Who’s going to win the presidential election?

My answer, as a journalist writing about politics who gets asked this a lot, is always, “It’s a coin flip.” That’s certainly what forecasting models based on polling suggest. As I write this, the Economist gives Donald Trump a 56 percent chance; FiveThirtyEight gives him a 51 percent chance; Nate Silver gives him 53.1 percent. That’s not exactly 50-50 but it’s pretty close.

But this year, a different way of judging the odds has become more popular than ever: prediction markets. These are real-money markets where people can bet on the ultimate victor of the presidential election, among other elections and events, from politics to sports to movies. While academic-sponsored markets like PredictIt and the Iowa Electronic Markets have been around for a while (decades in the latter case), two major new markets have emerged this cycle.

One is Kalshi, the first non-academic prediction market to be officially legal in the United States after the company beat regulators in court who tried to block it from allowing betting on elections. Right now, their market, with over $50 million bet in total, indicates that Trump has a 59 percent chance of victory.

But Kalshi is comparatively tiny next to Polymarket, an all-crypto market accepting bets from all over the world. It’s presidential market has over $2.3 billion invested and counts Elon Musk among its fans. (Technically, Polymarket disallows Americans from betting on it, but any moderately tech-savvy person with a virtual private network can get around that. As of Wednesday, it has started cracking down on US trading a bit.) Right now, it claims higher odds for Trump than anywhere else: 61.9 percent.

Like a lot of economists and political scientists, I’ve long been a big fan of prediction markets. They provide a useful complement to polling by summarizing conventional wisdom about candidate odds, and also function as a tax on bullshit. There’s a lot of cheap talk in political punditry, and I’m generally of the belief that if you say things like “Trump will definitely win Florida by 8 points,” good etiquette requires you to bet money on that proposition. Making the bet means you’re putting your own money behind your prediction, and if you have to do that, you’re probably going to make fewer garbage predictions.

But these markets haven’t had a test like 2024 before, and seeing them operate with billions behind them is giving us a sense of how they’ll work at scale — including to what extent they can be manipulated to produce a certain outcome. The big question: Can we trust these things?

Prediction markets for beginners

People have been betting on elections for centuries. Henry David Thoreau memorably noted in “Civil Disobedience” how often betting accompanied voting. The economists Paul Rhode and Koleman Strumpf have studied presidential election markets in the US that ran at large scale from 1868 to 1940; “betting activity at times dominated transactions in the stock exchanges on Wall Street,” they find. In 1916, the peak year of these Wall Street markets, betting reached $290 million in today’s dollars.

These markets do not work like typical sports gambling. Your neighborhood bookie (or, these days, FanDuel or DraftKings) sets odds on her own, obviously consulting what other bookies are setting but fundamentally deciding for herself what to charge. When you bet with her, you are betting against them, not against other bettors, and a good bookie will set odds such that they consistently profit.

Polymarket, Kalshi, PredictIt, and other prediction markets are genuinely markets: rather than betting against the house, you are buying “shares” that pay out if a particular event happens in the future. The market sells the shares to start, but then the trading is done between traders, not between traders and the house. The market-maker exists to broker transactions. That means odds can change much more quickly and fluidly than with a normal bookie. It’s more like a futures contract than a sports bet.

This structure does mean you have to be careful looking at these numbers. When you look at a site like Polymarket, it’s tempting to look at the headline numbers (hmm, 61.7 percent for Trump) and assume that this represents what bettors on the site think the odds of a Trump victory are.

What it really means is something like that, but a little bit subtler. As the pseudonymous finance blogger Quantian explains, the real purpose of a prediction market is to reach an equilibrium at which the price of a “share” in a given candidate matches demand for those shares. This equilibrium can mean the price exactly matches what people in the market think the probability of that candidate winning is, which is what you’d want the price to be if the markets are to be useful for forecasting. But those two can also diverge, especially if the market is restricted in various ways.

That said, economists Justin Wolfers and Eric Zitzewitz have found that, in practice, prediction market prices are close to participants’ aggregated beliefs, and you can set a kind of confidence interval and be reasonably sure that participants’ beliefs are within it. Zitzewitz notes that this is more of a problem with markets like PredictIt, which caps how much investors can put in at a relatively low level. Those rules can lead to major mispricings because they make it much harder to bet against low-probability events. “In a prediction market where you don’t really have those constraints,” Zitzewitz noted in a phone call, “then we’re much more likely to get a price equal to some average of beliefs.”

Meanwhile, be careful to never equate prediction market prices with polling results, as one since-corrected New York Times article did:

Polls do not give you a “chance of winning”; for that you need an actual model like FiveThirtyEight’s, the Economist’s, or Nate Silver’s.

The basic theoretical case for betting markets is the same as the case that normal markets are efficient. If shares of, say, Nvidia are mispriced, then you can make money by betting the price will move. In a market where millions of people leveraging trillions of dollars are all acting that way, there probably won’t be too many obvious mispricings. Any ones that once existed are quickly exploited by some investors to make money.

That’s a nice theory, though one challenged by the emergence of “meme stocks” whose prices seem totally out of whack with their actual value. The better case for prediction markets is that they’ve worked well in practice. Wolfers, Zitzewitz, and Erik Snowberg reviewed the evidence in a 2012 paper, and it’s fairly compelling. When it comes to forecasting economic outcomes (like economic growth or inflation), “macro derivatives” (a kind of prediction market for these outcomes) do as well or better as surveys of professional forecasters.

A paper by David Rothschild looking at markets during the 2008 presidential and Senate elections found that early in the election, prediction markets were significantly more accurate than polls; closer to the election, they were roughly equal. A 2008 paper by a team at the Iowa Electronic Markets, the longest-running prediction market in the US, found that the average error of polls (1.91 points) was higher than for markets (1.58), even late in the cycle.

A lot of companies have also adopted internal prediction markets to help make decisions. Zitzewitz and Bo Cowgill examined markets at Google, Ford, and an anonymous third company. These markets covered topics like “demand, product quality, deadlines being met, and external events.” Zitzewitz and Cowgill then compared the markets to the predictions of internal experts; the average error of the markets was 25 percent smaller.

In a very different context, economists Anna Dreber, Thomas Pfeiffer, Johan Almenberg, and Magnus Johannesson set up prediction markets where psychologists could make bets on whether specific pieces of psychology research would replicate when other researchers tried to reproduce it. They found that the markets were effective at predicting which studies would replicate and outperformed simple surveys of experts.

Economists Lionel Page and Robert Clemen marshaled evidence from nearly 1,800 prediction markets for a 2012 paper. They found that markets looking a year or more in the future were deeply flawed; they usually didn’t have enough traders to generate a price or they had systematically biased prices. But markets became increasingly well-calibrated as they got closer to the event in question.

Note that presidential elections are maybe the hardest context for judging prediction markets. If a market is “well-calibrated,” then, when looking at all the dozens or hundreds of markets it’s run, markets where an outcome has a 70 percent probability should see that outcome occur 70 percent of the time; markets where it has 20 percent odds should see that a fifth of the time; and so on.

To do that kind of analysis, you need a lot of markets and predictions. Presidential elections are one-offs, and aside from the Iowa markets, there haven’t been modern election-betting markets going back for more than a handful of elections. Polymarket, for instance, is only on its second presidential cycle, so we have no way of knowing if it, specifically, is well-calibrated in predicting presidential elections specifically.

Are the markets being manipulated?

So prediction markets are fairly accurate in general. Why are some people so opposed to them, then?

In the US, the main opposition to betting markets like Kalshi has come from a handful of Democratic senators led by Jeff Merkley (D-OR). “Billionaires and large corporations can now bet millions on which party controls the House or Senate and then spend big to destroy candidates to protect their bets,” Merkley bemoaned after the court ruling allowing Kalshi to operate election markets.

The story of the “French whale” on Polymarket seemed to give credence to these fears. About four accounts on the market (Fredi9999, PrincessCaro, Michie, and Theo4) have pumped about $45 million into bets on Trump since the beginning of September. That’s not a ton in the scheme of a market of over $2 billion, but it’s enough to raise eyebrows. The Wall Street Journal and Financial Times have written about it, but the best reporter on the topic is the pseudonymous Domer a.k.a. JustKen (he uses a shot of Ryan Gosling in Barbie as his avatar), another major Polymarket trader who was curious just who he was betting against.

Domer did some digging and linked the four accounts to a Frenchman named “Michel;” it seems like they were being controlled by the same person. “My best guess is it is a wildly risky-loving uber-wealthy Frenchman who is pretty damn sure that Trump is going to win,” Domer concluded.

But it could also, in theory, be an uber-wealthy Frenchman who is trying to manipulate the market to make Trump’s odds go higher. This is an often-hypothesized problem with prediction markets: someone could rush in with money to boost a particular candidate’s odds, which could then lead to media coverage and public perceptions that think the candidate is a favorite, which could in turn make them the favorite.

The nature of the markets is that if you do this on one market, you will probably have effects on all markets. If Polymarket gives Trump 65 percent odds and Kalshi gives him 55 percent odds, you can make risk-free money by buying Trump at Kalshi and Harris at Polymarket; if you only have to spend 55 cents to get a dollar if Trump wins, and 35 cents to get a dollar if Harris wins, then you can spend 90 cents to get a dollar if either Trump or Harris wins (and there’s basically 100 percent odds that one of them will win). This is called arbitrage, and it tends to close gaps between the markets over time.

A common argument from supporters of prediction markets is that they’re resistant to this kind of manipulation. If someone injects a huge amount of money into a market to make it look a certain way and that appearance is at odds with the underlying reality, then there’s money to be made taking the other side of that bet. That’s especially true on sites like Kalshi and Polymarket, which have larger investors, like hedge funds.

If I think Harris has a 60 percent chance of winning and Polymarket thinks it’s 35 percent, I’m not going to sink my life savings into Harris contracts. Sure, that bet is profitable in theory, but there’s still a 40 percent chance that I lose everything. But hedge funds exist more or less entirely to make bets like that, and have much deeper bankrolls and risk tolerances. That means they can help prevent manipulators from swinging markets. Sure enough, around 10 pm ET on Tuesday a single trader ​(Ly67890) bought over $2.1 million in “Harris wins” shares, indicating that at least one counter-whale has emerged to take the other side of the bet.

Past attempts to manipulate markets have tended to end badly for the manipulators. In 2012, a “Romney whale” who spent heavily trying to prop up Mitt Romney’s odds in prediction markets wound up losing $4-$7 million. 2008 saw a similar attempt to prop up John McCain’s odds, which led to a crackdown from the now-defunct market Intrade. In Domer’s words, “those people got BTFO” — blown the fuck out.

Rhode and Strumpf, the economic historians, have done the most careful academic investigation of manipulation I’ve seen, and found, “In the cases studied here, the speculative attack initially moved prices, but these changes were quickly undone and prices returned close to their previous levels. We find little evidence that political stock markets can be systematically manipulated beyond short time periods.”

It’s impossible to know whether French traders are actively trying to manipulate the market, or simply have a genuine belief that Trump will win and are putting huge amounts of money behind that belief. I asked Rajiv Sethi, a professor of economics at Barnard College, Columbia University, who writes a newsletter on prediction markets, what could explain the markets giving higher probabilities than models like FiveThirtyEight or Silver. “There are two possible explanations,” he told me. “Markets could be absorbing information faster than models. They see stuff that could be moving the models, so the price rises, and then the model adjusts a couple of days later. … There’s an alternative explanation that the market just adds a premium to the model.” That is, the traders could just be a bit more pro-Trump than the models overall.

It’s very very hard to know in real time which of these views is right. Sethi has put together some very preliminary evidence by creating virtual traders who buy and sell shares on prediction markets based on what the models from FiveThirtyEight, Silver, and the Economist are saying. None of these traders, he finds, made money; all lost money, by about the same amounts on Polymarket, and by a larger amount for Silver than other models on PredictIt. “Does this mean that models are performing poorly relative to markets? Tentatively, yes,” Sethi writes. “But this could change quite dramatically—in either direction—over the next few days.”

The real promise of prediction markets

In some ways, I find presidential elections the most boring use case for prediction markets. Such elections are probably the single event for which we have the most popular interest in and information around predicting. We already have polling and very sophisticated polling-based forecasts that tell us a lot about the race. The website Pollyvote has a very nice rundown of the literally dozens of forecasts available based on polls, models like FiveThirtyEight’s and Silver’s, political scientists’ models, and so on.

Do we really need prediction markets on top of those? Maybe not.

But the general idea of using markets to predict hard-to-predict events has merit outside an election context. Synthesizing information from a diverse array of sources is hard, and prediction markets show a lot of promise at that task.

Take natural disasters. Events like hurricanes, typhoons, and earthquakes have huge humanitarian and economic effects, and while we have some statistical models that can predict them a bit, these don’t exactly provide actionable guidance for businesses, residents, insurance companies, and other people with a stake in disaster-prone regions. Getting decent markets in place could help businesses prepare for supply-chain disruptions and give residents a strong signal that they need to get the hell out, or at least invest in more resilient housing.

The issue is that betting on where is going to have a hurricane next, and how bad it’s going to be, feels a bit ghoulish (though it’s not really any more ghoulish than the odds that actuaries have to sort through every day). It’s certainly not fun, like betting on sports or the presidential election.

I fear that means that the most high-value prediction markets might wind up not only not making their brokers money, but will have to be subsidized: insurance companies, say, would pay for a market where meteorologists help sort through hurricane odds, or the National Institutes of Health could subsidize a market where medical researchers bet on the results of clinical trials, giving them a better sense of which drugs seem most promising and worth investing in.

To date, prediction markets have been driven by two main forces: a sober-headed assessment of ways in which they can be a useful tool to understand the world and pure degenerate gambler instinct. As someone with a bit of the latter, I totally get how it’s fun. But if prediction markets are going to be of real social value, we need more of the sober force driving things as well.