Methodology — Electoral Models

The model simulates the election tens of thousands of times. Each simulation starts from the aggregated poll average, adjusts the uncertainty using the error polls have historically shown at this point in the electoral calendar, accounts for the movement that can still happen before the vote, and resolves the runoff from projections of how votes transferred in previous elections.

In the presidential model, we take into account the distribution of voters along the Brazilian ideological spectrum, estimated from a range of surveys and from recent electoral behavior. That distribution is not fixed: each simulation picks a plausible combination consistent with the available data. From there we estimate the voting intention each candidate manages to convert within each ideological bloc, and we also model the “non-vote” (blank, null and undecided).

Each simulation also takes into account how much the scenario can still change between now and the election. We estimate whether a candidate's movement points to a rise or a fall in voting intention, we simulate the trajectory of those changes, and we account for the chance of information shocks that change the pace of the campaign.

In the end, the simulations produce a complete first-round result and, where necessary, also simulate the runoff, whichever candidates reach it. The transfer of votes from the first round to the runoff is likewise based on historical estimates and current projections of how each candidate's electorate may behave.

In the state government model, each of the 27 state races is simulated tens of thousands of times, and published wherever there are enough polls.

Each simulation starts from an estimate of every candidacy's political structure within its own state, based on the electoral coalitions and on the vote won by the federal and state deputies backing the candidate. Being the sitting governor or vice-governor also counts, where that applies.

That structural projection is then corrected by each state's polls, as collected and presented in our aggregator. How much weight the polls carry in each candidate's probability varies by state. In some races the political structure already makes the favorites clear before any campaigning begins. In others the contest is closer, and only public opinion shows who holds the advantage.

To estimate the uncertainty around our projections, we take into account how far the polls sit from the result at each point in the campaign, including the probability of extreme departures from what the model expects — the unlikely election of Wilson Witzel as governor of Rio de Janeiro in 2018, for instance, which neither the polls nor the political structure saw coming.

Each simulation produces a first-round result: if a candidate takes more than 50% of the valid votes, the election is decided in the first round. If not, the model also resolves the runoff, drawing on how state runoffs have behaved historically. Once the first round is over, the model will also take the actual 2026 results into account when projecting the runoff.

In the Senate model, each of the 27 state races is simulated tens of thousands of times. In 2026 every state elects two seats — 54 of the 81. The other 27 were filled in 2022, run until 2031, and are not in play.

Each simulation starts from an estimate of every candidacy's political structure within its own state, based on the local party system and on each candidacy's electoral record — the offices it has contested and held.

That structural projection is then corrected by each state's polls, as collected and presented in our aggregator. How much weight the polls carry varies by state, for the same reason as in the governor model: in some races the political structure already points to the favorites before campaigning begins, in others the contest is closer and only public opinion shows who leads.

To estimate the uncertainty we take into account how far Senate polls sit from the result at each point in the campaign. That uncertainty is wider than the governor model's, and not by accident: a Senate seat depends on personal vote and on the combination of candidates the electorate selects, beyond what political structure can predict.

There is no runoff: each simulation produces a single-round result and the two most-voted candidacies are elected. So we publish each candidacy's individual probability of being elected, not the probability of a particular pair winning together.

Percentages follow the TSE's basis, as a share of valid votes. Because each voter picks two names, the candidacies add up to 100% and a candidacy shown at 25% is named by roughly half the electorate. The projected Senate for 2027 adds the 54 contested seats to the 27 already settled.

Building a predictive model for Brazil's Chamber of Deputies is a major challenge. Several complications make the undertaking radically different from models for other elected offices:

• The format of the race, in which seats are assigned according to the electoral result of each state list but distributed according to the individual votes cast for deputies, introduces several sources of uncertainty;
• There are few public opinion polls for federal deputy, and even the existing ones usually have far more difficulty estimating the size of candidacies correctly than polls for majoritarian offices;
• The result does not depend only on who receives the most votes, because the distribution of leftover seats means that a relatively small number of votes (in some cases, a few hundred) decides which parties take the last seats in a state.

Our model should therefore be understood as an experimental attempt to reduce the uncertainty about the relative size of each party in the 2026 elections. A higher error rate is expected than should be observed in the models for state governors and for the Senate.

The Plano Político model for the Chamber of Deputies is in fact made up of four different models. The first assesses the likely size of each state list: what percentage of the valid votes each party will receive in each state. The second and third estimate each candidate's vote by different routes: the second combines several detailed sources on each candidate's electoral record and campaign funding; the third, simpler, groups candidates by their electoral experience and estimates each group from its best previous result. These three models are purely statistical and were estimated on the 2014 and 2018 elections and tested mainly on 2022. They were then estimated again including the 2022 elections, for use in the 2026 elections.

The 2022 tests have a reasonable degree of accuracy, but still miss between 80 and 100 elected seats. After analysis, we confirmed that an important part of that error can be resolved with more information about the candidates: who each party's biggest bets are, which celebrities might receive large votes with no or almost no political experience, networks of support and incentives, etc.

This is where the fourth part of the model comes in: qualitative research on more than a thousand candidates in 2026, carried out with the help of artificial intelligence to gather information that changes vote expectations and their degree of uncertainty. With this information, we developed a probable distribution of each candidate's vote and of their chances of being elected, making the expectations consistent and running twenty thousand simulations of the election to produce estimates for each party list in each state.

Our analysis has varying degrees of uncertainty. For example, 70% of the projected seats are practically safe and go to the same party in 95% or more of the simulations. Getting 70% of the Chamber right is easy, because the level of electoral persistence is high: many of the candidates have been elected before, and these politicians tend to keep their vote, even when they change party. The hard part is estimating the other 30%. In our state view, seats with the most uncertainty about the party appear hatched or outlined.

How to evaluate the model: we expect at least thirteen of the fifteen parties or federations estimated to elect a number of deputies within the 90% confidence intervals shown on the main page. If every party falls within those intervals, the model was probably not confident enough; if more than three miss, the model was overconfident. For the state estimates, the expected error rate is correlated with the size of the state: states with only eight seats can in some cases be decided by just a few hundred votes, which is beyond the reach of any reasonable estimate. In states with more seats, estimation errors tend to offset one another, so the overall error rate ends up lower.