Just as voting was due to end in Brazil’s presidential runoff on 25 October, a supreme court justice ordered polling stations to stay open for another two hours, to give voters stranded by flash floods and bus shortages more time to cast their ballots. It added more drama to an already tense election. The final polls the night before showed Flávio Bolsonaro only marginally ahead, by about 51% to 49% – within the margin of error. A statistical tie.
Prediction markets told a different story. Polymarket priced a Bolsonaro victory at about 80% on election day, where it had hovered for three weeks. On the eve of the vote it touched 84%. Even as polls closed, the market gave the incumbent, Luiz Inácio Lula da Silva, roughly a one-in-five chance. A few hours later, Lula had won, by 51.8% to 48.2%.
Prediction markets claim to be highly accurate because they aggregate the independent judgments of many traders – the “wisdom of crowds” – and make each one put money behind their beliefs. Prices are more accurate than polls and faster than newsrooms, proponents argue. So what went wrong?
As Americans prepare to vote in the first midterms since prediction markets broke into the mainstream, the question has taken on new significance. The founders of Polymarket and Kalshi evangelise market efficiency as a kind of secular faith. They promise a more accurate way to forecast the future, one that harnesses the collective intelligence of traders with money on the line.
But Brazil was not a one-off. Across a string of high-profile elections, the markets have been confidently wrong. Meanwhile, the more money flows into election contracts, the more there is to gain from distorting them or exploiting privileged information. The US midterms – the night prediction markets have been waiting for – will be their biggest test to date and potentially a referendum on their authority.
The rise of political prediction markets
In recent weeks, traders have swarmed to wager on Brazil’s neck-and-neck presidential race, Israel’s election deadlock and the open question of who will control the US Congress. In October 2022, before the last midterms, trading volume on Polymarket and Kalshi combined reached just over $4m, according to data from DefiLlama. Over the same period before the 2024 presidential election, it surpassed $90m. This October: $21bn. A monthly record.
Prediction markets allow traders to buy and sell binary contracts on real-world events: will this thing happen, yes or no? As contracts change hands, their prices fluctuate, producing a continuously updating estimate of the probability of an outcome.
A few years ago, prediction markets were basically fringe betting sites. But under Donald Trump – whose pick for CFTC chair, Mike Selig, a former lawyer for Polymarket and Kalshi, has fought to bring and keep the industry onshore – prediction markets have emerged from the fringes of finance. Event contracts “help the public forecast, plan for, hedge, and even harness” perceptions of future outcomes, and can “sometimes forecast outcomes better than polling”, the CFTC states on its website.
Out of exile, the platforms have enmeshed themselves in public life. Newsrooms including the AP, CNN, the Wall Street Journal and CNBC have signed deals to integrate live election odds into their coverage. In a recent interview about the midterms, the former congressman Brad Carson cited one candidate’s “99% chance to win on Polymarket”. He made no reference to polls.
Polymarket is “the most accurate thing we have as mankind right now, until someone else creates some sort of a super crystal ball”, its founder and chief executive, Shayne Coplan, told 60 Minutes last year.
A run of misses
Prediction markets have a record of getting things right. An analysis of Polymarket transactions over two years found that markets priced at about 80% resolved “yes” roughly 80% of the time. Across the full range of prices, outcomes rarely strayed more than a few percentage points from the probabilities the prices implied.
But in several recent high-profile elections, markets have gotten things wrong.
The same day as Brazil’s runoff, Serbia held a snap parliamentary election. The day before, Polymarket gave Aleksandar Vučić’s United Serbia a 70% chance of winning. On election night, the opposition Student List, born out of nearly two years of anti-corruption protests, took 52.6% of the vote. Vučić conceded, securing just 39.2%.
In August, Polymarket and Kalshi gave Francesca Hong roughly a 95% chance in the Wisconsin Democratic governor’s primary. Her rival, David Crowley, won by less than half a percentage point. Polymarket then deleted an X post saying Hong had a 96% chance.
On the eve of the Los Angeles mayoral primary in June, both platforms gave the reality TV star Spencer Pratt about a 75% chance of advancing to the runoff. He finished third.
In March, on the morning of Illinois’s Democratic Senate primary, Kalshi priced Raja Krishnamoorthi – who had spent nearly $30m on the race – at 61%. His opponent, Juliana Stratton won, despite the crypto industry, which is closely tied to prediction markets, spending nearly $10m opposing her campaign.
The day before New York City’s 2025 Democratic mayoral primary, Polymarket gave Andrew Cuomo a 66% chance of winning. Zohran Mamdani beat him by 12 points.
Why did a “super crystal ball” diverge so drastically from polls on these occasions?
Firstly, it’s important to consider that prices and election polls measure different things. The former captures the odds that a candidate will win on election day, whereas polling margins reflect a candidate’s lead. For that reason, Polymarket percentages often accentuate small polling gaps. For instance, latest polling data from FiftyPlusOne shows a 2% margin between Pete Ricketts and Dan Osborn in Nebraska’s Senate race, but Kalshi and Polymarket have a Ricketts win at about 70% as of Monday afternoon. In Ohio’s governor race, polls show Vivek Ramaswamy within a few points of Amy Acton, whereas Kalshi and Polymarket have an Acton win at 82% as of Monday afternoon.
So, in cases like Brazil’s election, Polymarket’s 80% forecast of Bolsonaro victory looked wildly inaccurate compared with a two-point polling margin.
Prediction markets became popular in 2024 precisely because they saw what pollsters couldn’t. Polymarket and Kalshi predicted Trump’s victory weeks before results emerged, all while polls projected a toss-up. But researchers note the US presidential election wasn’t necessarily proof of the crowd’s wisdom.
To assess whether prices accurately and efficiently aggregate political information, researchers at Vanderbilt University analysed 2,500 political prediction markets during the final five weeks of the presidential race involving more than $2.3bn in transactions. The study found markets “often behaved in ways inconsistent with rational information aggregation”. Accuracy fell to 67% and 78% on Polymarket and Kalshi respectively, while prices “diverged sharply across markets, exchanges, and contract types”.
Deciphering why markets get things wrong in some elections, and right in others, can be challenging. Pollsters can scrutinise their sampling methods; on prediction markets, traders are anonymous and their reasoning largely opaque. Experts say it comes down to the breakdown of market volume between retail traders – casual, recreational users who use a vibes-based approach to betting – versus a skilled minority of full-time traders who build statistical models incorporating vast datasets. Growing research suggests that accuracy across the board owes less to the wisdom of the crowd than to a small minority of skilled traders.
The idea of the “wisdom of crowds” is commonly traced to the statistician Francis Galton, who in 1907 observed that the average of hundreds of guesses about the weight of an ox at a country fair came remarkably close to its actual weight. Aggregating many independent judgments, the principle goes, can beat individual experts.
But prediction markets don’t really work like that. A recent study found that just 1% of Polymarket accounts captured three-quarters of profits since 2022, with the top 0.1% taking half. Two-thirds of accounts have lost money. At the top are the so-called “sharps”: skilled traders who hunt for mispriced odds, exploiting advantages in knowledge, data and speed – and potentially insiders. At the bottom are recreational bettors, known to sharps as “dumb money” or “fish”, who often trade on instinct.
One Polymarket trader, a mathematics PhD known as “@gopfan”, has made nearly $2m over the past two years using statistical models to forecast elections around the world.
It is less the wisdom of crowds than “the wisdom of the informed”, as Theis Ingerslev Jensen, an assistant professor of finance at Yale University, puts it. A study he co-authored this year estimates that just 3% of Polymarket accounts belong to skilled traders, whose activity drives prices towards more accurate probabilities.
That model works in a quiet market, with a few well-informed gamblers. It breaks down in a loud one. When a high-profile election arrives, a wave of recreational money floods in – often from people treating a bet as a show of support rather than a prediction. The data-driven sharps can push back, but only with the capital they have. Once that is committed, nobody is left to correct the price.
High-profile elections attract “extremely, extremely high” volumes of recreational trading, says @gopfan, making the odds “sometimes pretty ridiculous”. Brazil was a case in point: “The prices were staying extremely distorted because all the sharp money was already allocated into the market. There was no one to move it back.”
Brazil’s crowd also had reasons to be confidently wrong. Bolsonaro won the first round by 47% to 45%, and since the two-round system began in 1989, no candidate had overcome a first-round defeat to win the runoff. His odds surged, suggesting traders may have treated a precedent drawn from a handful of elections as an iron rule. They may also have underestimated the significance of 33 million first-round abstainers, who leaned towards Lula. And Brazil’s April ban on prediction markets may have left the market dominated by foreigners, less attuned to local politics – though bans can be circumvented with VPNs, as Polymarket users in the US did until a year ago.
Why, then, did the markets correctly favour Trump in 2024, when a flood of new traders appeared to treat their bets as proxy votes? “Sometimes stupid people just get lucky,” says @gopfan. “They win just because they are randomly right. But it’s nothing more than a pure coincidence.”
“They were lucky to some extent, right?” Jensen agrees.
The flood also explains which way the markets lean. One trader who profited from the 2024 influx, @yungstalin, has made nearly $1m on Polymarket. Shortly after joining, he noticed a “very clear pattern” among new accounts: they were Trump supporters. “The vast majority of recreational bettors were just dumping money on Trump, and making the most basic mistakes,” he says. His strategy was simple: “Just follow the dumb people and bet against them.”
In 2024, the dumb money happened to back the winner. In Brazil, Serbia and Los Angeles, it did not. When it comes to using Polymarket to judge the midterms results, the wisdom of the crowds really depends on the wisdom of the individuals.
