Earlier this summer, I sat barefoot in Berkeley with a dozen aspiring “geopolitical traders,” learning how to make money from war. We were at a conference billed as a “festival for predictions, and markets thereof” – that is, a celebration of prediction markets. The venue, a hippie-ish compound, mostly forbade shoes indoors. Our instructor, Mike, claimed to have grown his money eightfold on Polymarket, mostly by trading on politics. Now he would share his wisdom.
Prediction markets were legalized in the United States in 2020 on the premise that they would “allow us, more often than not, to predict the future.” That, at least, is what Kalshi’s marketing copy claims. Reportedly valued at $40 billion, the firm runs the first exchange of its kind to win US regulatory approval. But Mike soon made clear that our predictive abilities had little to do with our ability to succeed on Kalshi or Polymarket.
Mike had a theory. “It would be in the interest of the US government to actively interfere in the information space to distort market prices,” he told us. Donald Trump, he noted, had repeatedly – and baselessly –promised an imminent deal to end the war in Iran, a claim prediction markets had largely taken at face value. The volatility now rippling through these markets, he suggested, might well be “actively planned and executed by the US government.” When the President is “abusing his position of trust,” what should a prediction-market trader do? “We can adapt,” Mike said, if we “embrace the slop.”
The markets are helping officials to quantify how effective they are at polluting the information space
Our forecasts are getting worse, he explained, “because the input we put into the forecast is also lower quality.” Which is to say: we should become “swing traders,” putting money on Trump’s every word no matter what we actually believe the outcome will be. I asked Mike whether “embracing the slop” defeated the social purpose of prediction markets – namely, generating reliable forecasts. He suggested they might soon find a new use: quantifying, for officials, how effectively their statements have polluted the information space.
Early last month, when moderate liberal David Crowley upset Democratic Socialist Francesca Hong in Wisconsin’s Democratic gubernatorial primary – an outcome Kalshi had priced at just 4.6 percent – I thought back to my exchange with Mike. That night, the markets had clearly gone haywire. After polls closed, Hong’s odds on Kalshi plummeted to roughly one in five, then spiked back to about 50 percent, briefly retaking the lead. Over the course of the evening, Hong’s and Crowley’s odds flipped nine times, five of them in a single 20-minute burst. Why might that be? Well, to borrow Mike’s phrase, users were trading the slop.
Traders “did not have the full read of what was going on when they were making those trades,” Lakshya Jain, the head of political data at the publication the Argument and the co-founder of the political-data site SplitTicket, told me. “They were just trading based on who was gaining in a drop, without looking at where the drop was from and whether the candidate outperformed” – or fell short of – expectations. On election night, candidates’ Kalshi odds are probably more predictive than throwing darts blindfolded, Jain told me. But these markets “should not be treated, especially on election night, as determinative, as predictive, or as especially well-calibrated.”
Can a prediction market ever be wrong? Not in any single case, according to Tarek Mansour, the chief executive of Kalshi. After Crowley upset Hong in Wisconsin, Mansour took to X to explain why the episode should not be regarded as a failure. A 5 percent likelihood of an event occurring, he explained, means the thing should happen one time in 20. If 5-percent candidates never won, the markets would be broken, too.
He’s right, of course. When prediction markets assign a political event a 60 percent chance of occurring, that event tends to occur about 60 percent of the time (at least for repeatable events, such as the release of government statistics or President Trump’s use of specific words in speeches). According to Kalshi’s own study of the topic, this pattern tends to hold three months, one month, one week and one day out from the event itself. (Still, Kalshi has not released data on the calibration of events in the “international category,” where geopolitical trading occurs.)
Then again, in the case of Wisconsin, Mansour is missing the point. Did the fundamentals of the race shift enough to justify reversing odds nine times? Or were traders merely gambling? However confident they may publicly appear in the technology, Wall Street thinks it’s probably the latter. “This whole slouching toward monetizing every outcome you can think of is basically no different than gambling online,” a Citigroup executive told me. The financial products that prediction markets sell are no different from existing predictive financial products, he told me. Consider, for instance, oil futures: “At the end of the day, how is it different to say that I’m betting on the price of oil than I am betting on the outcome of a presidential election?”
The difference, the executive told me, is your counterparty. Futures traders were “at least until maybe recently, sophisticated financial operators.” But prediction-market traders? Well, he’s squarely in their “target demo,” he says, which is “men under 40,” many of whom lack much financial or political knowledge.
Defenders of prediction markets often retort that, if their critics are wiser than the markets, then they should simply cash in. “Once someone notices that a prediction market has made a mistake,” writes the rationalist blogger Scott Alexander, “they’ll be incentivized to make bets in a way that corrects the mistake.” But when Kalshi and Polymarket collapse into slop trading, even the wisest trader has little way of correcting the unjustly high volatility.
Imagine yourself as a sharp trader on Kalshi or Polymarket the night of the Wisconsin gubernatorial primary. You buy Crowley and, when the race is called, you believe you’ll probably make money. Probably: if you believe Crowley has a 60 percent chance of victory, Hong still wins two times in five. So you bet only what you can afford to lose on a near coin-flip, and then you’re done. When traders’ next misreading of the returns sends Hong’s likelihood of victory to 80 percent, the market is offering an even better deal – but taking it means doubling down on the same trade, and a trader intent on a reasonable risk profile has little reason to add capital.
Correcting each swing would also require aggregating fresh information within minutes, in competition with counterparties who move quickly, precisely because they do not pause to verify their intuitions. Mansour is right, in his way: Crowley won, and 5-percent candidates should sometimes win. But nobody consults a prediction market to learn who won; the news tells you that for free.
‘How is it different to say I’m betting on oil prices than on the outcome of a presidential election?’
The point of prediction markets is to help elucidate ongoing events, and at this, on election night, they come up short. Then there’s the fact that the best interpreters of political data are barred by professional ethics from trading themselves. Jain cannot correct the prices; his firm, like nearly every other in the political-data business, forbids it. (He also cannot gamble, for religious reasons.) Professional observers’ views still move the markets, whether they intend to or not. “I’ve moved markets with my tweets, not intentionally,” Jain told me. “I will say something, and you’ll see that the odds will spike up for a certain candidate. I don’t like that, but I don’t control that.”
The most strident advocates of prediction markets argue that scattered traders with money on the line will, together, out-analyze any single expert – and faster than experts would. Per this theory, the markets should have been ahead of experts such as Jain, already pricing in what a competent reading of the drops would imply, so that his tweets arrived as stale news. Instead, the causality often runs the other way: on election night, experts’ tweets move the markets. In any event, calibration alone is not enough to establish that prediction markets are socially useful. A prediction market on a coin toss would look well-calibrated – bettors would never give heads much more than a 50 percent chance – but one hardly needs a market to predict that.
Similarly, a good deal of evidence suggests that prediction markets land in about the same place as more naive statistical models. At the moment, VoteHub’s “legacy” probabilistic estimate of the Texas Senate race – assembled from polls and fundamentals alone – gives Democrat James Talarico a 52.7 percent chance of winning. Kalshi rates him just 1.7 points less likely. The same model gives Abdul El-Sayed, the Democrat, a 64.8 percent chance of winning Michigan’s Senate seat; Kalshi gives him 64 percent. These markets mostly reflect conventional wisdom, which is also why they badly underpriced the winner of Florida’s recent Democratic Senate primary: Democratic Socialist Angie Nixon, whom Kalshi valued at 7.8 percent on the eve of her victory. There was little high-quality polling in her contest, and so little consensus for the markets to absorb.
The best-case scenario for Kalshi and Polymarket is for the money to get smarter as institutional investors pile in, seeking to buy bespoke insurance products – say, a hedge against the result of a presidential election – thus leaving the dumber money in the dust. Still, this seems unlikely. To entice institutions, prediction markets need deeper volume: even the $3.7 billion traded on Polymarket on the outcome of the 2024 presidential election is insufficient to provide hedging at a fair cost to all institutional parties with a stake in the outcome of a presidential election.
The best interpreters of political data are barred by professional ethics from trading themselves
An executive who reviews deals for a private investment group, and who was pitched on a Polymarket funding round, told me this leaves prediction markets in a catch-22. “You’re not going to get the volume just based on retail traders,” he told me. Retail bettors don’t wager enough on presidential races to be counterparties for institutions. Yet “to get the required liquidity” to hedge those institutions, “you kind of need the institutions there already.” He added: “Why are they going to join if the liquidity is not there first, right?”
For now, then, the markets will be reasonably well-calibrated when experts have already stated probabilistic views – or when insider trading occurs – and unreliable on everything else.
Back in Berkeley, Mike had some parting words of advice. The slop is here to stay, he believed, so we can learn to trade it. The approach wasn’t “that difficult, if you really pay attention,” Mike said, “because we actually understand how far the market swings.” Geopolitical markets, he explained, tend to lurch within a predictable range: say, “between, like, 30 and 60 percent.” So “you can just be in the range between, like, 30 and 60 percent,” he told us – buy “closer to 30,” sell closer to 60, collect the difference, repeat. You don’t actually need a view on the war itself; just a view of your fellow traders.
“You can probably make some money doing that,” Mike told us. “At some point, you’ll probably get screwed. But that’s just the reality of the markets.”
Comments