What prediction markets can't do

For four months this series has argued one thing from every angle: a market price is a probability, and a liquid market is the best forecaster we have. I believe that. But a case you can't argue against isn't a case — it's a sales pitch. So this is the chapter where I tell you where these markets break.

They are not magic. A prediction market is a machine with a known operating envelope: it does its remarkable thing only when a few conditions hold, and it degrades — sometimes gracefully, sometimes catastrophically — when they don't. Knowing the envelope is the difference between a tool and a toy. Here are the five places the machine fails: thin markets, long horizons, reflexivity, unresolvable questions, and markets that should never have been built.

1. Thin markets

Everything good about a prediction market comes from people trading against each other in volume. Take the volume away and the magic goes with it. Outside a handful of marquee markets — a presidential election, a Super Bowl, a Fed decision — most contracts are thin: a few traders, sparse orders, a wide gap between the best bid and the best ask.

A thin market is not a quiet version of a good one. It's a different, worse object. The price is noisy, jumping on a single small order. The spread is wide, so the "probability" is really a fuzzy band — is it 40% or 55%? The book can't tell you. And it is cheap to manipulate: when a few hundred dollars moves the line, anyone with a motive and a wallet can paint whatever number they want, at least for a while. This is the same depth problem I pulled apart in can you trust the price? — a price you can move with pocket change is not a forecast, it's a suggestion.¹

The hard part is that thin is the default. The famous, deep markets are the exceptions that make the news; the long tail of niche contracts mostly lives in the noisy regime. "The market says 31%" means something very different on a market with millions of dollars of depth than on one with four traders and a dream. Always ask how deep the book is before you trust the number on it.

2. Long horizons

Prediction markets are sharpest about the near future and weakest about the far one. The reason is money, the same as always — but here money works against accuracy. To hold a position, you lock up capital until the market resolves. On a question that settles next month, that's nothing. On a question that settles in 2040, you've tied your money up for fifteen years, and every one of those years has an opportunity cost: whatever you'd have earned investing it elsewhere.²

So long-dated markets distort in two ways. First, almost nobody shows up — who locks money away for a decade-and-a-half to be right about one binary outcome? — so these markets stay thin, with all the noise that brings. Second, the price stops being a clean probability. A "YES" share that pays one dollar far in the future is worth less than a dollar today, because a dollar today can earn interest in the meantime. The further out the resolution and the higher the interest rate, the more the price is pulled below the true probability by simple discounting. The market isn't telling you only what it believes; it's telling you what it believes, dragged toward zero by the time value of money.

3. Reflexivity

The first two limits are about thin or distant markets giving you a bad number. Reflexivity is stranger: it's when a perfectly liquid, well-built market gives you a number that changes the very thing it's measuring. A thermometer reads the room without warming it. Some prediction markets are thermometers that turn up the heat.

Take a market on "This bank fails — 80%." A price isn't a private guess; it's public, and it moves behavior. Depositors see 80%, panic, and pull their money — which is exactly the run that makes the bank fail. The forecast didn't observe the future. It caused it. That's a self-fulfilling loop. The mirror runs the other way too: a market on "X gets fired" at 75% becomes a public pressure gauge that nudges the board toward firing them, or — if they fight it — toward keeping them out of spite. Either way the forecast is now an input to the decision it claims to merely predict.³

This is George Soros's old idea, dressed for a new venue: when participants' beliefs feed back into the system they're betting on, the clean separation between observer and observed collapses.³ "Price equals probability" assumes the market watches the world. Reflexivity is what happens when the world watches the market back. It doesn't break every contract — most outcomes don't care what a trading screen says — but it quietly poisons exactly the high-stakes, decision-relevant questions a prediction market would be most tempting to point at.

4. Unresolvable questions

A market is only as good as its settlement. The whole edifice — price as probability, money as discipline — rests on one promise: when the dust settles, everyone agrees who won. Take that away and there's nothing to anchor a price to. I spent a whole chapter on this in who decides what's true?, and it's worth restating as a limit: if a question can't be cleanly and verifiably resolved, it can't be a good market.⁴

"Will it rain in Hanoi tomorrow?" resolves itself — a gauge, a number, done. "Was the policy a success?" does not. Success by whose measure, on what timeline, judged by whom? A contract on a vague, contestable, or unfalsifiable claim doesn't aggregate information; it aggregates an argument about what the words meant, and it settles in a dispute. Plenty of the questions we most wish a market could answer — Was this the right call? Did this idea actually matter? — are exactly the ones with no clean resolution. The mechanism needs a fact at the end. Where there's no fact, there's no market.

5. Markets that shouldn't exist

The last limit isn't technical — it's moral, and it's the one I take most seriously. A market pays people to be right. That is its genius, and on some questions it is monstrous. A market on "Will this named person die this year?" doesn't just forecast a death; it puts a bounty on it and hands a cash motive to anyone positioned to help it along. Reflexivity, again — but now the perverse incentive is the whole point. Some markets shouldn't exist because the act of pricing the outcome corrupts it.

This isn't hypothetical, and it isn't new. In 2003 the US Defense Department, through DARPA, floated the Policy Analysis Market — quickly nicknamed the "terrorism futures" market — where traders would bet on assassinations, coups, and terror attacks in the Middle East, on the theory that the price would aggregate intelligence. The public reaction was immediate and bipartisan revulsion: senators called it grotesque, the idea of a market that rewarded predicting an attack — and rewarded anyone with an incentive to cause one — and the program was killed within days of becoming public.⁵ The mechanism worked exactly as designed. That was the problem.

The edges aren't the refutation

So: thin markets give you noise. Long horizons give you a discounted, half-empty book. Reflexivity turns the forecast into a cause. Unresolvable questions have nothing to settle on. And some questions are ones we should refuse to price at all. Five hard edges — and none of them is small.

But notice what they are. Every one is a boundary of the same machine, not a flaw in it. The bull case from the last four months doesn't live in spite of these limits — it lives inside them. The markets that compound (and yes, today that's mostly sports — roughly 80% of the volume since mid-2024⁶) are precisely the deep, near-term, cleanly-resolved, non-reflexive ones: who wins on Sunday, where rates land next month, which candidate takes the state. That's the sweet-spot corner of the first figure, and it's a vast, real, fast-growing place. The edges don't shrink the category. They draw it. A tool you understand the limits of is the only kind worth building a business — or a belief — on.