Do Polls Err Toward Whoever Pays for Them? Five General Elections Under Review (Part One)

We compared the 391 polls taken in the four weeks before the five general elections from 2015 to 2023 with the actual results. The biggest error is shared by the whole industry: in 2023 it overstated the lead of the conservative People's Party (PP) over the Socialists (PSOE) by 4.4 points. Once that common error is removed, some pollsters and outlets lean the same way almost every time. The polls of La Razón and Okdiario gave the PP a bigger lead than average in every election; those of El Periódico and of RTVE–FORTA gave it a smaller one.

22 October 2026 · Miguel · Leer en español

Every polling firm makes choices that do not show in the headline: how it weights the sample by recalled vote, what it does with undecided voters, and where it finds its respondents. Those choices can push its estimates the same way every time. This is called a house effect, and our aggregator estimates and corrects it for each firm. Today we put the numbers in plain view, with a question that comes up often in conversations about polls: does it show who pays?

This is part one. Here we look at national polls for general elections. In part two we will compare regional pollsters with national ones in regional elections, where regional firms are often said to do better.

The error they all share

First, the most important point: the error of any single poll has two parts, the one it shares with all the others in that election and its own. The first is almost always the larger.

In 2023 the polls gave the PP, on average, a lead over the PSOE 4.4 points larger than the one it got. In 2016 they erred the other way: they underestimated the PP by 3.6 points and overestimated Unidos Podemos by 3.4. The common error changes sign from one election to the next, and no polling average removes it, because all the polls miss at once. That is why our model adds a wide-tailed industry error on top of each firm’s house effect.

With the common error removed: the firms

To see whether a firm deviates on its own, we subtract the industry’s error in that same election from its error. A +2 means that firm gave the PP a lead over the PSOE two points larger than the average of the others.

Firms with polls in at least three of the five general elections. “Average error relative to the industry”: mean absolute error across the five main parties, minus the industry’s in the same election; negative is better than average.

Two things stand out. First: NC Report gave the PP a larger lead than average in all five elections, with an average deviation of 2.5 points. It is the only firm with five of five in the same direction. On the other side, IMOP did so in three of three and GESOP in four of five.

Second: leaning one way is not the same as being more wrong. NC Report’s average error is practically the industry’s. When everyone overestimated the PP (2019 and 2023), its lean took it further from the result; when everyone underestimated it (2016), its lean brought it closer. In 2023, 40dB, the firm in the table that moved furthest from the average toward the PSOE, was also the one that came closest to the PP’s actual lead.

And the outlets

Almost every outlet commissions its polls from the same firm election after election (the newspaper La Razón always from NC Report; the newspaper ABC from GAD3; the newspaper El Mundo from Sigma Dos), so in most cases the outlet effect and the firm effect are the same thing. Two exceptions help tell them apart: Okdiario and RTVE–FORTA (the public broadcasters RTVE and FORTA, the latter the association of regional public broadcasters) changed firms and still kept the same deviation.

Okdiario commissioned its polls from three different firms in three elections, and all three gave the PP a larger lead than average: 2.1, 2.2 and 3.0 points. RTVE–FORTA went through three firms in five elections and its polls always came out somewhat below average on the PP’s lead, though only slightly (1.1 points on average). In the other cases there is no way to know whether the deviation comes from the firm or from the commission.

With these data we cannot say why this happens. It may come from methodological decisions that the outlet requests or that the firm applies to all its clients. It may also be selection: an outlet decides which polls to publish and when. And there may be chance. With five elections, a given firm deviating the same way all five times by pure luck happens once in sixteen, and here we are looking at more than fifteen firms and as many outlets. That some look consistent by chance is to be expected. That La Razón (five of five), Okdiario (three of three, with three different firms) and El Periódico (four of four) do is harder to put down to chance, but it is not proof either.

Is it still happening in 2026?

For the current cycle there is no result to compare against, but we can measure how far each firm sits from the others on the same dates. We compared each poll published since January with the average of the other firms’ polls around its fieldwork date.

Firms that deviated in 2023 tend to deviate the same way now: among the eight with polls in both cycles, the correlation between their 2023 deviation and their 2026 one is 0.63. NC Report was 2.8 points above the average in 2023 and is 3.1 above it now. But departing from the others is not the same as being wrong: on 29 November it may be the others that miss, as in 2016.

Are regional pollsters more reliable?

In general elections two regional firms have polls in all five: Sondaxe, of La Voz de Galicia, and GESOP, from Barcelona. Both err slightly less than the industry average (0.1 and 0.2 points less in average error), but they do not stand out as unusual. The idea that regional firms are more accurate is usually applied to regional elections, and testing it requires regional polls we have not yet collected. That will be part two.

What the model does with all this

Our aggregator discards no poll. It estimates each firm’s effect from its track record and subtracts it, and it also adds the industry’s common error, which weighs the most and which no average can remove. The details are in the methodology and the full list of polls is under polls.

How we calculated it

• Polls. Those compiled on Wikipedia for the general elections of 2015, 2016, April and November 2019, and 2023, with their firm and outlet, excluding the CIS (we cover it in its own article). Only those whose fieldwork ended within the 28 days before the vote: 391 polls from 20 to 25 firms per election.
• Error. Published estimate minus result, in points of the valid vote. The left of the PSOE groups Podemos, IU, Unidos or Unidas Podemos, Más País and Sumar, depending on the election. Each firm (or outlet) weighs the same in each election, using the average of its polls in the window.
• Industry error. The average of the firms in each election. A firm’s deviation: its error minus the industry’s in that election, averaged over the elections in which it published.
• 2026 cycle. Polls since 1 January 2026 (excluding the CIS). Each is compared with the weighted average of other firms’ polls using a 10-day Gaussian kernel around its fieldwork date. Firms with four or more polls are shown.
• Code. articulo_casas.py, in the altnalysis repository.