BR

About Pollster Performance

The goal of the Pollster Performance series is not to analyze the quality of polls or of the pollsters that produce them. The idea is to quantify only how far the polls of the final week of each race were from the official election result, without speculating about the reasons or deciding which polls "got it right" or "got it wrong".

Analyzing quality, or identifying the possible causes of the gaps, would require every pollster to publish not only precise details of its methodology and the weights used in its samples, but also its complete data. Common practice in the industry is to keep data and weights as intellectual property, and therefore confidential, which is understandable: in the same way, we choose not to disclose every parameter behind our forecasting models, for example.

Despite these limitations, we believe in the value of the descriptive exercise that is Pollster Performance. On one hand, we take care not to focus only on comparing the results of the main candidates, but also on how each pollster mapped the election as a whole. On the other, we make clear that much of the variation in performance is noise: polls that miss by just one percentage point are not really distinguishable from those that miss by two.

Which polls are included

In each race we consider each pollster's last poll with fieldwork ending up to seven days before the election. Older polls are left out: the further from the vote, the more a poll describes an election that has not yet happened, so comparing it with the ballot box makes little or no sense. Using only one poll per pollster also keeps those who published more often in the final stretch from carrying more weight in the comparison.

The numbers come from each poll's original release, in the pollster's report or in a news report, identified by its registration with the Superior Electoral Court (TSE). At the moment 96% of the polls considered have a link to their source in the table itself. The rest come from historical archives, and their original release could not be found.

The series covers presidential elections since 2002, governor elections since 2014, Senate elections since 2018 and state-capital mayoral elections in 2016, 2020 and 2024. Runoffs appear for president (2006 to 2026), governor (2014 to 2026) and state-capital mayor (2016 to 2024).

Valid votes

The TSE publishes the final result in valid votes, excluding blank and null votes and abstentions; the comparison has to be made with that in mind. Many pollsters publish their own valid-vote calculation, and in those cases we use their numbers, since each pollster has its own way of handling undecided voters and voters who intend to spoil their ballot. When a poll publishes only total votes, we apply what we call the standard conversion: we remove blank, null and undecided responses and distribute the rest proportionally, as if voters who did not state a vote split the same way as those who did. It is an imperfect conversion, but it is the choice with the fewest theoretical assumptions for polls about which there is not enough information.

For Senate polls we use the same basis as our Aggregator: the share of all votes cast, without undecided voters. In years with two seats each voter votes twice, and the TSE records the result by adding all votes into shares that sum to 100 in each state.

Candidates are compared one by one with the poll's numbers; those the poll did not list separately form a single group ("Others"), which also enters the calculation. When a candidate's candidacy was annulled and the poll did not include them among the options, we compare the poll with the result without that candidate, redistributed among the others.

Weighted error

The term "error" here describes the difference between poll and result and should not be read as more than that. There are several ways to measure this number, and none is completely neutral. Our calculation starts from a simple premise: the error that matters is the one that changes how an election is read. Missing the vote of a candidate with only 2% is not the same as missing who goes to the runoff, for example.

The weighted error combines two factors and two adjustments, all measured in percentage points. The rules below apply to president, governor and mayor in the first round; the Senate and runoffs have adaptations, explained later. To show how the calculation works, we use the same example throughout: the official result had Candidate A with 45%, Candidate B with 35%, Candidate C with 12% and the others combined with 8%. A poll showed A with 48%, B with 33%, C with 13% and the others with 6%.

Decisive margins

The first factor is the error in the margins: the difference between 1st and 2nd place and, when a runoff spot was at stake, the difference between 2nd and 3rd. In the example, the poll showed A 15 points ahead of B, when the real gap was 10, an error of 5 points; it showed B 20 points ahead of C, when the gap was 23, an error of 3 points.

We measure margins rather than each candidate's share because the margin decides the election and is what the public reads first. A poll that underestimates the top two candidates by the same amount (for example, because it has more undecided voters) still described the race correctly.

The two margins do not weigh the same: the closer the race in the result, the greater the weight. A 3-point error in a race decided by 2 points changes who wins; in a race decided by 30 points, it changes nothing in the story of the election. The weight falls continuously as the real gap grows: for example, 0.88 for a gap of 2 points, 0.51 for 10 points, 0.26 for 20 points and 0.14 for 30 points. In the example above, the margin between A and B (10 points) weighs 0.51 and the margin between B and C (23 points) weighs 0.22, and the decisive-margins error comes to 4.4 points, closer to the error in the tighter margin.

The margin between 2nd and 3rd only counts when the runoff spot was actually in play. If the winner had less than 50% of the valid vote, it counts in full; above that, its weight falls quickly (0.82 with the winner at 50.5%; 0.37 at 52.5%; 0.14 at 55%). We prefer this gradual approach because in some races whether there was a runoff came down to tenths of a point: in 2026, in Paraná, the winner Sergio Moro had 50.1% of the valid vote.

The decisive margins are measured in percentage points of margin: when one voter switches from the 1st to the 2nd candidate, the margin between them changes by two points, while the misplaced electorate changes by only one. In practice, the decisive margins weigh about twice as much per voter, which is intentional: it is the error that changes how the election is read.

Misplaced electorate

The second factor measures the election as a whole: how much of the electorate the poll put in the wrong place. We add up the difference between poll and result for each candidate and for the group of others, and divide by two, because each misplaced voter is counted twice (once for the candidate who "gained" the vote and once for the one who "lost" it). Continuing with the example above, the differences are 3, 2, 1 and 2 points: a total of 8, or 4 points of misplaced electorate. In other words, 4% of respondents would have to change their vote for the poll to match the result exactly.

This factor exists because the margins only see the top of the race. A poll can get the margins between the top three right and still inflate or deflate the smaller candidates, which is also a way of describing the election badly.

In general, the two factors weigh the same. When the winner is close to 50% of the valid vote, the weight of the misplaced electorate rises to 75%, because in that case every point of the leader decides whether there will be a runoff. In other words, the size of each candidate matters as much as the margins. With the winner at 47%, the weight is 64%; at 50%, 75%; at 60%, it falls to 53%. In the example, with the winner at 45%, the weight is 59%, and combining the two factors (decisive margins and misplaced electorate) with these weights gives 4.2 points.

Wrong side

The first adjustment deals with errors that change the story of the election. We consider three "decision lines" in a first round: who leads, which two candidates go to the runoff, and whether the winner passes 50%. When the poll sees the electorate on the other side of a line that the result did not cross, we count how many points of electorate it placed past the line. For who leads, for example, that is half the inverted margin: the number of voters who would have to switch sides to undo the inversion.

This value is discounted by how far the result itself was from the line. It makes no sense to require a poll to get the right side of a line that the result crossed by a hair: that is within any reasonable margin of error. If the result was 0.25 points from the line, we charge only 22% of the value; at half a point, 39%; at one point, 63%; at two points, 86%; from three points on, practically all of it. Half of the result enters the weighted error.

In the example, the poll got right who leads, who went to the runoff and that there would be a runoff, so there is no adjustment. In another case, if Candidate A had finished with 49.5% and the poll had shown 51%, it would have pointed to a first-round win that did not happen. The poll was 1 point past the 50% line, but the result was only half a point from it: we charge 39% of that point, and half of that, 0.2 points, enters the weighted error.

So calling the wrong winner in a race decided by less than two points costs little: the difference between poll and result was within what chance explains.

Date correction (President only)

The second adjustment applies only to the presidential election, in the first round and the runoff, because the larger number of polls makes it possible to calculate. Some pollsters publish their last poll four or more days before the election, so their numbers lag behind the movement of the electorate. Comparing those polls directly with the result would charge them for movement that had not yet happened when they were in the field.

To estimate that movement, we compare the average error of other pollsters' polls with fieldwork ending on the same days (up to two days apart) with the average error of other pollsters' polls ending one to three days before the election. If the polls from the same period missed by more than those from the eve, the difference is discounted and the weighted error falls. In each window, each pollster's latest poll counts, and the discount is only calculated when there are at least five pollsters in each one. The discount appears in parentheses next to the weighted error; it never increases the error and never takes the result below zero.

Senate

In Senate races, the decision line is who gets the seat. The decisive margins measure the error in the margin of each winner over the best-placed loser, with the same greater weight for closer races. The misplaced electorate is calculated the same way. The wrong-side adjustment charges a poll that put in the seat a candidate who ended up outside. As there is no 50% line, the two factors always keep the same weight.

Runoff

In a runoff, only polls fielded after the first round count. For the 2026 runoff, we use the polls published in Plano Político's Agregador, plus PoderData's last poll, released on election day.

In a runoff there are only two candidates, and the poll's error comes down to the margin between them. The misplaced electorate is not calculated, because in this case it would repeat the previous measure. The wrong-side adjustment still applies to who wins.

In the runoff, the weighted error is the error in the margin between the two candidates, in percentage points of margin. That is why the numbers are not comparable with those of the first round.

How much of the error is noise

Even a perfect poll, with a truly random and correctly calibrated sample, constantly misses by pure chance. For each race, we simulate thousands of perfect samples from the official result and calculate the weighted error of each one, exactly as we do with the polls. The "chance band" shows how far the error of 9 out of 10 of these samples goes. In the 2026 presidential election, for example, half of the simulated perfect samples missed by up to 2 points, and 9 out of 10 by up to 4 points. That is why we show the weighted error in whole numbers: a decimal place would suggest a precision that chance itself does not allow.

As for the size of the perfect samples: the number of interviews each poll declares is not its effective size. Weighting, sample design, collection method: all of these add noise, making the polls' effective samples smaller than their actual samples and increasing the margin of error significantly.

In the vast majority of cases it is not possible to measure this reduction precisely, pollster by pollster, with the published data. The best option with the available information is to analyze how much each pollster varies from one poll to the next, to estimate its movement. However, some pollsters vary less from one poll to the next than pure chance would allow, which is typical of panel polls (when the same people are interviewed repeatedly) or of weighting to fixed targets (extremely common in Brazilian polling).

Using such a pollster-by-pollster estimate would reward stability, not precision. That is why we use the same effective size for every poll, estimated from the data themselves: we compare 515 pairs of consecutive polls by the same pollster in the same race, discounting the movement recorded by the other pollsters in the same period. On average, the polls fluctuate as if they had about 970 effective interviews, although the median declared sample is about 1,500. The chance band is calculated on that number.

The levels

Each poll is placed in one of three levels: below-typical error, typical error or above-typical error. The classification is derived as follows:

First we divide each poll's weighted error by the median error of the simulated perfect samples of the same race. This puts different races on the same scale: an election with many competitive candidates produces more error by pure chance than a two-way race.

Then we compare that value with the final polls for the same office in all the other elections in the series, leaving out the election being analyzed to keep the numbers internally consistent. First rounds are compared with first rounds, and runoffs with runoffs.

A poll has a below-typical error when it missed by less than 3 out of 4 polls in that comparison, and is classified as above-typical when it missed by more than 3 out of 4; by construction, typical error lies between those two thresholds. If a particular election was hard for pollsters, the levels show it: more polls from that year appear with above-typical error.

Formulas and data

Below are all the formulas used for the calculations described above. \(P_c\) is candidate \(c\)'s share in the poll and \(R_c\) the same candidate's election result, both in valid votes. Candidates are numbered by their order in the result (1st, 2nd and 3rd).

First round: president, governor and mayor

Errors in the margins:

$$e_{1\to 2} = (P_1 - P_2) - (R_1 - R_2) \qquad e_{2\to 3} = (P_2 - P_3) - (R_2 - R_3)$$

Weights for the closeness of the race and for the runoff spot:

$$w_{1\to 2} = e^{-(R_1 - R_2)/15} \qquad w_{2\to 3} = g \cdot e^{-(R_2 - R_3)/15} \qquad g = \begin{cases} 1 & \text{if } R_1 \lt 50 \\ e^{-(R_1 - 50)/2.5} & \text{if } R_1 \ge 50 \end{cases}$$

Decisive margins:

$$D = \frac{w_{1\to 2}\,|e_{1\to 2}| + w_{2\to 3}\,|e_{2\to 3}|}{w_{1\to 2} + w_{2\to 3}}$$

The misplaced electorate uses \(L\), the candidates the poll published separately (the others form a single group on both sides of the equation):

$$F = \frac{1}{2}\left(\sum_{c \in L} |P_c - R_c| + \left|\Big(100 - \sum_{c \in L} P_c\Big) - \Big(100 - \sum_{c \in L} R_c\Big)\right|\right)$$

Weight of the misplaced electorate:

$$a = 0.5 + 0.25 \cdot e^{-|R_1 - 50|/5}$$

Wrong side: for each decision line \(\lambda\), \(d_\lambda\) is how far the poll went past the line and \(\delta_\lambda\) is the distance of the result itself from it, both in points of electorate:

$$X = \sum_{\lambda} d_\lambda \left(1 - e^{-\delta_\lambda}\right)$$

Who leads: \(d = \tfrac{1}{2}\max\big(0,\ \max_{c \ne 1} P_c - P_1\big)\) and \(\delta = \tfrac{1}{2}(R_1 - R_2)\).

Who goes to the runoff: \(d = \tfrac{1}{2}\, g \cdot \max\big(0,\ \max_{c \ne 1,2} P_c - \min(P_1, P_2)\big)\) and \(\delta = \tfrac{1}{2}(R_2 - R_3)\).

The 50% line: if \(R_1 \ge 50\), \(d = \max\big(0,\ 50 - P_1,\ \max_{c \ne 1} P_c - 50\big)\); if \(R_1 \lt 50\), \(d = \max\big(0,\ \max_c P_c - 50\big)\); in both cases \(\delta = |R_1 - 50|\).

Weighted error:

$$\text{WE} = (1 - a)\,D + a\,F + \tfrac{1}{2}\,X$$

Senate

With \(W\) the winners (one or two seats) and \(\ell\) the best-placed loser, in shares of all votes cast:

$$e_w = (P_w - P_\ell) - (R_w - R_\ell) \qquad D = \frac{\sum_{w \in W} e^{-(R_w - R_\ell)/15}\,|e_w|}{\sum_{w \in W} e^{-(R_w - R_\ell)/15}}$$

\(F\) is calculated as in the first round, and \(a = 0.5\). The same rule, with one seat, applies to the capital where no runoff was possible (Palmas, 2020). Wrong side: \(d = \tfrac{1}{2}\max\big(0,\ \max_{c \notin W} P_c - \min_{w \in W} P_w\big)\) and \(\delta = \tfrac{1}{2}\big(\min_{w \in W} R_w - R_\ell\big)\).

Runoff

$$\text{WE} = |e| + \tfrac{1}{2}\,X \qquad e = (P_1 - P_2) - (R_1 - R_2)$$

with \(d = \tfrac{1}{2}\max(0,\ P_2 - P_1)\) and \(\delta = \tfrac{1}{2}(R_1 - R_2)\).

Date correction (President)

For polls with fieldwork ending four or more days before the election:

$$\Delta = \min\Big(\text{WE},\ \max\big(0,\ \overline{\text{WE}}_{\text{same days}} - \overline{\text{WE}}_{\text{eve}}\big)\Big)$$

"Same days" is the average weighted error of the latest poll of each other pollster with fieldwork ending up to two days before or after; "eve" is that of the latest poll of each other pollster ending one to three days before the election. The discount is only calculated if there are at least five pollsters in the relevant window.

Chance band, effective size and levels

Chance band: we draw 20,000 perfect samples of \(0.9 \times 970\) valid votes from the result (a multinomial distribution, with the candidates listed in the result and the others grouped) and calculate the weighted error of each with the same formula. \(A_{50}\) is the median; the 90th percentile is the edge of the band.

Effective size: for each pair of consecutive polls by the same pollster in the same race (3 to 14 days apart, first round for president, governor and mayor), \(\Delta m\) is the change in the margin between the top two minus the change in the other pollsters' average margin (up to five days before or after). With \(p_1\) and \(p_2\) the two candidates' average shares in the pair, as proportions:

$$n_{\text{eff}} = \frac{2 \cdot \operatorname{median}\big[\,10^4\,(p_1 + p_2 - (p_1 - p_2)^2)\,\big]}{\sigma^2} \qquad \sigma = 1.4826 \cdot \operatorname{median}\big|\Delta m - \operatorname{median}(\Delta m)\big|$$

Levels: \(r = \text{WE} / A_{50}\). The thresholds are the 25th and 75th percentiles of \(r\) among the final polls for the same office and round in all other elections in the series; below-typical error if \(r\) is below the 25th percentile, above-typical if it is above the 75th.

All the numbers used in the calculations above are available here. The file has one row per poll, with ";" as the separator and a comma as the decimal mark (Brazilian spreadsheet conventions; column names are in Portuguese). It contains the polls' shares and the result of the election they refer to, each component of the calculation, the date correction and the windows used for it, plus the chance band and the thresholds for each level.

Final remarks

We stress again that the Pollster Performance numbers do not seek to infer the quality of a pollster's polls in general: each race is analyzed on its own, and errors in one cannot be carried over to another directly. Moreover, historically the correlation between a poll's error and the next one's accuracy is not high. Yesterday's "perfect poll" often does not do as well tomorrow. Like the polls themselves, these are imperfect indications of trends.