How the polling averages work

Polls are compiled from the opinion polling tables on Wikipedia for each election, which list published polls with their fieldwork dates, pollster, sample size and results. The tables are read again every morning, so a poll usually appears here the day after it is added there.

Only polls of the whole electorate count. Regional subsamples, seat projections and hypothetical questions (for example, how people would vote under a different leader) are left out. Each poll is dated by the end of its fieldwork. Results are used as published; the site doesn’t reallocate undecided voters.

After an election, the count starts again: only polls taken since the most recent election go into the average. Until new polls arrive, the landing page shows the run-up to that election and says no polls have been published since.

The average on the landing page, and the summary sentence under each polling chart, is a weighted average of recent polls, recalculated for each week and on the date of the latest poll. A poll’s weight depends on three things:

• How recent it is. Weight falls off smoothly with time, following a bell curve with a standard deviation of 21 days. A poll three weeks old counts about 60% as much as one taken today; one six weeks old, about 14%.
• Its sample size. Weight grows with the square root of the sample, which is how sampling error shrinks. Samples above 3,000 count as 3,000, so a very large online panel can’t dominate, and a poll with no published sample counts as 600.
• How often its pollster publishes. A pollster with several polls in the same six-week window shares weight between them (each counts 1 divided by the square root of the number), so a firm that publishes weekly doesn’t outweigh one that publishes monthly.

Where a party has no polls for more than 90 days in a row, no average is drawn across that gap. When there have been no polls at all for 120 days, the site says so instead of reporting a lead.

The lead is the difference between the two leading parties’ averages, each rounded to the nearest point. When the rounded averages are equal, the race is described as tied.

The band around each line on the landing page is one weighted standard deviation of recent polls either side of the average. It shows how much the polls disagree with each other. It is not a margin of error or a confidence interval, and it doesn’t capture errors that all polls share.

The full polling chart on each election page plots every poll and draws a trend line for each party using LOESS, a local regression: at each date, nearby polls are fitted with a straight line, weighted so that closer polls count more. The window adapts to how many polls there are, from 8% to 40% of the period on either side. Polls more than 120 days apart start a new line, and a party needs at least five polls for a line to be drawn.

The trend lines don’t weight by sample size or pollster, so they can differ slightly from the average on the landing page, especially when one pollster publishes much more often than the rest.

The averages aren’t forecasts or seat projections. They don’t adjust for differences between pollsters (house effects), survey methods (online panels, phone or interactive voice response), likely-voter screens or regional quirks, and they treat every poll in the tables as equally reliable apart from the weights above. Polls have missed Canadian results by several points before, as in the 2012 Alberta and 2013 British Columbia elections.

The compiled polls from the past 15 months, for every jurisdiction on the landing page, are published as a JSON file, updated every morning. Each poll lists its end date, pollster, sample size and results. Spotted a missing or wrong poll? The tables on Wikipedia are the source, and corrections there reach this site the next morning.