How we calculate
who's going to win

Every poll published by major firms (Léger, Abacus, Nanos, Mainstreet, EKOS, Angus Reid, Pallas, Innovative…) is entered with its field dates, sample size, firm, collection mode and regional breakdowns when available. Federally: 240+ polls aggregated. In Quebec: 60+ provincial polls, with region and language breakdowns when published.

Recent polls count more: time weighting follows a 14-day half-life — a two-week-old poll counts half as much as one from today. When the model detects a wave (fast, coherent movement across several firms), the half-life shortens automatically, by up to half, to follow the movement without over-smoothing it.

Before any weighting, a poll-quality module audits each poll's methodology individually — not the firm's reputation, the poll itself: declared weighting variables (age, gender, region, education, language, past vote), the source of weighting targets (which census), presence of regional breakdowns, and the effective sample size once the design effect is accounted for.

Each check produces an auditable flag and a quality score that modulates the poll's weight, with a conservative floor: no poll is discarded, but a national topline without breakdowns weighs less than a complete study. Older polls, published before these fields were captured, get neutral values — missing information is not treated as a methodological fault.

This is the first building block of our data integrity initiative, which strictly separates three questions: is this poll methodologically sound (quality), does this firm historically hit the mark (performance), and are there anomaly signals in the data (integrity).

Each poll's final weight combines several documented multiplicative factors: the firm's methodological transparency (disclosure score), its historical performance (mean absolute error at past elections: a firm averaging 1 point of error gets a bonus, a firm at 5 points is heavily reduced), a penalty for polls commissioned by partisan clients, and sample size on a logarithmic scale — going from 1,000 to 2,000 respondents helps, but not twice as much.

Then comes the house effects correction: for each firm, the model measures the systematic gap between its numbers and the cross-firm consensus over the same period. The correction uses adaptive shrinkage: a small gap (compatible with sampling noise) is barely corrected; a large, persistent gap is corrected at full strength. We don't punish a firm for being different — we correct what is demonstrably systematic.

Finally, historically poorly calibrated firms aren't just down-weighted: the uncertainty of their observations is widened inside the model, so that doubt propagates all the way to the final confidence intervals.

The model's core is a Bayesian Structural Time Series (BSTS), fitted separately for each party. Vote intention is treated as a hidden state evolving over time — a level and a local trend — observed imperfectly through the weighted polls. The model picks its own structure based on data richness: a cycle rich in recent polls allows a more reactive trend; a sparse cycle forces a more cautious structure.

Estimation runs by MCMC (Markov chain Monte Carlo): thousands of successive draws explore the full set of trajectories compatible with the data. The result isn't one number, it's a complete posterior distribution for each party at each date — “the Liberals are probably between 41% and 46%” — and every draw keeps cross-party coherence (the shares have to stay plausible together, not just separately).

A national average doesn't produce seats. The Monte Carlo simulation isn't what sets the model apart — it's simply the translation tool: once intentions are estimated and corrected, we replay the full election tens of thousands of times — 50,000 simulations federally, 20,000 in Quebec. Each simulation draws a plausible national result from the BSTS distributions, respecting the cross-party correlation matrix estimated from the data: when one party rises, the model knows whom it is taking votes from.

That national result is then cascaded region by region using swing elasticities: each region and riding type (urban, rural) reacts differently to the same national movement, calibrated on past elections. Regional swing anchors on the polls' regional breakdowns when they exist, on history otherwise. Federally, by-election results serve as an extra local anchor. Riding-level noise completes the picture — because every riding has its surprises.

Published numbers are statistics over these simulations: the seat projection is the average, the 80% interval covers the central scenarios, and “majority probability: 90%” literally means the party wins a majority in 9 out of 10 simulations.

After every election, firm profiles are updated: measured biases, mean absolute error, confidence. These profiles feed directly into the next cycle's weighting and corrections — the model learns from every vote.

The model is also replayed on past elections using only the polls available at the time (backtests): the 2019, 2021 and 2025 federal elections, the 2024 UK general election, the 2017 and 2022 French presidential elections, and the 2022 and 2024 French legislative elections. The observed errors — by party, in seats — are published on our Track Record page. Predicting the past is easier than predicting the future; that is exactly why we impose the exercise on ourselves and show the results.

Every run exports public static JSON — the same files that power these pages. Each projection shows its latest run date, with the number of polls and simulations used. As election day approaches, recent polls carry more weight and the intervals tighten naturally.