Even famed statistician Nate Silver couldn’t have forecast the first-round results of Brazil’s presidential election, which heads to a final runoff vote this Sunday. And not just because the campaign has unfolded with Dickensian complexity—down to the colorful cast of characters, tragic death, and political rebirth. From a scarcity of polling data to the very way elections are held in Brazil, political forecasters here face unique challenges, the results of which have been on display during this volatile election season.
Much of the unpredictability is a product of Brazil being a young democracy. The country’s lack of historical election data makes it especially difficult to divine voters’ intentions, according to Clifford Young, President of U.S. Public Affairs and former Managing Director for Brazil at Ipsos, a global market and public opinion research firm.
“Nate Silver would have had the same troubles here,” Young said in a telephone interview.
In the 2012 U.S. presidential election, Silver accurately predicted the winner of all 50 states and Washington, D.C. thanks to the input of some 13,000 polls. According to Young, in Brazil he would be limited to no more than 75 polls.
“A luxury we have in the U.S. is it’s a poll-rich environment, we have thousands of polls that can all be aggregated,” said Young. “Brazil doesn’t have as many polls—the lower your sample size, the more prone to error are your estimates.”
Brazil’s presidential election is heading to a runoff on October 26 between incumbent Dilma Rousseff and challenger Aécio Neves, who were the top two vote-getters in the initial 12-candidate contest on October 5. Neves trumped third-place finisher Marina Silva by nearly 13 percentage points, even though polls showed them in a dead heat just a day earlier. In what amounts to further evidence of Brazil’s polling and forecasting troubles, an exit poll by the respected Brazilian research group IBOPE projected that Rousseff had won 44 percent of the vote, though she ultimately won just 42 percent. The same exit poll gave Neves 30 percent of votes cast, though he actually won 34 percent—outside the two-point margin of error.
The polling discrepancies continue. On October 11, Sensus, a pollster based in Minas Gerais, published a poll that gave Neves a 16-point lead over Rousseff. Meanwhile, all polls conducted since October 5 by Datafolha and IBOPE have put the candidates in a statistical tie. Heading into the first-round vote, wild swings in day-to-day polls roiled the stock market and local currency while raising questions about the integrity and accuracy of local pollsters.
Young agreed that IBOPE’s inaccurate exit poll underscored the need for refinement in survey methods. For instance, pollsters need to start asking voters if they intend to cast a ballot, as about one-fifth of the electorate stayed home on October 5 despite the fact that voting is mandatory. Young said the exit poll’s inaccuracy could also be due to voters thinking they were voting for one candidate when they were actually voting for another, implying that the ballot itself may need revamping.
All this hasn’t stopped João Manoel Pinho de Mello of Insper Business School in São Paulo from building a forecasting model based not just on polling results, but also on fundamental data points such as the bias toward incumbency and the health of the economy. But even here Brazilian pundits are at a disadvantage. Whereas U.S. forecasters can strengthen their models with 240 years’ worth of election results, data for Brazil is only available since 1989, the year it held its first open vote under the nation’s new democratic constitution.
“Brazil is a relatively new democracy,” Professor de Mello said. “There’s much less accurate data in Brazil. We’re talking about five to six historical data points, including this one.”
Brazilian elections are also inherently more volatile and unpredictable because of the country’s multiparty system. That’s why the former favorite Silva, for example, could lose “change” voters to Neves, low-income voters to Rousseff, environmentalist voters to Green Party candidate Eduardo Jorge, and evangelical voters to the Christian minister Everaldo Pereira. The U.S.’s two-party system makes it less likely that voters will shift their allegiance at the last minute.
An accurate political forecast in Brazil would need to gauge the intentions of all 142.8 million voters, as the president is elected by direct national vote. But in the United States, the president is chosen by an electoral college that awards state votes on a winner-takes-all basis, which is why forecasting basically boils down to analyzing several “swing” districts in Florida and Ohio. This allows U.S. pollsters to hone their survey sets and collect more nuanced data.
Even the way polling is done in emerging nations is more challenging. In Brazil all polls are conducted face-to-face because—traditionally—only more affluent voters have had access to landlines or mobile phones. In contrast, in the U.S. pollsters can conduct cheaper and more numerous polls with live telephone surveys and robocalls. According to Young, that could be changing. Stronger mobile phone penetration may make it possible for Brazil’s next presidential election to include telephone surveys.
But it’ll be a while before election forecasting in Brazil is as accurate as in the United States, he notes. “I’m a pollster saying take the polls with a grain of salt.”