Our note of 14 September set out four ways of evaluating the presidential forecast model's performance. Although the election is not over yet, here is our partial assessment of how the model behaved in the first round, historically the hardest part of Brazilian presidential elections to predict.
How the model did
The last update published on the eve of the election projected medians of 44.9% of valid votes for Lula (PT) and 44.3% for Flávio Bolsonaro (PL), with a 99.6% probability of a runoff between the two. The official result was 47.0% for Flávio and 45.2% for Lula.
Were the results within the model's intervals? Largely, yes. In the last update, the official result was inside the 90% confidence ranges of five of the six main candidates: Romeu Zema (NOVO) finished slightly below his. However, Flávio's result ended up close to the upper limit of his range (47.4%), and it had been above the range for most of the campaign.
How far from the median? For Lula, only 0.3 percentage points above the projection; for Cury, Renan and Caiado, between 0.3 and 0.7 points below. For Flávio, 2.7 points above. In other words, the model's central projection underestimated Flávio and slightly overestimated the smaller candidates.
Which direction did the model point in over the final days? Flávio grew steadily in the projections of the last few days. Between 30 September and 3 October, Flávio gained 2 points in the projection, while his chance of finishing the first round in front rose from 29% to 39%. The direction was right, but the model underestimated its strength.
How did the model arrive at this result? Our poll aggregator showed a 1.6-point lead for Lula in valid votes, in line with other aggregators such as BBC/PollingData (list of aggregators). The model's median projection, however, saw a lead of only 0.6 points for Lula. Most of that difference comes from a mechanism in the model that estimates the likelihood of voters anticipating the runoff, the so-called "tactical vote" (voto útil), in which the pattern of transfers from the smaller candidates to the two leaders was decisive in narrowing the gap.
One way of seeing this is through the two leaders. The combined valid-vote share of Lula and Flávio was 87.2% in the aggregator, while the sum of their medians in the model was 89.2%. The final result of that sum was 92.2%. In other words, the model's term for the consolidation of the vote around the two leaders helped reduce the imprecision of the aggregated polls and better anticipate the chances of a first-round result favourable to Flávio.
How the polls did
Our model depends heavily on opinion polls, which, despite constant complaints about their accuracy, are still the best tool for anticipating how the electorate is likely to behave. Consequently, if the picture the polls present on the eve of the election differs significantly from the result, the model will find it harder to project.
The gap between Lula and Flávio in the aggregator was +1.6 points, while the result was -1.9 points, an error of about 3.5 points in valid votes (setting aside the fact that the polls also slightly overestimated the smaller candidates' support). Although it looks like a large error, it was not evenly spread: four of the fifteen institutes that released numbers in the final week had Flávio ahead of Lula, while others had numbers compatible with the official result within their margins of error. In other words, there was reasonable uncertainty about who would come out in front on Sunday.
Many people's surprise, and the idea that "the polls got it wrong", come from a hasty and selective reading of the institutes. First of all, it is not possible to say with certainty that a particular institute "got it wrong": what changes are the assumptions needed to reconcile the projected number with the result. For example, can we say Quaest "got it wrong" if its estimates for every candidate were within the margin of error (2 points) of its valid-vote projection? Sampling error alone explains the difference. On the other hand, more divergent projections such as MDA's depend on other assumptions, such as high differential abstention or large last-minute shifts in the electorate.
In other words, the polls' "error" has to be explained in terms of its plausibility under statistical criteria. In the aggregator, we see movement in favour of Flávio (more) and Lula (less) over the last two weeks, which supports the idea that tactical voting consolidated and likely accelerated at the end. In addition, the polls as a whole improved in 2026 compared with 2022 on three important counts:
- They estimated the distance between the two main candidates better: in 2022, the average of the final week's polls missed the gap between Lula and Jair Bolsonaro by about 4.2 points; in 2026, our aggregator's error was 3.5 points.
- The polls correctly estimated the real chance of a first-round victory as close to zero: even with the result he got, Flávio would have needed to "take" almost 4 million votes from the other candidates, a very implausible shift given the polarisation and his rejection rates.
- They also better identified the plausible direction of the last-minute movement in the electorate. In 2022, the polls on average moved away from the final result in the last week; in 2026, almost all of those that published on the eve of the election (seven out of eight) moved towards the result.
The terms in the forecast model helped show how these factors amplified the chance of seeing something like the official result.
Second, much of the public's astonishment was caused by a selective reading of the data. On one side, many Lula supporters expressed disbelief and questioned the methodological integrity of some polls (and of our model) based solely on the results they showed, despite these being established institutes, with national polls in previous elections and a track record that can be assessed, such as Gerp and Futura. On the other, the preference for looking at one or two institutes instead of checking the aggregators led to undue confidence in one or two numbers, undermining the assessment of the uncertainty surrounding this race.
And now, in the runoff?
Uncertainty falls drastically in the runoff, and the forecast model becomes simpler. With the official results in hand, what remains is to estimate a few things: how much of their electorates the two finalists will retain in the runoff, how many voters from other candidates they will persuade, how abstention and blank and spoiled votes will behave, and finally how much movement can happen in the last three weeks of the campaign.
Since 5 October, the model has projected the runoff taking into account how these factors may play out. Historical patterns of vote transfers from the first round to the runoff are applied, with uncertainty about the transfer profile of each type of voter: those who chose other candidates, those who chose to stay at home and also those who voted for Lula or Flávio. In addition, we give some weight to the polls that simulated a runoff between the two before 4 October.
As runoff polls conducted after the first round are published, they will be incorporated into the model. Even so, the uncertainty about the result will still be calibrated according to how much the polls have historically shifted in presidential runoffs. While in the first round the difference between polls and results is relatively large and has always tended to overstate the PT candidate's lead, in the runoff the polling average is much more accurate and shows no particular direction.
The model's first publication has Flávio winning in 83% of simulations and Lula in 17%, a probability comparable to rolling a given number on an ordinary die. The average valid-vote share is 51.8% for Flávio and 48.2% for Lula. Flávio's chances rose by 20 points entirely because of the first-round result. In the simulations run on the eve of the vote in which the first round ended much like the official result, Lula's probability was between 14% and 19%.