Three ways in which prediction markets are oversold

Don't get me wrong, I think these are great tools for aggregating a weighted average of people's beliefs. This is a valuable tool. Most of the time they will provide strong signal, clearly indicating the information they are designed to indicate, that is, a conviction-weighted average of beliefs about a subject.

But I'm hearing a few boosters go way beyond that and say things like these are using markets to learn truth. They are only using markets to learn truth when insider trading is faithfully used (caveats discussed elsewhere). But most of the time they are not actually bringing in new information. It's weird to call them "truth machines."

1) This not Hayekian Price Discovery.

Even if you don't identify as an Austrian Economist, it's good to be familiar with one of the seminal pieces of the 20th century economic literature, Hayek's 1945 paper "The Use of Knowledge in Society".

Hayek argues eloquently that the use of markets for determining, say, the price of tin, is aggregating and combining massive amounts of economic data which is not publicly available. Thus the market itself carries useful information that would not be available to a central planner. This is unambiguously true and nobody should be questioning this in 2026. (It needed to be argued in 1945, because, who knew, maybe, centrally planned economies with rigid prices controls could have been the way of the post WWII future? By now we fully understand that markets do perform a lot functions quite well.)

Each individual economic agent is responding in real time to his or her economic conditions, their preferences, availability of alternatives, constraints, etc – and by virtue of their participation in the economy useful information is being dispersed from node to node in the decentralized economic network which includes everyone. One important thing to note here is that the information is propagating through the medium of real economic activity, which has high fidelity. It's pure signal and it's transferred through the network like electricity is through a circuit.

The other key thing here is that each agent is acting on knowledge (it's in the title of the paper) not guesses or beliefs. If your business demands you spend more money on copper and less on grain, you just do that, you don't try to guess what people are the world are also spending on copper or alternatives.

Prediction markets distinctly do not do this. Yes, there is a market involved – but it's not tethered to economic reality in the same way. You don't run out of Seahawks Super Bowl Futures and scramble to find an alternative to put in your product to make it marketable. You don't decide to start producing more Marco Rubio Presidential Futures because you find they are cost effective to grow on your land.

At the end of the day, prediction markets are based on guesses, and while the guess might be made based on your local information – it's not passed throughout the network in an economic node to economic node way, it's all communicated via a centralized market of guesses.

Yes, there is "skin-in-the-game" which will select for those with more conviction. But this does not make the market essentially Hayekian. It's still just an ensemble of guesses.

There's a similar conflation that goes on with blockchains – touching a blockchain does not makes something truthful or contain unforgeable information. Touching a market does not give something market-driven fidelity in the same way that the price of tin does.

Galton's Ox.

When speaking of "Wisdom of the Crowd" people are often referring to the phenomenon in which the mean or median of guesses tends to be surprisingly accurate. The famous example was demonstrated by the statistician Galton who had 787 people estimate the weight of a 1182 lb ox - the average was off by 1 lb.

What this proves is not that prediction markets are magical, but rather, the market itself is almost unnecessary. There is no magic Hayekian invisible hand guiding the market to the true solution- the underlying principle is that when you aggregate guesses, you are likely to get better results than any individual guess.

At the end of the day, a prediction market is just Galton's Ox with an extra weight on conviction. What the market is actually accessing is Galton's Ox, not unrevealed individual information in the way that Hayek describes.

2)Arbitrage on future markets is massively different from arbitrage in spot markets.

Suppose you believe the market has been deliberated manipulated to make it appear that the probability of an event ia 10% higher or lower than you believe it is in reality - Naively one might respond "Well, certainly someone will arbitrage that!"

The math is not the same: Yes, you may be getting a future that is "worth" $0.50 for $0.40 – but if the market is buying and selling that future for $0.40, you can't just execute some trades and lock in a profit. You can make a trade which has positive expected value (or you believe does), which naively might seem attractive, but there's many reasons why one might not do this.

Simple example: The coin flip on next years Super Bowl. If someone offers you a Heads future on this flip for $0.40, how many do you buy? Obviously you wouldn't bet your house – there's a 50% chance you lose all the money you wagered.

Hedge funds think similar way. They aren't maximizing expected value with each bet - this leads to ruin in the short term, rather, they are maximizing expected growth rate over time. A simple calculus computation, introduced by a mathematician John L Kelly in 1950s, gives a formula for how much someone should bet if they are making a series of bets over time. This formula is called the Kelly Criterion.

Similar calculus determines that a hedge that is trying to grow their wealth over time is going to try to make as many short-term uncorrelated bets as possible, bets in which the hedge funds believes they have some expected positive returns. If the expected returns are small and uncertain, it's not worth their time.

Suppose a market for a political market opens, and a certain candidate, say even before polls, is given a 10% chance of winning the primary. Boosters for the politician dump thousands into the market and move it 25%. Will hedge funds arbitrage that? It's not so clear. Even if they have really good information to suggest that the real probability is 10%, it might not be their long term best strategy to put aside a lot of capital – if they lose they lose big.

3)Bad comparisons and bad statistics

Some misunderstanding of statistic is forgivable; some is just dumb.

To say that prediction markets have performed "better than polls" is really a nonsense statement that should never be made by a statistically literate person.

This is like saying that fresh baked bread tastes better than stalks of wheat. Of course it does.

To begin, the comparison doesn't even make sense: Polls don't produce predictions, they produce data. Analysts (Nate Silver, Sam Wang, or whoever) make prediction, and use the polls as the ingredient. If you see news that a poll shows that 60% of respondents will vote Yes on a measure and believe this means there's exactly 60% of the measure passing; please seek mathematical help. If you see this news and wonder "Was this a big sample? How does this poll usually track" you're asking the right questions.

Each poll is an uncorrected raw data point. Anybody with any sort of numerical literacy is going make some attempt to average the polls in a meaningful way. It doesn't take a market to do this. 1992 Microsoft Excel should outperform most poll.

The other thing is that anything 40-60% range is a toss-up, more or less, in the sense that you would have to run dozens of these events repeatedly to determine with any sort of statistical confidence whether a 51% or 59% prediction is more correct. One outcome tells you almost nothing about whether a prognosticator who gave an event a 58% probability is better than someone who gave it a 54% probability.

Finally, let's talk about 2024. Polymarket gave it Trump a 57% chance the day of the election - after being up and down the weeks leading up to that. The actual results were not at all close. A estimate 57% was a massive miss. Could you really look at the results and believe that the market knew something, and think that if we had that election again on a random day of the week, 3 of the 7 days would produce Kamala as the winner?

Trump massively overperformed most estimates, including the prediction markets.

If prediction markets really had good insights into what was going on, the number should have been much higher.