The free money on Polymarket lasts 200 milliseconds

The dollar rule is as old as the markets. The Iowa Electronic Markets, a research exchange at the University of Iowa, issued contracts only in unit portfolios (Berg et al., 2008). A unit portfolio is one of each contract in a market, bought from the exchange or sold back to it for their total payoff. So a full set priced above its payoff could be bought whole and sold in pieces, and one priced below bought in pieces and sold back. Each trader could typically put in no more than $500.

The plumbing did not stop the mistakes. In the 1992 presidential vote-share market, Kenneth Oliven and Thomas Rietz found that 5.4% of the new best bids and offers broke the no-arbitrage rule (Oliven and Rietz, 2004). Price takers did worse, with 37.7% of their trades made at a worse price than the market offered. Prices stayed sound because, in the authors’ words, “marginal (in this case, market making), not average, traders set prices”.

By the 2012 presidential race on Intrade, the fixing had become a trade of its own. David Rothschild and Rajiv Sethi found 40 accounts in that market running arbitrage strategies, under 1% of its traders (Rothschild and Sethi, 2016). Those accounts made 16% of the market’s volume.

Where the rules kept arbitrageurs small, the gaps stayed open. PredictIt capped each trader at $850 a contract (Stershic and Gujral, 2020). It also kept a tenth of every profit and charged a fee on withdrawals. With those fees and a minimum price of a cent, a set’s bids had to sum to $1.11 before selling one of each paid. In its 2016 nomination and presidential markets, days with mispricings of more than 20 cents added up to as many as 433, while the same markets on the Iowa exchange had none (Stershic and Gujral, 2020). Eric Zitzewitz found that PredictIt’s bids often summed to more than 110 cents for long stretches, which he reads as an arbitrage even after fees, and traced the bias to the cap (Zitzewitz, 2026).

Polymarket started without either brake. Its documentation limits an order by the trader’s balance rather than by a cap like PredictIt’s (Polymarket, 2026c), and it charged no fee per trade before January 2026, when it began charging takers on its 15-minute crypto markets. On March 30 the fee reached every category but geopolitics (Polymarket, 2026d). The studies of its arbitrage count it three ways (Table 2). Saguillo and coauthors read every executed fill on Polygon and counted the trades that locked a payoff, which treats every recorded profit as one that was there to take. Gebele and coauthors kept only the profit they could tie to a mechanism that executes it, so their total is far smaller. Cheng and coauthors, and this report, read the order books instead, because a book shows the prices a taker could have hit and how long they stayed.

Table 2. Arbitrage in prediction markets as the studies measured it, from the Iowa Electronic Markets in 1992 to Polymarket in 2026; the last row is this report’s.

The scale runs from months to milliseconds. On PredictIt a mispricing could sit for a season, because the cap and the fees kept anyone from betting enough to remove it; on Polymarket the S&P 500 and SPY books closed half their gaps within 216 milliseconds.

Every rung of a ladder settles on one number. “Bitcoin above $77,000 at 2 a.m. ET on August 22” and “Bitcoin above $76,800” both read the close of the same one-hour Binance candle (Polymarket, 2026k), so the higher rung can only pay if the lower one does. Buying YES on $76,800 with NO on $77,000 therefore pays at least $1. The same holds within every close ladder on stocks, indices and oil, and within every “reach” or “dip to” ladder that counts a touch of a price during a day, a week or a month.

Each rung settles on its own. Polymarket resolves its prediction markets through the UMA Optimistic Oracle: anyone may propose the outcome by posting a bond, typically $750, and the proposal stands if no one disputes it within two hours (Polymarket, 2026e). Nothing in that process checks one rung against its neighbour, so the dollar a pair promises rests on every proposer reading the same number off the same source.

Polymarket kept every one of those promises. The 2,414 ladder events that ended in the window held 294,826 linked pairs, and in none of them did the more likely rung pay nothing while the less likely rung paid $1 (Table 3). Because pairs share rungs, a bound belongs to whole events: with none out of order, at most about one ladder event in 800 settles out of order, at 95% confidence. That bound is pooled over every kind of ladder, most of them hourly crypto ladders; the 317 touch ladders alone support about one in 110. Each of the 217 multi-outcome sets that ended in the window paid exactly one of its outcomes.

Table 3. Linked ladder pairs that settled in the window, by asset class and contract kind, and the pairs in which the more likely rung paid nothing while the less likely rung paid $1.

The one trap is in the reading of the rules. A touch rung counts a price reached “at any point after market creation”, so a rung Polymarket adds in the middle of a week watches a shorter stretch than an older rung at a nearby strike (Polymarket, 2026j). Pairing rungs by strike alone, without the listing times, produces 631 apparent failures out of 302,675 pairs, and every one of them pairs a touch rung with an older one, so the two watched different stretches. The settlement risk in these trades is the risk of misreading the contract, and in these five weeks Polymarket’s own settlements gave them no other reason to fail.

Polymarket’s S&P 500 market settles Up when the official index close beats the previous day’s, and its SPY market settles Up when the fund’s last-minute close on Pyth beats the previous day’s (Polymarket, 2026h; Polymarket, 2026i). Because SPY tracks the index, S&P 500 Up plus SPY Down pays exactly $1 on any day the two agree, and they agreed on all 23 trading days in the sample. When that pair costs less than $1 after both fees, the books have crossed.

The books crossed 5,492 times in regular trading, a median of 149 times a day. They came in bursts, with 1,945 on August 28 alone. Half closed within 216 milliseconds. A quarter were gone within 81 milliseconds, while one in ten lasted longer than 3.3 seconds.

An order takes at least one round trip to land, because the change in the book has to reach the trader and the order has to travel back. From a home broadband connection that round trip took a median 192 milliseconds on October 5, 2026, and from a rented cloud server 170. By the time an order from home could land, nearly half the crossings had closed (Fig. 2). By 350 milliseconds more than three in five had.

The fee takes most of what a crossing offers. The median crossing priced the two legs 1.1 cents under a dollar before fees, and the median fee bill for the pair was 0.9 cents. What was left, the edge, came to a median 0.16 cents a share. More than nine in ten crossings offered less than a cent of it. Summed over every crossing, the fees took 72% of the gap before fees.

The crossings that offered the most were the first to go (Fig. 3). Crossings worth under a quarter of a cent a share, three in five of them, lasted a median 201 milliseconds, and those worth a quarter of a cent to a cent lasted longer still. Those worth 2 to 5 cents lasted a median 88. The few worth 5 cents or more lasted 70.

The replay tries every crossing. At its first instant it plans both legs at the best prices, up to 100 shares and at least $1 a leg, and skips a crossing with less than that or with quotes it has already traded against. Both orders reach the book a set delay later and fill only what is still there at those prices or better. Matched shares pay $1 at the close, and an unmatched leg is sold back into the book after a second delay, or held if too small to sell.

With instant orders, every one of the 2,212 planned trades filled in full. In all, the replay made $480.50 on $133,865 spent. That is 0.36 cents on every dollar (Fig. 4). Pooled over the five weeks, fifty milliseconds of delay removed the profit without yet making a clear loss. Both legs filled on five trades in six, and the replay lost 0.19 cents per dollar, too close to zero to tell from breaking even. From 100 milliseconds on, the loss is clear of zero at 95% confidence. At 0.35 seconds, the delay this report uses for an order from home, both legs filled on only about a third of trades, and one leg came up short on more than half (Table 4). The replay lost 1.98 cents per dollar.

The loss comes from the missing leg. At 0.35 seconds the replay held three unmatched shares for every two it could match, and an unmatched share is a one-sided bet bought at a price that has just moved away. Selling it back costs more than the matched shares earn at under a cent each, and longer delays only deepen the loss, to 2.92 cents per dollar at 2 seconds.

Table 4. The replayed taker at the S&P 500 and SPY crossings, by arrival delay: the share of trades with both legs filled and with one leg short, the remainder filling nothing, and net profit per dollar spent with its 95% interval over trading days.

Polymarket adds a wait of its own on SPY’s side of the pair. On a market it flags, an order that would trade at once is held before it is matched, and its documentation tells traders to look for that flag, itode, in each market’s record (Polymarket, 2026c). Read on October 6, the record of every SPY Up or Down market in the sample carries it, and none of the S&P 500 markets does (Polymarket, 2026g). The flag on markets that closed in August and September is the evidence that the hold applied in the window. Its length there is an inference: the documentation gives 250 milliseconds. The changelog has since cut the crypto hold to 50 milliseconds on August 17 and raised it to 150 on September 4, and it records no change for finance markets (Polymarket, 2026d).

The hold also falls on the leg that went missing more often in the replay. A quarter of a second after a crossing opened, the SPY leg of a planned trade was still there in full 61% of the time. The S&P 500 leg, which carries no flag, was there 72% of the time. However fast its sender, an order for the SPY leg waits out the hold, so on these two books the instant taker in Table 4 is a ceiling. The replay delays both legs equally; under the hold the S&P 500 leg would fill first and the SPY leg a hold later, a case between the rows of Table 4.

Bigger contradictions, where a buyer bids more for a harder rung than a seller asks for an easier one, come in the hour after a ladder lists. On September 18 at 12:01 UTC Polymarket listed SPY’s ladder for September 21, with 11 rungs. Forty-six minutes later a buyer bid 56 cents for “SPY closes above $755” (Fig. 5). At that moment the four easier rungs still offered YES at 48 to 50 cents. Buying YES on $750 at 48 cents and NO on $755 at 44 cents locked a profit. After fees it came to 6.02 cents a share. The four pairs against the $755 bid stayed open for 202 milliseconds to 4.8 seconds.

Mistakes like SPY’s are rare: in the hour after each listing we watched 37,497 linked pairs, and their prices contradicted each other after fees only 43 times (Table 5). Ninety-nine percent of the pairs with quotes on both rungs never came within 1.19 cents of a lock. Bitcoin’s and Ether’s hourly ladders list 20 and 16 rungs every hour, so the watch plan takes three whole days of them to stand for the rest. Pendulum Flow’s record covers 1,104 of the plan’s 1,297 batches of new rungs.

Table 5. From the pairs watched in the hour after each listing to the contradictions that could be traded.

Most of the 43 were too small to take. Twenty-nine were penny quotes on touch rungs added to a live ladder. At the minimum order of five shares, each locked under 5 cents. Eleven came from Ether’s weekly touch ladder for September 7 to 13. There the bid on the dearer rung held only 5 shares, so at most 75 cents could go into the cheaper leg.

Three contradictions held at least $1 a leg, and two of them were gone before a home connection could react (Table 6). All three came 35 to 55 minutes after their ladders listed. Bitcoin’s 2 a.m. hourly ladder for August 22 offered 2.42 cents a share for 53 milliseconds. The WTI ladder for September 21 offered 10.57 cents, which lasted 183 milliseconds. Trading up to $30 at a time, an instant taker would have made $3.43 from all three, since every quote was still there. One arriving 0.3 seconds late would have lost $2.98, because by then the Bitcoin and WTI quotes were gone and only SPY’s remained.

Table 6. The three listing contradictions with at least $1 a leg at the contradictory prices: gap after fees, how long it lasted, and the replayed result at up to $30 a trade, instantly and 0.3 seconds late.

The ladders carry Polymarket’s hold as well. Every market behind the three tradable contradictions is flagged, and so are 70 of the 90 ladder rungs we sampled across asset classes. On them a taker waits out the hold on both legs, on top of the round trip. Bitcoin’s gap, on a crypto ladder, lasted 53 milliseconds, barely longer than the 50 milliseconds the changelog then gave for crypto markets (Polymarket, 2026d).

The other kind of lock lives in sets where exactly one outcome pays: brackets on where Bitcoin or Nvidia will close on a date, rankings and brackets of companies’ market value, and central-bank decisions. Buying every outcome’s YES for under $1 locks a profit, and so does selling every one for over $1 by converting a single NO. These sets charged takers a rate of 0.04 to 0.07.

Gaps in these sets last far longer than on the S&P 500 and SPY books, even with all three kinds measured by a common rule. Each gap is a stretch that locks at least half a cent after fees, with closures under two seconds merged. Of the 25 such stretches in the sampled sets, 92% were still open a second after they appeared (Fig. 6). Twenty of the 25, though, came from one Bank of Israel set. After new listings, 82% were still open after a second, against 44% on the S&P 500 and SPY books. None of the 33 markets we checked in five of the sets carries the hold.

Locked baskets were rare all the same. We drew three hours a day at random and added the three around the Fed’s decision on September 16, 108 hours in all. Across the 217 sets they make 4,888 set-hours. A locked basket appeared in 4 of them (Table 7). In 99% of the set-hours with a full basket on offer, the cheapest one stayed at least 0.55 cents short of a lock.

Table 7. The four set-hours in the sample with a basket that locked at least half a cent after fees: how long it lasted, its best gap, and the most any of its stretches could lock at once and 0.35 seconds later.

Meta’s market value set, seven brackets on where Meta would end August, shows a slow lock and a deep one within half an hour; it traded $17,725 in all. At 14:21 UTC on August 18 its seven bids summed to 118.8 cents. Most of it was 85 cents for the $1.25 to $1.50 trillion bracket and 32 for the next. Selling one YES of each locked about 17 cents after fees, and the gap stayed open 23.7 seconds. The bids behind it held a single basket, so that was all it was worth.

Half an hour later the gap was wider still and twenty times as deep. At 14:51:28 a buyer bid 35.4 cents for the $1.00 to $1.25 trillion bracket while the favourite was bid at 94. The seven bids now summed to 131.3 cents. A trader could buy NO on the favourite for 6 cents and convert it into YES on the other six brackets, which buyers were bidding 37.3 cents for. After fees the lock came to about 30 cents a basket. The books held 20 baskets, $6.02 in all, and the gap lasted 33 milliseconds. A third stretch soon after lasted 10.6 seconds and was worth 17 cents in all. The stretches that lasted held one basket or pennies, and the one with depth behind it was gone before any order from home could land.

The Bank of England basket in Table 1 was the only other lock worth a dollar. It opened at 19:36:14 UTC on September 16, the day of the Fed’s decision, when no change was offered at 86.5 cents (Table 1). The four other outcomes cost 6.3 cents together. The basket was worth $1.03 and was gone within 26 milliseconds (Table 7). The slowest basket shows what lasts. In the Bank of Israel’s August decision set, the offers summed to 98.3 cents, a lock of 0.95 cents after fees. Three of the five were offers of a tenth of a cent. The gap came and went 20 times over 44 minutes. Where its depth was recorded, the books behind it held about five baskets, 5 cents in all.

The largest set offered nothing. The Fed’s September decision traded $208 million. A full basket was on offer in 84 of its 93 sampled hours. Neither of its baskets came closer than 1.12 cents to a lock, and in the median hour the better one was 2.41 cents short (Fig. 7). In its last two sampled hours, around the decision on September 16, the five outcomes were never all on offer at once, so no full basket could be bought. Of the ten central-bank sets in the window, nine had a full basket on offer at some point. Across those nine, a locked basket appeared in 2 of 463 quoted hours: the Bank of England’s dollar and the Bank of Israel’s pennies.

Arriving a tenth of a second late lost money in every week of the sample (Fig. 8). On time, the replay made between 0.20 and 0.45 cents per dollar each week. Fifty milliseconds late, every week but one still made a little. The exception was the week that held August 28’s burst of crossings, which lost 0.55 cents per dollar. At 0.35 seconds the smallest weekly loss was 0.45 cents per dollar.

Taking only the fat gaps makes lateness worse, because they close first (Fig. 3). Crossings worth a cent or more paid an instant taker 1.71 cents per dollar. At 0.35 seconds they lost 4.79, more than twice the loss on all crossings. For crossings worth 2 cents a share or more, the late loss reached 11.14 cents per dollar.

The fee rebate changes the instant result and little else. The instant replay’s own trades come to about $47,000 of weighted volume over the five weeks, enough for Polymarket’s Silver tier. Its 8% rebate lifts the instant result to 0.45 cents per dollar. The top tier, which takes $10 million of weighted volume in 30 days, returns half the fee (Polymarket, 2026f). At that rate the instant result would have been 0.90 cents. At 0.35 seconds the buying fees came to about 0.98 cents per dollar. Even with half of those buying fees back, the late taker loses about 1.48 cents.

The census holds in every corner of the ladder market. All 17 of its cells, from hourly crypto close ladders to monthly touch ladders on single stocks, settled without an inconsistent pair. The smallest, daily close ladders on single stocks, still held 240 pairs. Locked baskets were rare even where the books were full. A complete basket could be bought or sold at some moment in 76% of the set-hours, and only 4 of them held a lock.

The strongest case for the studies is that their profits come from the record of executed trades, so someone took them. Saguillo et al. (2025) found politics dominant, led by the 2024 US election, and most of the money in sets of mutually exclusive outcomes, in a year when Polymarket charged no fee per trade. Held to a mechanism that executes it, the same record yields one 136th as much (Table 2). Jonas Gebele, Timm Mutzel and Florian Matthes put that kind of profit at about $1.12 million across their full sample, nearly all of it through Polymarket’s converter for multi-outcome sets (Gebele et al., 2026). Their median profit per conversion fell from about $1 before August 2024 to about 8 cents by early 2026 (Gebele et al., 2026).

The gaps in those sets also last far longer. Gebele et al. (2026) measured a median life of 16.15 seconds for 253 YES-side gaps in multi-outcome sets. Guang Cheng, Jiaxin Yang and Haoxuan Zou measured 16 seconds for combinations across NBA game markets and 3.6 seconds for single NBA markets (Cheng et al., 2026; Fig. 9).

The studies are right that slower gaps exist, but those gaps are small. Cheng et al. (2026) found 76.9% of their combination episodes too thin to absorb $100. Their single-market medians sit at the 3.6 to 5.5 seconds between their snapshots of the books, so those lives are set partly by how often they looked. Our sample of sets shows the same split. Its basket runs lasted a median 123 seconds, mostly the Bank of Israel’s pennies, and the baskets that lasted seconds or minutes held 5 to 17 cents. The two worth a dollar or more closed in 26 and 33 milliseconds. The ladders and the S&P 500 and SPY pair are two-leg contracts with no converter, and their gaps worth a dollar mostly close within tenths of a second, so on them any such profit needs a reaction measured in milliseconds.

The other case is speed: a trader fast enough still earns the money, more of it with the top rebate tier. The instant replay made a profit on every one of the 22 days it traded (Fig. 10). In all it made $480.50, a median of $9.13 a day. Three busy days, August 26 and 28 and September 1, brought in 65% of it.

Those busy days are also where a small delay hurt most. Fifty milliseconds late, the two best days became the two worst (Fig. 10). August 28 alone lost $239. The late taker still made money on 15 of the 22 days. Its two worst days lost more than the $250 it lost over the whole window. SPY’s hold also rules out the instant taker of the left panel on that book.

Three readings of the pair would show the gaps slowing or the fee giving way (Fig. 11). Each pools the five trading days to its date, because a single day can mislead. On September 17 only 6 crossings had a dollar a leg, and their median life of 1.76 seconds says little about the month.

The median life of the crossings a dollar-a-leg trade could take ran between 182 and 235 milliseconds through August and September. A month above half a second would leave most of them open for an order sent from home, and replaying Pendulum Flow’s later hours tests it directly (Pendulum Flow, 2026). A taker arriving 0.2 seconds late, about one home round trip, lost money in every five-day window. A month in which that taker breaks even would mean the gaps lasted long enough for an order from home, before SPY’s hold.

Between 87% and 99% of each window’s crossings cleared both fees by less than a cent. Polymarket’s finance rate of 0.04 sets that line (Polymarket, 2026a). A window in which fewer than half did would mean the gaps had grown well past the fee, as a cut posted on its fee page would make them.

Polymarket’s hold is the fourth thing to watch. Its changelog moved the crypto hold twice in three weeks, on August 17 and September 4 (Polymarket, 2026d). A shorter hold on finance markets, or one added to the S&P 500 market, would change who can take the pair (Polymarket, 2026g).

Two of the three tradable listing mistakes lasted under 200 milliseconds. A listing contradiction that holds $30 a leg for longer than a second would be the first in this sample, and so would a multi-outcome basket worth a dollar that stays open as long.

Gebele et al. (2026) prototyped a two-way converter that would let a complete set of YES be turned into collateral before settlement, as a set of NO can be today. If Polymarket adopted one, more multi-outcome baskets would become executable, and their median of 8 cents per conversion in early 2026 is the number that adoption would have to raise. Polymarket’s changelog is where it would be announced (Polymarket, 2026d).