I keep two tabs open when I am looking at any event that both Metaculus and Polymarket happen to cover. Same question, roughly the same resolution, two numbers that are supposed to mean the same thing. And they almost never match. Sometimes the gap is a couple of points, which is noise. Sometimes it is fifteen points, which is a trade, or at least a reason to go read more.
The interesting part is that the two numbers disagree in a predictable direction most of the time. That predictability is the whole game. If the difference were random you could not do anything with it. Because it is structural, you can build a rule around it.
Why the two numbers are built differently
Metaculus is a forecasting platform. People submit probability estimates, the community median gets tracked, and a separate Metaculus prediction weights forecasters by their track record. Nobody is putting money down. The incentive is your reputation and your Brier score, so people tend to answer the literal question, think about base rates, and update calmly when new information shows up. What you get is a fairly clean read of what informed, reasoned opinion thinks the probability is.
Polymarket is a market. The number you see is a price, and a price is whatever the last person was willing to pay. That is a different object. A price carries things a forecast does not. It carries the cost of capital, because your money is locked up until the market resolves and that time has value. It carries fees and spread. It carries whatever it costs to actually get in and out at size. And it carries the psychology of people betting real money, which is not the same as the psychology of people logging an estimate.
So even if both crowds were equally smart, you would still expect the price and the forecast to sit in different places. They are measuring related but distinct things. One is an estimate of truth. The other is an estimate of truth after you subtract everything it costs to hold the position.
The skews that push the price around
A few forces bend Polymarket prices away from a clean probability, and knowing them tells you which direction to lean.
- Favorite-longshot bias. This one is old and shows up in almost every betting market ever studied. Longshots trade rich and heavy favorites trade cheap relative to how often they actually happen. People overpay for the small chance of a big payout and underpay for the boring near-certainty. So a contract sitting at three or four cents is usually still too expensive, and a contract at ninety-five is often a little cheap.
- Capital cost on the long tail. If a market does not resolve for a year, locking a dollar into the near-certain side to earn a few cents is a bad use of capital. Rational traders skip it, which leaves favorites underpriced. This effect gets worse the longer the time to resolution.
- Attention and flow. Prices move when a market is in the news and money floods in. Big directional flow can push a contract well past any reasonable probability for a while, especially on politically charged questions where people are betting a feeling.
- Resolution ambiguity. Sometimes the price is not wrong about the world, it is pricing in the risk that the market resolves in a weird or contested way. Metaculus does not carry that same tax because there is nothing to lose.
None of these are mysteries. They are just the standard reasons a market price is not a naked probability.
Using Metaculus as a fair-value anchor
Here is the way I actually use it. Treat the Metaculus community median as a fair-value estimate of the underlying probability. Not gospel, just a reasoned anchor. Then look at where Polymarket is trading and ask what would explain the gap.
If Polymarket is far above the Metaculus median on a longshot, that is favorite-longshot bias doing exactly what it always does, and the market is likely overpriced. If Polymarket is below the median on a near-certain outcome, that is often capital cost leaving the favorite cheap, and there may be value in buying it, as long as you are comfortable tying up money until resolution. Those two cases are the bread and butter.
A rough workflow that keeps me honest:
- Find a question both platforms cover with resolution criteria that genuinely match. This is the step people skip and it is the one that ruins everything. If Metaculus resolves on one definition and Polymarket on another, the gap is fake.
- Read the Metaculus median and, if it exists, the weighted Metaculus prediction. Note how many forecasters and how recent the updates are. A median from six stale forecasters is not an anchor.
- Compare to the Polymarket mid price, not the last trade. Check the spread and the order book depth so you know whether you could actually fill.
- Decode the gap. Is it favorite-longshot skew, capital cost, a news spike, or resolution risk? If you cannot name a reason, assume the market knows something you do not.
- Only act when the gap is large enough to survive fees, spread, and the cost of locked capital until resolution. A five-point edge on a market that resolves in a year is usually not a trade.
When the play-money number is the smarter one
The anchor cuts both ways, and this is where people get burned. Sometimes Polymarket is right and Metaculus is stale. Markets update in seconds. A forecaster community can take a day or two to move, especially on a fast-breaking story that lands overnight. If there is a sudden shock and Polymarket has repriced hard while Metaculus has not budged, the price is probably the better number and the gap is not an edge, it is a lag. Trading against it is trading against fresh information.
So the direction of the gap matters less than the reason for it. A gap caused by a durable structural bias, like longshot skew, is tradeable and repeats. A gap caused by one side simply being slow to update is a trap. The honest question every time is which crowd currently knows more, and on breaking news the money usually knows first.
One more failure mode worth naming. Thin Polymarket markets can show a price that no real size ever touched. If ten dollars set the last trade, that number is barely information. Always check depth before you treat a gap as real.
When I am comparing signals like this across a lot of markets at once, I lean on tooling rather than tabs, and inside Blockcircle I keep prediction-market feeds next to the rest of the signal stack so the same contract can be checked against other reads before I do anything. But the mental model is the part that matters and it does not need any software. Two numbers, built differently, disagreeing for reasons you can usually name. The forecast tells you roughly what should be true. The price tells you what it costs to be right. Most of the edge lives in the space between those two, and most of the losses come from mistaking a lag for an edge.