so they are used for Ontario only. Who held office before 2025 comes from the 2022 results and the by-elections since; first-term members could be identified in five of the ten tests.
• Uncertainty. The province-wide polling error is 1.5 times its estimated size rather than 1.25 times, so that the 90% ranges of the new method still contained the actual seat totals at least 90% of the time in the tests. Regional and riding errors are as in British Columbia, over seven regions: Toronto, the 905 belt, Hamilton-Niagara, Southwestern, Eastern, Central and Northern Ontario.

The Alberta seat projection estimates how many of the 87 seats each party would win if an election were held at the time of the polls. No election has been called (the fixed date is 18 October 2027, earlier if the legislature is dissolved), so each projection describes opinion when its polls were taken. It uses the British Columbia method, with the differences set out here.

• Ridings. The projection uses the 87 ridings of the 2023 election and their 2023 results. The legislature is redrawing the ridings before the next election; when the new map is enacted, the 2023 results will be recast onto it and the projection will move to the new ridings.
• Parties. The polling average covers the UCP and the NDP; all other parties share the remainder, as most Alberta pollsters report them together. Those parties, among them the Progressive Tory Party (the Alberta Party, renamed in December 2025), the Republican Party of Alberta and the Wildrose Loyalty Coalition, had 3.3% of the vote in 2023. Their gain since is added evenly across the ridings, because Research Co.’s poll of 24 to 26 August 2026 puts them at 10% in the Calgary metropolitan area and 11% in the Edmonton area and the rest of the province. It is taken from the UCP and the NDP in the proportions that poll reports among 2023 voters: 9% of 2023 UCP voters and 8% of 2023 NDP voters now choose one of them, so 57% of the gain comes from the UCP and 43% from the NDP. A sensitivity run places it where the UCP vote is instead.
• Members who changed party. Peter Guthrie, elected for the UCP in Airdrie-Cochrane in 2023, now leads the Progressive Tory Party. He is treated like a member who left their party and ran again (39% of his 2023 vote carried forward, drawn at random around that median in each simulation). Five by-elections have been held since 2023, the last in Calgary-Shaw, which the NDP gained from the UCP on 14 September 2026; those ridings start from their 2023 result.
• Regions. The regional errors use three regions, as pollsters report Alberta: the Calgary and Edmonton census metropolitan areas and the rest of the province. Each riding is in the area holding at least half of its 2021 population, from Statistics Canada’s 2021 dissemination areas, which gives 29 ridings in each.

The Quebec seat projection for the general election of 5 October 2026 was computed after that election, as the projection would have run from the 2022 election: one projection for each day with a new poll, each from only the polls published up to that day, the latest from the polls published before voting day. It uses the British Columbia method, with the candidates who stood in 2026 (every party ran in all 127 ridings).

• The 2022 results on the 2026 map. The 2026 election used 127 ridings, two more than in 2022, and new boundaries for many others. The site found no official recast of the 2022 results onto the new map, so it made one from the 2022 results by polling section (the site’s maps, from Élections Québec): each section’s election-day votes go to the 2026 riding that holds it, a section cut by a new boundary is divided by area (such sections cast 2% of election-day votes), and each 2022 riding’s advance, institutional and other votes go to the 2026 ridings in the proportions its election-day votes went. Every 2022 vote is counted once. On the new map the 2022 result gives the Coalition Avenir Québec 92 seats, the Liberals 21, Québec solidaire 11 and the Parti Québécois 3. These are estimates, not results.
• Regions. The regional errors use three regions, as pollsters report Quebec: the Montreal and Quebec City census metropolitan areas and the rest of the province. Each riding is in the area holding at least half of its 2021 population, which gives 60 ridings in Montreal, 11 in Quebec City and 56 elsewhere.
• No incumbency allowances. The Ontario projection’s allowances for members running again made the Quebec 2022 test worse, so they are not used.
• The result. The latest projection, from polls to 4 October, put the Parti Québécois at 52 seats (90% range 26 to 74), the Liberals at 35 (27 to 46), the Coalition Avenir Québec at 16 (2 to 41), the Conservatives at 13 (6 to 26) and Québec solidaire at 7 (2 to 12). Élections Québec’s final results gave 59, 40, 0, 19 and 9: four of the five inside their 90% ranges, the Coalition Avenir Québec below its range. The projected leader won 105 of the 127 ridings (82.7%); keeping each riding with the party that won it in 2022 would have won 26%.