Model log

Probabilistic projections such as Plano Político's presidential model should not be read as a guaranteed forecast of the election result. On the contrary, the purpose of a statistical model like this one is to quantify the uncertainty that exists around the outcome, informed by the record of past elections. If the model is perfectly calibrated, a candidate given a 10% chance will win one election in ten.

Presidential elections, however, are rare events: they happen only once every four years, each with a different cast of candidates and a different political context. It would be easy to artificially inflate the model's uncertainty so that it is "never wrong", but that would undermine its analytical usefulness. On the other hand, an overconfident model creates the opposite problem: mistaking a temporary trend for inevitability.

We are twenty days from the first round of the presidential election. A great deal can happen in twenty days, and the electorate reacts increasingly quickly the closer it gets to polling day: in 2014, for instance, at this same stage of the campaign, Marina was favourite to face Dilma in the runoff; in the final week Aécio overtook her and finished the first round twelve percentage points ahead.

A probabilistic model cannot predict the exact result; instead, it maps out those possibilities so that they are on the radar of anyone following the election. The best way to evaluate the model, then, once we know the result on 4 October, is to be guided by the following points:

• Were each candidate's individual results within the model's confidence intervals?
• If so, how far were they from the median projection?
• Did the direction the model pointed in over its final days indicate a greater or a lesser likelihood of the eventual result?
• Finally, which terms in the model were decisive in landing closer to — or further from — the final result?

In the days following the election, Plano Político will publish an analysis of the presidential model's results in light of these questions, pointing out what it got right and wrong and, should there be a runoff, an explanation of how the model will work in the race's second phase. The reason is simple: although it is technically a "new election", the runoff takes place in the shadow of the first round's results, which allow the two remaining candidates' probabilities to be read with far greater clarity.

For the first time, Plano Político's presidential model gives Flávio Bolsonaro (PL) a probabilistic advantage over Lula (PT). The Rio senator's chances have been climbing over the past two weeks, driven above all by his gains in runoff scenarios across practically every poll released during September. Today, with three national polls released — two showing Flávio ahead in the runoff and one a tie — Lula records his worst probabilistic result since the model was published, and across the reconstructed record going back to 5 April of this year.

There is no way to separate out how much of this movement comes from recent news coverage. The swing towards Flávio is not without precedent: in 2022 his father, then-president Jair Bolsonaro, also gained ground in the polls over the same period, in late August and early September. That year the polling movement reversed in the second half of the month, which did not stop the former president from finishing above what the final week's polls had projected.

Our model projects an average runoff result of 50.5% for Flávio against 49.5% for Lula. That is the average across all simulations in which both candidates reach the second round. Many simulations therefore return results more favourable to each of them, which leaves the outcome still highly uncertain and makes for a very close election. Even so, there is no denying that the current picture favours the PL candidate.

This version of the model fixes defects in the previous one that let the detection of movement in the polls overstate the estimates of candidates who are gaining ground, by failing to identify correctly when that growth has stopped. The revision has very small effects on the current estimates, except on the confidence intervals of the smaller candidates, Augusto Cury in particular. The movement in the probabilities seen today is not caused by these fixes, but by the new information coming from the polls published this week. The history has not been recalculated.

The new version of the model we are publishing today also shows the chances of Augusto Cury, of AVANTE, of winning the presidential election. The changes required keep the model's underlying logic intact: Cury as a candidate is mapped onto our base according to his target audience within the Brazilian electorate, and compared directly against the polls to estimate his room to grow and how his voters would behave in hypothetical runoff scenarios.

This change was planned from the model's creation, though not specifically for Cury. The five candidates we have been tracking here since the model was published launched their candidacies many months ago, before the April resign-to-run deadline. Historically, candidacies like Cury's are not necessarily confirmed until the party conventions. Once they are confirmed and registered, we monitor every minor candidacy in case one or more of them shows growth capable of changing the course of the campaign. That is Cury's case now.

In our poll aggregator, Cury held steady through all of July and the first half of August, moving between 1.48% and 1.80%. Since then he has moved above that band, reaching 2.07% on 26 August and closing the month at 2.04%. On top of that, over the past week several indicators of rising interest in his candidacy have run above their normal levels, including Instagram follower counts and Google searches, as well as growth in voting intention in private trackers — noting that we do not publish figures that are not publicly available.

Given the evidence, both the Presidential Model and the Aggregator will show voting intention and win probabilities for Augusto Cury from today, 30 August, onwards. The changes to the model do not affect the rest of the reconstruction we published, but they do directly affect the numbers from today onwards. Previously published estimates stand as the historical record, which has not been recalculated to account for his inclusion.

Pablo Marçal's candidacy is unlikely to be approved, but while it stands, a TSE resolution requires pollsters to include him in stimulated first-round scenarios until the case is decided. Our aggregator always prefers the scenario without Marçal, but not every pollster publishes both. So we built a two-stage correction: in the polls that carry both scenarios — Datafolha, Quaest and Indexa among them — we measure how much of Marçal's support goes to Flávio Bolsonaro, and we only apply the correction if the estimate is statistically significant. For now the readings vary far too much from poll to poll, from −40% to +75%. On top of that, Marçal polls at 1% to 2% in most scenarios, too small to measure against. The correction is in the model but is not being applied for the time being; it will kick in if and when the data supports it.

Testing that correction surfaced a bug in our code, since fixed: in some state polls, minor candidates were being counted as blank, null and undecided votes, which artificially inflated Lula's and Flávio's first-round valid-vote shares. With the fix, the chance of the race being decided in the first round fell from 10% (on 24 August) to 4%, and Lula's probability of winning went from 64% to 61%. That shift also reflects the most recent polls, which were more favorable to Flávio. We also adjusted how much weight the Aggregator gives pollsters that were accumulating influence simply by publishing more often. The model's entire history has been rebuilt with these corrections.

After several legal actions against Instituto Veritá's polls — including one suspension over suspected duplication of respondent records — we decided to remove their polls from the Aggregator and the Presidential Model. The effect on the Presidential Model's numbers was below rounding margins, but we're reporting it here to explain the change in the number of polls the model considers.

The judicially suspended polls were never part of our database to begin with: we only include polls whose numbers were published in the press, and a suspended poll cannot legally have its results disclosed. The decision to also remove Veritá's other polls — the ones no court has suspended — is a precaution, and it mainly affects state-level estimates in a handful of states.

It's worth stressing that we only remove polls when strictly necessary. Many legal challenges against polls, especially state-level ones, are frivolous. That is not the case here: the action cited above rested on statistical analysis that found data duplication in a poll conducted in Amazonas.

Welcome to Plano Político's 2026 Presidential Model. This model is a probabilistic projection of Brazil's 2026 presidential election, built from public data on election results and opinion polls, with the aim of mapping the probabilities and the uncertainties of the presidential campaign. To read more about the model, see the methodology. This page will publish every change to the model's results, as well as any corrections to its architecture that prove necessary over the course of the campaign.

As of 15 August, the win probabilities are as follows: 63% for Lula (PT), 33% for Flávio Bolsonaro (PL), 2% for Renan Santos (MISSÃO), 1% for Ronaldo Caiado (PSD) and Romeu Zema (NOVO). The probability of the election being decided in the first round is 12%.