Method

Each model produces three results: the trend in voting intentions, the likely distribution of seats and the uncertainty of that distribution. The detailed methodology of each model, with its validation figures, is given at the bottom of its page (Quebec, British Columbia, Canada).

Data

Voting intentions come from the polls released by polling firms, collected from Wikipedia tables and from the pollsters' reports. Results come from the electoral agencies (Élections Québec, Elections BC, Elections Canada) and riding profiles from the Statistics Canada census. The tables, with the source of every row, are published in the votovision/data repository.

Trend in voting intentions

Party shares form a composition: they are positive and sum to 100%. They are transformed into isometric log-ratios, then smoothed over time by a Gaussian process that weights each poll by its sample size. The hyperparameters (time scale, noise, breaks at leadership changes, acceleration during campaigns) are estimated from the data.

From voting intentions to seats

Each riding starts from its result at the last general election. The provincial or regional swing estimated from the polls is carried over to it, with adjustments whose effect was measured on past elections: language composition and regional poll breakdowns in Quebec, vote transfers between parties according to their local strength at the federal level, vote transfers according to where each party is running in British Columbia. An adjustment is kept only if it improves the forecast of elections that were not used to estimate it.

Uncertainty

The seat distribution is obtained by simulation: each draw combines the uncertainty of the trend, a common polling error, regional deviations and a riding-specific deviation drawn from the errors observed at past elections. The reported seat count is the median scenario of the joint distribution (medoid); intervals contain 90% of the draws. Their coverage is checked on the eve of past elections, each one evaluated with parameters estimated without it.

Limitations

A forecast is a probability distribution for election day, conditional on the polls published and the results known on the date of computation. Its uncertainty includes polling error and the change in opinion until election day, both estimated on past elections. It does not account for events the polls do not yet measure; far from an election, it mostly describes the state of opinion.