Here’s the Economist’s tracker of polls leading up to Brazil’s October 2 first-round presidential elections, with the actual result of those elections added as horizontal lines.

Polling predicted challenger Luis Inácio Lula da Silva’s vote with reasonable accuracy. His 48.4 percent vote share—not enough to avoid a runoff election on October 30—was well within the band of probability that polls foresaw.

Polling utterly failed, however, to predict controversial President Jair Bolsonaro’s share of the vote. A consensus of surveys had pointed to October 2 being a blowout. Instead, Bolsonaro ended up just 5.2 percentage points behind Lula, with an outside chance of winning re-election in the second round.

Bolsonaro’s result was better than what he got in any of the dozens of polls that the Economist tracked (the blue dots in the chart). Only a handful came close.

Brazilian and international press will no doubt publish analyses over the next few days trying to explain how polling missed so bad, and what this means for the future of the opinion-surveying industry.

In the United States, where Donald Trump has outperformed his poll numbers by a few percentage points, analysts talk about “shy Trump voters.” That may have happened in Brazil, too: a lot of respondents who supported the far-right, often boorish populist president appear to have declined to say so in interviews with pollsters.

It’s also possible that pollsters under-sampled a pro-Bolsonaro sector of the population—although with compulsory voting (and nearly 80 percent turnout), the electorate’s makeup should have been easier to predict than in the United States.