How our averages are built
A poll average is not a single poll but an estimate of voter sentiment drawn from all published surveys. A single poll has a sampling error of several percentage points, so we read the trend from many polls together.
1. Sources
We use voting-intention polls for national parliaments published by polling agencies and media, with sources stated. For every poll we record the date, the agency and, where known, the sample size. All data can be downloaded (CSV/JSON) next to each chart.
2. Trend (the line on the chart)
The trend is a smooth line obtained by Gaussian kernel regression: each poll influences the estimate for a given day the closer it is to that day. The kernel width adapts to the length of the period (roughly 1/28 of the total span, at least 10 days). Larger polls carry more weight (proportional to the square root of the sample size, limited to a 0.5–2× range).
3. Uncertainty range
The shaded band around the line shows an approximate range (≈95%) that takes into account the spread of polls and their effective number. It is an indicator, not a formal confidence interval: actual election results can deviate more, especially long before an election.
4. Pollster deviations
Some agencies consistently give a party more or less than the others (“house effect”). We estimate it as the average difference of each poll from the trend of the other polls, shrunk towards zero when an agency has few polls. In the chart you can switch on an adjustment for these deviations.
5. Numbers next to parties
The percentage next to a party is the weighted average of polls from the last three weeks (counted from the latest poll), and the arrow shows the difference from the previous three weeks.
6. Seat projection
Where the number of seats and the electoral threshold are entered, we show an illustrative D’Hondt allocation of seats from the poll average. It ignores constituencies, bonus seats, minority rights and parties not shown; it is a simplification, not a forecast.
Limitations
Polls are not forecasts. They do not measure turnout, late voter decisions or systematic errors shared by all agencies.