Key Takeaways
- Prediction markets priced Flávio Bolsonaro's win probability between 80% and 87%, treating a close race as a near certainty.
- Political risk analyst James Bosworth pegged Bolsonaro's odds at 64% after viewing the first-round data, having viewed the race as a coin flip earlier.
- Luiz Inácio Lula da Silva won the runoff by 4 million votes despite public polling pointing toward a Bolsonaro victory.
- A shift of under 2% in voter turnout and another 1% to 2% swing from blank and invalid ballots determined the entire race.
- Betting markets failed because traders did not generate independent data; they simply recycled herded, correlated survey results.
Prediction Markets Trade the Same Bad Data as Everyone Else
Bettors love to treat prediction markets as truth machines. When people put real money on an outcome, the theory goes, they cut through bias and arrive at the true probability.
The Brazilian runoff proved that prediction markets have a glaring blind spot: they rarely produce new information. When pollsters make systematic mistakes, bettors take those mistakes and compound them into extreme odds.
James Bosworth, a political risk analyst, watched this unfold during the presidential contest between Luiz Inácio Lula da Silva and Flávio Bolsonaro. As Bosworth recalled, “I had Bolsonaro winning at about 64%. So prior to the first round, I was essentially coin flip.” (play from 8:17) But betting markets went much further: “And the prediction markets were at 80%. In fact, there was a moment where the prediction markets hit 87% for Bolsonaro.” (play from 8:26)
Traders were not uncovering private intelligence. They were looking at public polls, assuming those polls were independent data points, and bidding up Bolsonaro contracts.
Correlated Errors Turn Small Poll Biases into False Certainty
When multiple polling firms run surveys with similar methods, they often copy each other to protect their reputations. No pollster wants to risk their business on an outlier result. The result is herding.
Bosworth admitted that he was trapped by the same consensus: “All of the data, all of the polls, said Flávio Bolsonaro was going to win. And so, based on the first-round data, it looked like Flávio winning. But you have to factor in that the polls seem to be herding together and correlated.” (play from 8:56)
If five polls show the same candidate ahead by three points, bettors treat it as five distinct confirmations. In reality, it is often one flawed model repeated five times. If survey firms undersample working-class neighborhoods or misjudge voter enthusiasm, every survey errs in the exact same direction. The market reads that false consensus as ironclad certainty, pushing odds past 85%.
A Two Percent Shift Swings Four Million Votes
Elections in large nations rarely hinge on grand ideological conversions. They turn on whether tired voters stay home or decide to show up.
Lula did not need a sweeping change in public opinion to defeat Bolsonaro. He only needed fractional movements on the margins. As Bosworth pointed out: “The change in turnout from the first round to the second round was less than 2%. You had a slight shift in terms of non-voters, people who cast blank or invalidated ballots the first round. You had another 1% or 2%. But in Brazil, that's a couple million votes. And in an election won by 4 million votes, a shift of 1% or 2% across those rounds really did make the difference.” (play from 9:20)
Polls and prediction markets are too blunt to detect a 1% shift among citizens who blanked their ballots three weeks earlier. When a forecasting model cannot track movements that small, betting markets that price an outcome at 87% are trading on fiction.
What to Do With This
Audit your key operating metrics for correlated inputs before making your next capital allocation decision. If your product team cites three customer surveys, check whether all three drew from the same email list or user cohort; if they share an origin, treat them as one single opinion rather than three separate validations.