Methodology

Our method, in plain terms

What the simulator calculates, where its numbers come from, and what it does not claim to know.

The simulator does not guess: it transposes, swings, corrects and simulates. Every number that enters the calculation carries one of three statuses, shown throughout this page:

Measured observed in official data or published reports Calibrated tuned against real elections Judgment documented, adjustable analyst choice

1. The foundation: official data, at the finest grain

Everything starts with the Chief Electoral Officer of Québec: the official results of the 2014, 2018 and 2022 general elections; the detailed results by polling station for the 125 electoral divisions of 2022, which our automated checks validate at 125 out of 125; the official geometry of the polling divisions, in its two boundary sets — the 2022 set (17,924 polygons), which carries the results, and the July 15, 2026 set (16,756 polling divisions, of which 15,531 are drawn on the map — the other 1,225 are institutional polling divisions, without territory), which is the grain of the product; the 2026 electoral map and its 127 ridings; and the 2021 census (age, language) by riding. Measured

The new 2026 map requires transposing the 2022 results: we first locate them polling division by polling division in the 2022 boundary set — 16,523 ordinary polling divisions — then each polling division allocates its votes to the 2026 polling divisions that overlap its territory, in proportion to the shared areas. The 2022 result shown for a 2026 polling division is therefore an estimate; the riding total, however, is exact. A polling division covers a few hundred electors (414 at the median): the grain is fine enough to follow real electors, not averages. And each 2026 polling division carries its actual number of electors registered as of July 15, 2026, published by Élections Québec — not an estimate. Measured Checks confirm this: the transposition keeps provincial totals within four votes out of 4.1 million, and recalculated riding by riding, it differs from the whole-polling-division method by at most 0.17 points (median: 0.00).

2. From polls to the 127 ridings

Published polls are aggregated with weights that are written down and dated — by firm, by type, and a weight that halves every 14 days Judgment — then broken down into targets for three regions (Montréal, Québec City, rest of Québec), defined exactly as polling firms weight their own samples. Measured The list of polls used in the current update — firm, field dates, sample, link to the source — is published continuously. Recognized firms weigh equally: none receives a privileged weight. What separates two polls is their age and their type. Judgment The calibration in section 4 tests the last Léger polls before the 2014, 2018 and 2022 elections against the official results: that is a test of the model, not a weight in the aggregate.

A proportional swing — chosen after comparing five swing methods across the 127 ridings — applies the regional shift in opinion to the 2022 local base. It is damped: the proportional component carries four-fifths of the movement, a uniform shift the remaining one-fifth — because replaying the method on 2014→2018 and 2018→2022 shows that a pure proportional swing stretches a party's strongholds and deserts too far, and that this damping is the only setting that reduces the error across both transitions. Measured Four families of corrections then refine the result:

• Age and language. The census gives the profile of each riding; polls publish voting intentions by age and language in 2026 as in 2022. We therefore correct for the measured shift of these gradients — never an assumed gradient. Measured
• Second choices. When a party loses ground, its points go first to the parties that its own electors name as their second choice, according to the matrix published with each poll. Measured
• The incumbent running again, and the leaders. An elected member's personal vote is already in their 2022 base — but replaying the method on 2014→2018 and 2018→2022 measures a 3.5-point gap between the party of an incumbent defending their seat and that of an incumbent who is leaving, stable across both transitions. The model therefore adds 1.5 points to the party of the incumbent who is running again in the riding they represented and subtracts 2 points from the party that loses its member, the party under which the incumbent was elected in 2022 even if they later sat as an independent (72 and 53 ridings as of August 23, 2026, kept up to date with the candidacies). The split between the two was chosen to add nothing on average. Party leaders, meanwhile, carry a bonus in their own riding — 5 points for a major party, 3 for a smaller one, and the reverse where a leader is stepping down — this is the strongest case of the candidate effect in the research, and a judgment inherited from our first version, adjustable. The effect of the local candidate is real but weak in Québec — decisive for 2 to 5% of electors. Measured Judgment
• By-elections. Five by-elections have been held since October 2022 (Saint-Henri–Sainte-Anne, Jean-Talon, Terrebonne, Arthabaska-L'Érable, Chicoutimi). A real, recent vote says more than a four-year-old result: in these ridings, the projection targets the by-election result anchored in the voting intentions of its own period, then carried forward by today's polls, blended with the standard projection since 2022 — two-thirds when the by-election is recent, one-third once eighteen months have passed: the information from a vote goes stale, and by-elections draw fewer voters, and different ones. The adjustment runs through the model's local bonuses, recalibrated at each registry update, and each affected riding page shows its by-election result. Measured Judgment

Regional rebasing

When an exceptional regional measurement allows it — a large pooled sample, such as the Léger poll Les grandes tendances régionales from August 25, 2026 (n = 11,305) — the model recalibrates its projections region by region. This measurement is used for its regional structure, never for its national level: a pooled run of waves lags behind recent waves, and each regional target is therefore rebased to the current national vote before any comparison. Measured

The correction follows a rule of caution: a region is revised only if the gap exceeds the margin of error of its sub-sample, and only by the excess beyond that margin — the noise from small samples does not pass through. Corrections are capped at ±8 points per party and per region; any magnitude beyond the cap is logged for review, not applied. Each correction is entered as a riding-level bonus, justified and logged, and the change log of the published version lists them. A poll of this type enters the registry with a weight of zero: it is shown, it is not counted twice. Judgment

3. Probabilities: simulating error, not chance

(The technical terms in this section — Monte Carlo, median, ranges, percentiles — are explained in plain language in the glossary.)

A projection says who is ahead; it does not say the odds. For that, we simulate 10,000 scenarios: each draws a plausible set of regional voting intentions — the size of the polling error is calibrated on three real elections Calibrated — and adds three layers of uncertainty to it: the noise specific to each riding, whose amplitude and correlations between parties are measured on 253 observations drawn from the 2014→2018 and 2018→2022 transitions Measured — the error in the swing itself, measured by replaying the method on past elections with perfect regional targets (a standard deviation of 5.2 points per riding and per party): even with exact polls, transposing a regional voting intention into a riding gets it wrong, and this error enters the distribution instead of being suppressed Measured — and a systemic error per party, drawn once per scenario and applied everywhere at once, calibrated on the gaps between the last polls and the actual results of the last three general elections: the documented risk that the margins of error of all polls lean the same way at once, entered into the distribution rather than left in a footnote. Calibrated Each scenario then passes through the same calculation engine as the published projection — not a copy, the same code — and the average of the scenarios reproduces the published projection.

A turnout correction by age, based on official turnout rates from Élections Québec, converts the population into a likely electorate Measured — applied to the projection shown as well as to the simulations. It applies to our aggregate; the "according to Léger" or "according to Synopsis" scenarios remain the figures published by each firm, untouched.

The count shown on the map and the simulated range do not say the same thing, and that is deliberate. The map counts each riding at its centre: it answers "who is ahead where, today." The simulations incorporate the measured local noise — and because a party running second in thirty ridings ends up winning somewhere in nearly every scenario, its simulated median can run well above its centre-point count. This is not an inconsistency: it is the difference between a readable scenario and an expectation under uncertainty. Both are published, each under its own label.

4. Calibration: three real elections as the only judges

Before publishing a first number, the model passed two tests. First, an out-of-sample back-test: tuned on the single 2014→2018 transition, it had to predict 2022 — an election it had never seen; the measured noise proved optimal, without adjustment. Calibrated Then, the hardest test: the last Léger poll before each of the three elections, against the official result.

This table contains the fact with the heaviest consequences for 2026: the PQ has been overestimated by its last poll in all three elections — three for three, from +1.7 to +2.9 points.

5. Who is accountable for the method

The simulator is designed and overseen by Stéphane Gobeil and Alexandre Gobeil-Hébert, co-founders of ASGO. Stéphane has taken part in the campaign direction of eleven federal and Quebec general elections: the parameters marked Judgment are his own — set, documented and reversible — and it is this field experience that led the model to explicitly address the variable that voting intentions do not measure: turnout. Alexandre specializes in integrating artificial intelligence within organizations. Development has drawn on Anthropic's Claude models, under their supervision.

6. What the model does not say

Our numbers are a conditional snapshot — "if the election were held today, and if the polls are unbiased" — not a campaign prediction. Not captured: an opinion shift that the polls do not yet see; the actual scale of voter mobilization in 2026; independent candidacies and any sixth party; and events between now and the election. Every published simulation can be replayed identically (the random seed and the parameters are archived with the result): a number that cannot be replayed is not a defensible number.

Reuse and citation

Our projections are made to circulate. Cite them freely, naming Route 127, and share our images as they are — they carry their source and their disclaimer. Journalists do not need to ask us for permission to cite a number or reproduce one of our images with credit; we gladly respond to interview or fact-checking requests before publication. What does require our consent is the systematic extraction of our output and its reuse in a commercial product: the detail is on the terms of use page.

Contains open data granted under the Chief Electoral Officer's Open Data User Licence, available online at dgeq.org/en. The fact of granting the licence in no way implies that the Chief Electoral Officer approves the use made of the open data. Poll sources: published reports from Léger-Québecor (2014, 2018, 2022, 2026), Synopsis Recherche–La Presse (2026), Pallas Data (2026); 2021 census via Élections Québec's socioeconomic profile.