Polymarket's TWAP Settlement: Data & Strategy Analysis

TWAP means time-weighted average price. Instead of reading the price once, you read it every second for a while and take the average. That is the entire change.

Every one of these markets asks the same question: is the coin higher or lower than when the market opened? That does not change. The opening reference does not change. The payout does not change. The only thing changing is which number you compare against the open.

In symbols: P(t) is the coin price at second t, the market opens at P(0) and closes at second C (300 on a 5-minute market, 900 on a 15-minute one). The averaging window is the last W seconds, either 30 or 60.

Deciding who wins needs no model at all. It's arithmetic:

That's the whole of it, and nothing in it is estimated. We ran both formulas over every cycle in our archive and compared the winners. That's where 6.22% and 4.35% come from.

The harder question, what the order book would have looked like, does need a model, because nobody ever quoted these markets under the new rule. That's the next section.

Nobody ever quoted these markets under the new rule, so there is no real book to show. Rather than invent a theoretical one, we move the real one:

The bracketed term is the only thing that moves: the shift. Everything the real market got wrong (its spread, its lag, its bias, its depth) carries across untouched, because it is still the same book, just re-centred on a new rule.

The shift itself comes from asking, at each second and under each rule, what is the chance Up wins given only what has happened so far? Under the snapshot rule the best forecast of the settlement price is simply the current price. Under TWAP, once you are inside the window, part of the average is already locked in. Those seconds have happened and cannot change. That single fact is what drives everything that follows.

About 6% of markets reverse. That average hides the shape completely. Reversals sit almost entirely in markets that finish close to the open. Across the 11,496 markets we sampled from five June days, of the 448 that moved 50 bp or more, not one reversed. The biggest move that still reversed was 31.6 bp.

The tables on this tab and the next come from that five-day sample, so they run slightly hotter than the full month: 6.40% and 4.71% here, against 6.22% and 4.35% across all of June.

The headline figures on the previous tab are unconditional: every cycle counted once, knife-edge markets mixed in with markets that were decided two minutes ago. Once we split that same moment by how far the price had actually travelled, the average falls apart into something you can trade on.

If your position is more than 10 bp ahead with ten seconds to go you were already safe under the old rule and the change does nothing for you. If it is under a basis point ahead, you were facing a 34.62% chance of losing a position you thought you had won, and the new rule cuts that to 9.21%.

Everything above reduces to a single comparison: what the market is charging you versus what actually happens from a position like this. The difference is the edge.

The same position one minute earlier, at 60 seconds left with the same 5 to 10 bp lead, is only worth 93.4%, which makes $0.90 a +3.4 trade barely worth the spread. The edge is created by the window opening, not by the price.

Pick a market and press play. The cycle runs one tick per second: the first chart draws the coin price, the second draws both order books at once. The dashed lines are the book priced for snapshot settlement, the solid lines are the same book priced for TWAP. The strip above recomputes every tick from the measured table, so you can watch the edge open up as the window fills.

Try DOGE · 5m and scrub to the last ten seconds: spot goes above the open while the projected TWAP stays below it, the book is still lagging at 0.87, and the signal fires.

Every panel below is measured data. The faded lines are the real order book as it traded; the solid lines are the same book re-priced for the TWAP rule.

• Spreads do not tighten here. Under TWAP market makers face far less end-of-cycle gap risk, and Polymarket is adding $1M in liquidity rewards to encourage tighter quotes. These books hold spreads at their observed values, so any strategy that pays the spread will look worse here than it should after August. Treat backtested spread costs as a floor, not an estimate.
• No behavioural response is modelled. This is the same order flow re-priced, not a simulation of how traders will act once the rule changes. Everyone gets this arithmetic; the lock-in edge is real but it is not proprietary and it will compress.
• The window edge is an assumption. The lengths (30s and 60s) are confirmed. Chainlink does not publish the feed's sampling boundaries, and whether the market open also becomes an average is unstated. A strategy whose edge lives entirely in the last second or two of the window is resting on an unverified detail. Test its sensitivity to a ±1 second shift before sizing it.
• The closing second is an assumption too. The old rule reads one price at the close, and the feed does not say whether that is the last second inside the cycle or the first tick of the next one. On a quiet close it makes no difference. On a violent one it decides the market. We rejected one candidate example for this article because 78% of its move landed in its final second, so which of the two seconds you read flipped its result. Every example that survived was re-tested against both readings and against shifting the averaging window five seconds either way. If your edge depends on that single tick, it is resting on something nobody has published.
• The probability tables are measured under the old rule's order flow. The price path is rule-independent, so the arithmetic holds, but the thinnest cells (under 1 bp, few seconds left) rest on a few hundred samples. Treat that column as directional.