Prediction markets beat the experts on average, sure, but that stat hides more than it tells you. Accuracy swings hard depending on what's actually being predicted, and if you're using market prices as a decision input, knowing where they're sharp versus where they fall apart matters more than the headline number.
Where they actually work
The best case is an event with a big pool of informed participants and a clean, verifiable outcome. US elections are the obvious one. Millions of people follow US politics closely, plenty of them professionally, the outcome is binary, and everyone agrees on the resolution source, which is the official result. That combination is about as close to ideal as you get.
Major economic data prints are similar. GDP, unemployment, inflation, all of it has a deep well-informed pool, a clear resolution source, and standardized definitions. Markets on whether the Fed moves rates have consistently beaten both economist surveys and Fed funds futures.
Sports is the other strong one, high participation plus real domain expertise. Sports betting markets are among the most efficient forecasting mechanisms ever built. The closing lines are remarkably well calibrated, so a team priced at 70 percent to win really does win about 70 percent of the time across thousands of games.
Where they fall apart
The moment the participant pool stops having relevant expertise, accuracy drops. A market on whether some specific scientific paper gets retracted might pull decent volume, but most of the people trading it aren't experts in that field. The price ends up reflecting general sentiment instead of informed analysis.
Long-dated contracts are shaky too. A market on something three years out is pricing an outcome that depends on a pile of intervening developments. The price captures what's known today, but the world will have moved a lot by resolution. That doesn't make the price wrong, it just means its predictive value decays the further out you go.
Events with almost no precedent are the third weak spot. If a thing has never happened, there's no base rate to anchor to. Contracts on novel geopolitical scenarios or first-of-their-kind regulatory moves tend to show wide disagreement and poorly calibrated prices.
The thin market problem shows up again
Accuracy tracks liquidity closely. Deep markets pull in more participants, including more of the informed ones, and that pushes prices toward the right answer. Thin, low-volume markets can sit on a mispricing for a long time because nobody's willing to put up enough capital to fix it.
There's a selection effect baked in here. The contracts people cite as proof that prediction markets are accurate are the liquid ones, which happen to be the easiest for a market to get right in the first place. The long tail of low-liquidity contracts barely gets studied and is almost certainly less accurate.
Calibration versus discrimination
Accuracy has two parts and they're worth separating. Calibration means events priced at 70 percent actually happen 70 percent of the time. Discrimination means the market gives genuinely different probabilities to genuinely different events instead of parking everything near 50 percent.
Prediction markets tend to be well calibrated but sometimes poorly discriminating. Take biotech trial outcomes. The market might price most contracts between 40 and 60 percent because participants can't tell trials with very different odds apart. In aggregate those prices are calibrated, the ones at 50 percent do resolve about half the time, but they're not discriminating, when the truth is some belong at 30 and some at 70.
How I use this before trading a contract
Before I put money on a prediction market contract, I run the category through a few quick checks. Big informed pool? Unambiguous resolution? Historical data to anchor on? If it's yes across the board, the price is probably right and you need a genuinely strong contrarian read to justify a trade. If it's no, the price is more likely to be off, and your own analysis is worth a lot more. We built the same instinct into how Blockcircle surfaces prediction data, but you can do this in your head in about ten seconds, which is the whole point.