BR

Methodology

The goal of this exercise is to describe the behavior of Brazilian election polls in presidential and gubernatorial races, offering multiple metrics to analyze each pollster's performance in recent Brazilian elections. Each election's page presents these metrics as calculated within that electoral cycle. The meaning of these metrics and what they aim to measure is explained below.

Weighted Error (WE)

There are several ways to measure the size of an election poll's error. Reasonable criteria need to account for what polls actually claim to do, though. Pollsters often compare their output to "snapshots" of public opinion during their fieldwork window, which makes their predictive power dependent on the fluctuations that occur during the campaign.

Our main metric is Weighted Error (WE), which uses the last poll a pollster published before the election under analysis. It isn't a perfect match, but it's the closest approximation available between the public opinion expressed at the ballot box and the one captured by a pollster. Final polls are usually published on the eve of the election, reducing the chance of significant shifts in the electorate afterward.

After converting every available number into an approximation of "valid votes" (i.e. excluding blank and null votes) — the criterion the TSE uses to report and certify election results — we calculate the margins between pairs of the leading candidates (at least 10% of valid votes in the election) in percentage points and compare them to the margins shown by the polls.

For example, if the official result had Candidate A at 40%, Candidate B at 30% and Candidate C at 15%, the margins between pairs are 10 and 15 percentage points, respectively. A poll showing Candidate A = 42%, Candidate B = 28% and Candidate C = 13% got the B-C margin right but missed the A-B margin by four percentage points. These numbers are weighted to give more weight to errors between pairs of candidates in tighter races, which carry bigger narrative consequences.

We also apply two extra penalties: one for pollsters that over- or underestimate the share of the electorate that concentrated on minor candidates (below 10% of valid votes), and one for polls that incorrectly estimate how close a candidate came to winning in the first round (more than half the valid votes). For example, a poll showing Candidate A at 53% of the vote frames the election as likely decided in the first round; if the final result shows 45% for Candidate A, the poll overestimated the chance of a first-round win and is penalized slightly. The same applies if a poll underestimated the chance of a first-round win. The penalty grows as the official result gets closer to 50% for the leading first-round candidate.

This method aims to size up polling error appropriately without requiring a perfect match on each candidate's percentage; what matters is getting the magnitude of the margins between the leading candidates right, since that's the information most relevant to political parties and the public at large. Other ways of calculating error tend to overweight errors in minor candidates' percentages, or focus unduly on numeric accuracy for one candidate or another.

Bias

Using only the last poll, however, limits the analysis to a small slice of the information pollsters make public. So we use every poll released in a cycle to build other metrics. While Weighted Error assesses how close a pollster got to the final result, Bias measures the tendency to inflate or deflate a specific candidate's performance across the whole campaign cycle.

It's important to note that Bias does not mean manipulation. Different polling methodologies produce different results, and some methods tend to capture certain candidates more than others. Our definition of Bias is a mathematical one.

To calculate this metric, we pick a reference candidate and compare their share of valid votes in each of the pollster's polls to the percentage they actually won in the election. The average of that difference, weighted to give more weight to polls closer to the election, is the pollster's Bias for that cycle, expressed in percentage points. A positive sign means the pollster, on average, showed the reference candidate above the final result; a negative sign, below. Calculating bias against a single candidate — rather than an average across the whole field — is necessary because deviations relative to the field always cancel out: if one candidate was overestimated, another was necessarily underestimated.

For example, if in a given election the reference candidate finished with 35% of valid votes and the pollster, across the cycle, showed the candidate averaging 39%, the pollster's Bias is +4 percentage points. If it showed the candidate averaging 32%, the Bias is −3.

The choice of reference candidate follows a fixed rule, applied the same way to every cycle. The PT candidate is the first choice whenever their share of valid votes reaches the eligibility threshold (10%): the party has run in every presidential election since 1989, offering a consistent point of comparison across cycles, and it's also the party with the most solid ideological identity among Brazil's major parties. When the PT candidate doesn't reach the threshold, we use the incumbent seeking reelection, or the sitting government's political heir, as the reference for state executive elections. When neither option is available, the cycle appears with no Bias calculated.

Worth repeating: the Bias measure isn't necessarily a negative judgment on the pollster. It can reflect a genuine feature of the methodology — for example, a sampling mode that tends to reach a demographic more aligned with one candidate — or a choice not to update the weighting model during the campaign. What matters is understanding what the number describes: how much the pollster, over time, counted the reference candidate above or below what the electorate actually decided.

Late entries

In some elections, relevant candidates only enter the race after the campaign has already started — Marina Silva in 2014, Fernando Haddad in 2018, and so on. Polls that predate a candidate's entry don't make sense as a basis for measuring the pollster's Bias: the candidate simply didn't exist as an option when those polls were fielded.

When there's a late entry, we apply a transition in the Bias calculation. Polls that predate the entry are ignored; polls from the following two weeks get increasing weight (from zero to one), giving pollsters time to incorporate the new candidacy; after that window, every poll is weighted normally. The transition applies only to the Bias calculation — the other metrics are unaffected.

Movement Capture (MC)

Weighted Error and Bias assess the pollster against the official result. Movement Capture (MC), by contrast, describes how the pollster portrayed the campaign's dynamics — how the electorate moved, and in which direction. The measure splits into three complementary components; two of them are computed for every election and expressed in percentage points per week.

The first component, Direction, measures which way the margins between eligible candidate pairs moved across the pollster's polls. A positive sign means the predominant direction was widening margins between the candidates above and below the TSE's final result; a negative sign shows the opposite. Combined with Weighted Error, Direction reveals different performance profiles: a pollster with low WE and positive Direction described a campaign whose dynamics were confirmed at the ballot box; a pollster with low WE and negative Direction got the final read right but showed a different trajectory than the final result.

The second component, Volatility, measures the average size of swings between consecutive polls, without a sign — regardless of the direction of movement. Each swing is corrected for the poll's statistical uncertainty (which depends on sample size), so small fluctuations caused by statistical noise aren't counted as real movement. A pollster whose polls showed large swings between releases will have high Volatility; one that showed stable readings will have low Volatility.

Both components weight each interval between polls by the number of days it covers — longer intervals weigh more than short ones, reflecting the fact that they cover more of the campaign and offer more stable estimates of the rate of change. These are descriptive metrics, not evaluative ones: what these numbers tell you is the narrative the pollster told during the campaign — useful for identifying who captured movements others didn't, who held a stable read while the rest of the field swung around, and who drew a curve that would or wouldn't be confirmed at the ballot box.

Synchrony

Synchrony is the third component of Movement Capture. It analyzes how much the swings a pollster identified were also captured by the rest of the field. In dense electoral cycles — with several pollsters publishing regularly — it's possible to compare each pollster's swings to the field average.

The measure ranges from 0 to 1. A pollster that always moved in the same direction as the field, and by a proportional magnitude, gets a Synchrony near 1. A pollster that moved with no consistent relationship to the field gets a Synchrony near 0.5. A pollster that systematically moved opposite to the field gets a Synchrony near 0. The score combines two elements: whether the direction of changes matched the field's (weight 2/3), and whether the magnitude of changes was proportional to the field's (weight 1/3).

To be informative, the Synchrony metric requires a minimum number of pollsters and polls: at least four eligible pollsters and a total of twenty polls in the cycle. When there are fewer polls or pollsters, Synchrony isn't calculated. That's the case for most state-level cycles, where polling density is lower.

High Synchrony doesn't necessarily mean the pollster got it right — it means it moved along with the majority. Likewise, low Synchrony doesn't mean error — just that the pollster drew a different curve than the consensus. Combined with Weighted Error, the Synchrony metric lets you distinguish between a pollster that diverged from consensus and was right, one that diverged and was wrong, one that followed consensus and was right, and one that followed consensus and was wrong along with everyone else.

Who's included

The universe of polls we analyze comes from Poder360's poll aggregator, which compiles polls registered with the TSE and made public since the 1990s. The official results used as the comparison baseline come directly from the TSE, in the final consolidated version after vote counting.

To appear in an election's main table, pollsters need to meet two criteria. First, having published at least two polls during that specific campaign: Movement Capture needs at least one interval between consecutive polls to have anything to measure. Second, having published their most recent poll within the seven days before the election: older polls would be judged against the wrong snapshot — not the electorate's final portrait, but an outdated one from earlier in the campaign.

Pollsters that don't meet both criteria appear in a separate "unranked" table below the main one. Their Weighted Error is still shown as a diagnostic reference — it can be calculated, but it has limited meaning because it was measured under different conditions than the rest and so can't be directly compared.

Rationale

The metrics described above stay separate and aren't combined into a single score. The intent is for viewing all of them together to reveal the "story" each pollster told during the election. For example, a house showing high Weighted Error, low Volatility, high Synchrony, and a strong Bias in one direction supports a reading like: "got it wrong along with the rest of the field, without wavering, systematically favoring one side." A single composite number would lose that interpretive possibility.

In short: the model is focused on describing the polling field and avoiding rushing to rank one pollster as better than another. After all, a pollster might have a perfectly sound methodology but still perform poorly for a variety of reasons — including statistical swings that plain sampling randomness would predict. Likewise, a pollster with a questionable methodology might get lucky and land on the right result. The model avoids making that judgment call. The partial exception is Weighted Error, which compares each pollster to the official result; even so, it was designed to use only each pollster's last poll, precisely because that's the only point where a direct comparison to the ballot box makes sense.

We tested the model's structure with sensitivity and robustness analyses. More importantly, we applied the finished model — with no adjustment — to 22 governor-race cycles from 2014 that hadn't been seen during its design, and it correctly reproduced the relative size of errors documented at the time, such as Sartori's late surge in Rio Grande do Sul.

Limitations

← Back to the list of cycles