Polls: Thermometer or Weapon? What They Measure, How Wrong They Get and How to Read a Headline

A poll is both a measure of the public mood and a political message. Both are true, and the useful question is not which one but how much of each. Using the five general elections from 2015 to 2023 and the 526 polls of this legislature, we measured it: the average of last-week polls missed by 0.7 to 2.4 points per party; each pollster has its own stable bias, of up to 1.5 points for or against a party; and seat headlines exaggerate what polls know. Polls are useful. What isn't useful is reading a single one, or reading the seats.

31 October 2026 · Miguel · Leer en español

Priors or posteriors: both

In statistical language, a prior is what you believe before seeing the data, and a posterior is what you believe afterward. The question of whether polls are one or the other has an honest answer: they are both, in proportions that can be measured.

As a posterior, a poll is a measure of the electorate’s mood on a given date, with its sampling error and the bias of whoever conducts it. It is real information: in every general election since 2015, the poll average got the order of the parties and the size of the shifts right.

As a prior, a published poll influences what it measures. It does so through three channels the literature knows well: the bandwagon effect, whereby some undecided voters join whoever is rising; tactical voting, whereby supporters of small parties switch to a large one when a poll says their vote doesn’t count; and demobilization, when a result looks settled. All three effects exist, are small and are concentrated among the undecided and small parties. In an electorate like Spain’s, with strong partisan voting, they move tenths of a point, not points.

Then there is political use, which is something else. A party that leaks an internal poll, an outlet that headlines seats rather than votes, a pollster that publishes only what favors its client: none of that changes what the poll measures, but it does change what the reader thinks it measures. That is where the weapon lies. There is a simple defense against it: look at many polls, correct each pollster’s bias and don’t read the seats.

How wrong they get

In the five general elections from 2015 to 2023, we compared the average of the polls from the last two weeks with each party’s actual result, for parties above 3%:

The average error per party was 2.0 points in 2015, 2.4 in 2016, 1.2 in April 2019, 0.7 in November 2019 and 1.5 in 2023. There are two lessons:

• The error is not random. In 2016 the polls underestimated the PP by 3.6 points; in 2023 they underestimated the PSOE by 3.4 and overestimated the PP. This is the industry-wide error: in any given election, all pollsters tend to miss in the same direction, because they share methods, samples and assumptions about who will vote. That is why our forecast includes an error common to all polls, calibrated with those five errors, and why its intervals are wider than a single poll’s sampling error suggests.
• The error shrinks over the campaign. Last-week polls were more accurate than those from a month earlier in all five elections. That is why, since October 6, our aggregator gives more weight to recent polls and forgets older ones sooner.

Each pollster’s bias

Each polling firm has its own deviation from the average of the others, which repeats from poll to poll. Estimated with the 526 polls of this legislature and corrected with its 2015–2023 track record:

The deviations of private pollsters are small, between half a point and a point and a half, and run in both directions: NC Report, InvyMark and DYM give the PP more than the average; GESOP and 40dB, less. The CIS is a case apart, with a direct vote intention that gives the PP 5 points less and the PSOE 3 more than the rest, for reasons we analyze in “The CIS under scrutiny.” The point is not to single anyone out: a stable bias is correctable, and our aggregator corrects it. The bias that can’t be corrected is the one common to all, the industry’s, and only the election reveals that.

Seats: where headlines detach from the poll

The most misleading figure in a poll headline is not the percentage but the seat count. A poll of 1,000 or 2,000 interviews measures the national vote with a margin of about two points. Translating that into seats requires distributing it across 52 provinces, and there each pollster applies its own method, which is not published and which usually ignores uncertainty. The result is a figure with the look of precision, “PP 135, PSOE 103,” that hides the fact that the last 20 or 30 seats are up in the air.

We saw this two days ago with a concrete example. Polls published after the Tempi accident in Greece and the Lisbon funicular accident measured drops of 3 to 5 points that the ballot box did not confirm. Polls overestimate recent change, especially in their seats, because the allocation model amplifies every tenth of a point of vote.

That is why our forecast uses no pollster’s seats, only their percentages, and distributes the votes with 50,000 simulations that carry within them sampling error, pollster error, industry-wide error and territorial error.

How to read a headline

• Look at the percentage, not the seats. And compare it with the same pollster’s previous poll, not with another pollster’s.
• Look at the sample size and the dates. A poll of 1,000 interviews has a margin of ±3 points for the big parties. One of 3,000 with three weeks of fieldwork mixes different moments.
• Distrust polls that publish only some parties, or only seats without percentages, or that are commissioned by the party itself. Our aggregator no longer uses them.
• A change of a tenth of a point is not a change. Neither is one of a point in a single poll. It is one when several pollsters measure it in the same direction.
• Read the average, not the poll. And read the interval, not the midpoint.

So, thermometer or weapon?

Mostly thermometer. Polls measure, and they measure reasonably well: between one and two points of error per party in the last five general elections, with small, correctable pollster biases. They are a weapon only to the extent that they are published without context, with seats instead of votes and without saying who pays for them. The antidote is not to stop looking at them, but to look at all of them, correct them and measure their uncertainty. That is what we try to do here every morning.

How we calculated it

• Historical errors. Simple average of polls with fieldwork in the 14 days before each election (excluding the CIS), against the official result, for parties with more than 3% of the vote (m5.error_historico, out/m5/error_historico.csv).
• Pollster effects. Deviation of each pollster from the aggregator’s state in logit, shrunk toward its historical deviation relative to the industry in the 2015–2023 general elections (out/m5/casas.csv, data/casas_sesgo_historico.csv), converted to points using each party’s average share. Pollsters with eight or more polls in this legislature.
• Excluded polls. Since October 26, polls commissioned or released by a party, and those that do not publish PP, PSOE and Vox together, are not included.
• Code. m5.py, casas_historico.py and backtest_campana.py, in the altnalysis repository.