What’s the Difference Between a 20% Chance and a 0% Chance?

This article originally appeared in Galaxy Research's weekly newsletter. Subscribe to get timely insights delivered to your inbox every Friday morning. Find our past reports and newsletters at galaxy.com/research. Send feedback to [email protected].

Prediction market skeptics are crowing (again) after Luiz Inácio Lula da Silva upset Flávio Bolsonaro in Brazil’s runoff election on Oct. 25.

Heading into the election, prediction markets had Bolsonaro holding a 75% to 80% chance of winning the election to Lula’s 11% to 20% probability after Bolsonaro won the first round of Brazil’s presidential election on Oct. 4. Lula ended up winning the runoff and remains Brazil’s president, and Bolsonaro fled to the United States.

Traditional polls had shown a statistical dead heat. The polling firm Datafolha had the two candidates tied at 50% each, while AtlasIntel/Bloomberg had Bolsonaro ahead 51.1% to 48.9% and an average compiled by The Rio Times put Bolsonaro at 50.9% to Lula’s 49.1%. Lula ultimately won with 51.8% of valid votes to Bolsonaro’s 48.2%, so most polls leaned slightly toward Bolsonaro and understated Lula’s share by one to four points. Even so, the polls never suggested Bolsonaro was a heavy favorite. Prediction markets offered a far more opinionated stance on the outcome of the election compared to polls that, when compensating for margin of error, offered no tangible takeaway. This setup was similar to that of the 2024 U.S. presidential election, where polls showed an ultra-close race with Kamala Harris barely in the lead while Polymarket gave Trump roughly a 60% chance of winning, a figure that stood at 62% on Election Day.

The difference was the ultimate outcome. In 2024, the markets’ bet that polls were understating the Republican candidate’s chances paid off, while in Brazil the equivalent bet failed. This gap is where critics point to an inaccuracy and apparent lack of efficacy in prediction markets’ implied odds.

OUR TAKE

The picture is not as black and white as the skeptics make it out to be. The take that prediction markets failed in this setting breaks down when taking into consideration what these markets actually tell you: predictions on probabilities. What the detractors miss here is that implied odds are not an implied prophecy. A prediction market price is a statement about frequency, not a guarantee that something will or won’t happen (e.g. a probability >50% does not imply the event must always happen just because the market suggests it is more likely than not). A contract trading at 80% is saying that if you could replay this event many times, the outcome would occur about four times in five, and fail the fifth. A 20% chance is expected to happen one time in five, not never. Lula winning at roughly 11% to 20% was a longshot, not an impossibility, and one longshot landing tells you very little about whether the price was wrong.

Human brains struggle to cope with this idea when looking at probabilities. Psychologists have documented for decades that people judge probability with mental shortcuts, and that these shortcuts lead to systematic and predictable errors (see Amos Tversky and Daniel Kahneman in Science, 1974). The problem shows up in elections specifically. In a Journal of Politics study, Sean J. Westwood, Solomon Messing, and Yphtach Lelkes found through experiments that probabilistic forecasts increase certainty about an outcome and confuse many readers. When one candidate is ahead, win probabilities convey substantially more confidence than vote share estimates do. The authors tied this observation to 2016, when forecasts gave Hillary Clinton a 70% to 99% chance of winning and left voters thinking the race was far less competitive than it was. A 60/40 or 80/20 market reads to many people as a verdict, when it is still a description of uncertainty. This misconception turns Bolsonaro’s 75% to 80% probability into “the markets said Bolsonaro would win,” and a Lula victory into a breakdown in market efficacy, when an 80/20 market is still just a description of some uncertainty.

Historically, over many markets, prediction platforms’ prices hold up. Kalshi Research tested more than 2.2 million resolved markets, and found prices generally behave like genuine probabilities, with a contract at 70 cents resolving “Yes” about 70% of the time. The calibration plots below show this. Each dot groups markets by their price (horizontal axis) and shows how often those markets actually resolved “Yes” (vertical axis). The dashed diagonal is perfect calibration, where a 70% price means a 70% hit rate. The dots hug that line, and they sit closest to it in the panels measured one week and one day before resolution, which is the window most relevant to the Brazil runoff. Prices are also more accurate as resolution nears, with Brier scores (a standard error measure where lower is better) falling from roughly 0.08 to 0.09 at three months out to about 0.02 at close. This shows that even in the markets that were well calibrated, the favorite didn’t always win. For elections, the favored outcome came in about 90% of the time across most horizons, which means roughly one favorite in 10 lost. The authors also note that in a race with more than two viable candidates, a favorite priced at 60 cents should win only about 60% of the time, and that near-perfect hit rates would point to underpriced favorites rather than better forecasting.

Source: Kalshi Research (https://kalshi.com/research/kalshi-research-calibration.pdf)

Brazil’s stock market made a similar bet to that of prediction market odds. After Bolsonaro won the first round, the Ibovespa index jumped almost 8% on Oct. 5, from 192,138 to 206,912, and peaked at a close of 209,067 on Oct. 9. It chopped around this level until after the runoff, when it fell 13.9% in a single session, from 205,083 to 176,608, and kept sliding. As of Oct. 29, it was trading below 170,000 which is about 17% below its pre-runoff close and below where it stood before the first round of the election. The Brazilian Real also lost more than 8% after an election result the market seemingly did not expect. Equity investors didn’t have a secret read on Lula that the prediction markets missed. Everyone seemed to price Bolsonaro as the favorite.

Taking this into consideration, prediction markets worked as a viable hedge. A trader long Brazilian stocks could have bought Lula “Yes” contracts as insurance. A contract priced near 20 cents paid five times the stake, so the Lula position would have paid out many multiples as the Brazilian stock market entered a free fall. A 20% price was never a claim that Lula couldn’t win. It was the market rate for protection against this outcome.

The data suggests prediction markets were not wrong in the way their doubters are claiming. A price of 75% to 80% means the favorite loses one time in four or five, and Kalshi illustrated this is roughly the case in aggregate.

Then again, if the election was stolen, as Bolsonaro has claimed, maybe the markets were wrong, at least about the ultimate outcome (even if they were right about his popularity).