Mean Reversion in Prediction Markets: When a Price Is Just a Spread

Mean reversion in prediction markets is real, but most of it is a bid-ask bounce rather than a forecast. A price that snaps back after a thin-book blip is behaving normally. A price that snaps back after genuine news is usually a market that has not finished repricing. The difference decides whether you are earning the spread or buying a lottery ticket against yourself.

The practical version of the strategy is market making: rest a bid below your estimate of fair value, get filled by an impatient seller, and sell into the spread. It is not "buy anything that fell." In a binary market, "fell" is often just the last trade moving from the ask to the bid.

Three things people call the mean

The last-trade mean. If a market has a spread, transaction prices alternate between the bid and the ask. Roll showed in 1984 that this alone induces negative autocorrelation in observed returns, even when the underlying value never moves. Compute reversion from trade prints and you will find it whether or not it exists.

Your model's fair value. This is a forecast. It can be right, wrong, or right but early. Reversion to your own estimate is only a strategy if the estimate has an independent basis, such as a base rate, a poll average, or a mechanical relationship to another market.

Structural convergence. As resolution approaches, a price must converge toward $0 or $1. That is not mean reversion; the mean is moving. A contract at 0.70 that drifts to 0.95 over three months has not "reverted," it has been repriced by accumulating information.

Bid-ask bounce, in cents

Suppose the best bid is 0.48 and the best ask is 0.52. The midpoint is 0.50. A buyer lifting the ask prints 0.52; the next seller hitting the bid prints 0.48. That is an 8% decline in the printed price with no change in the underlying forecast.

Now add costs. At 0.50, a politics taker fee is 0.04 × 0.50 × 0.50, or one cent per share per side. A round trip therefore costs two cents in fees plus the four-cent spread, about six cents on a 50-cent contract, or 12% of the stake. Fading a three-cent move in that market is not a trade; it is a donation.

There is a display detail that makes this worse. Polymarket shows the midpoint, but if the spread is wider than $0.10 it shows the last traded price instead. In an illiquid market, the quoted price can therefore jump when the book widens, with no trade behind the change. If your reversion signal reads the displayed number, it is partly reading quote width.

Overreaction is a hypothesis, not a default

The behavioral case for fading is intuitive: thin books, salient headlines, a crowd that overweights the latest story. But there is no general rule that prediction markets overreact. Traders often anchor on the previous price and update slowly, which produces underreaction rather than overshoot. The academic literature on prediction-market accuracy points to good aggregation of public information, not to a reliable overreaction premium.

The tradable version of overreaction is narrower. It appears when a headline is ambiguous relative to the resolution rule: an unconfirmed report, a paraphrase of an official statement, a story about a related event that does not meet the market's stated criterion. In those cases the price can move on something that will not decide the contract. That is a genuine gap between headline and resolution, and it is different from "the market overreacted to real news."

Structural limits near 0 and 100

Near the boundaries, the payoff becomes brutally asymmetric and the usual reversion intuition breaks.

At 0.98, buying Yes risks 98 cents to make two. A one-cent spread consumes half the remaining upside. The fee is small in absolute terms, 0.04 × 0.98 × 0.02, or 0.078 cents per share, but it is almost 4% of the two-cent edge. More important, the downside is not limited to the price move.

Resolution risk is the reason. Polymarket resolves prediction markets through the UMA optimistic oracle: a proposer posts a bond, typically $750, and there is a two-hour challenge period. If a proposal is disputed twice, the question goes to a UMA token-holder vote, and the process can take four to six days. The documented outcomes include "Unknown/50-50," in which each token redeems for $0.50. A 98-cent position that resolves 50/50 loses 48 cents. That is not a remote theoretical tail; it is a published part of the process.

The mirror image is selling a longshot. Buying No at 0.98 collects two cents per share and risks 98. One resolution surprise erases dozens of good trades. This is the "picking up pennies in front of a steamroller" problem, and in binary markets the steamroller has a name: an ambiguous rule.

Negative-risk markets add a useful but double-edged tool. In a multi-outcome event where only one outcome can win, one No share can be converted into one Yes share in every other outcome. That makes hedging cheaper, but it also means the price of one outcome is linked to all the others. A "cheap" 2-cent longshot may be cheap because the No side is expensive, not because the market is wrong.

Event risk ends the trade

Mean reversion assumes no new information. A scheduled catalyst invalidates that assumption completely. A court ruling, a data release, or a debate can turn a 0.48-to-0.52 oscillation into a 0.48-to-0.20 gap in one print.

The rule is blunt: do not hold a fade through a scheduled event. Close it, hedge it, or accept that you are no longer trading mean reversion. The same applies as a market approaches its end date: liquidity thins, the price becomes a step function, and the "mean" you were fading is now the resolution value.

The honest version: get paid for liquidity

Consider a market with a 0.30 bid and a 0.34 ask. Your estimate of fair value is 0.33. Buying at the ask is already a loss of one cent against fair value, plus 0.86 cents of fee. Selling at the bid is worse. There is no reversion trade at the touch.

The trade exists only if you rest a bid at 0.31 and an impatient seller hits it. Your mark against your 0.33 estimate is two cents. If you later sell at 0.33, you gross two cents, pay 0.86 cents on entry and 0.88 cents on exit, and net about 0.26 cents per share, or 0.8% on a 31-cent stake.

Now the risk: if the price falls to 0.25 before you exit, you lose six cents. One bad fill erases more than twenty good ones. That is the real shape of the strategy. You are not predicting reversion; you are being paid to stand in front of order flow, and you keep the spread only if the flow is uninformed.

A rule set

1. Fade order flow, not news. If a headline touches the resolution criterion, assume the move is information.
2. Require the expected reversion to exceed twice the round-trip cost of spread plus fees.
3. Never fade a market whose resolution source or rule is ambiguous. The 50/50 outcome is a real cost.
4. Cap size near the extremes. At 0.02 or 0.98, a resolution dispute can cost you most of the position.
5. Time-box the trade. If the price has not reverted within your stated window, exit and redeploy.
6. Keep a calendar of catalysts and end dates. No fades through them.

What the data cannot tell you

Trade-print data overstates reversion because of the bounce. Midpoint data is better but can be stale or one-sided. Without order-book depth, you cannot tell whether a move was a single impatient order or a genuine repricing. Most retail post-mortems of "mean reversion strategies" fail at exactly this step: they measure printed prices, not executable ones.

The strategy is real, but it is a liquidity business with a thin margin, not a forecast with a wide one.