What one paper documented about last-second settlement on Polymarket’s five-minute Bitcoin markets

Public write-ups about Polymarket’s five-minute Bitcoin markets often skip the paper and keep the folklore: a five-second push on Binance drained millions, the study found fraud, and the same window is sitting there now.

One working paper measured a version of that activity. It is Dai, Jia, and Yu, Settlement Manipulation in Prediction Markets (arXiv:2606.31675; the opened page’s current version is 11 August 2026). This note states what that paper documents, and what public reporting plus official docs say the venue changed on 7 August 2026. It is not a method for repeating the measurement.

The sample is one contract, ending 8 April 2026

The authors study Polymarket’s Bitcoin five-minute up/down contract: a binary claim that pays $1 if Bitcoin is higher at the close of a five-minute window than at its open, and $0 otherwise. They date the public launch of that contract to 12 February 2026. Their tape runs 1 July 2025 through 8 April 2026. The five-minute-live slice is 12 February to 8 April 2026 — 56 days on the page, which the intro also calls “two months.”

They count over 60 million fills on that Bitcoin five-minute contract, 16,073 cycles with Polymarket data, and roughly 243,000 wallets that traded it.

That window is the fact. It is not a statement about books on 15 August 2026.

In the sample, the close was a single print

In the period they measure, the five-minute, fifteen-minute, and four-hour crypto contracts resolve against a Chainlink oracle that aggregates many venues — not against Polymarket’s own book, and not against one named exchange. The strike is the oracle price sampled at the contract’s open. The settlement value is the oracle price at its close. The “up” side wins if the close print exceeds the strike.

They treat Binance as a tight proxy for that print: its mid sits about two and a half basis points from the oracle and finishes on the same side of the strike as the resolution about 85% of the time. Their later design section calls this what it is: a contract that “settles on the spot price at a single instant—the close.”

A general Polymarket resolution page retrieved the same day still describes a different machine — the UMA Optimistic Oracle, with a bond and a challenge window — for ordinary event markets. Those two paths are not the same rule.

They classified cycles. They did not publish a fraud verdict.

After the five-minute launch, they report a footprint in the last ten seconds before each close on Binance: the magnitude of net order flow jumps, about 50% above the pre-launch level. Their Table 2 states that last-ten-second order flow is 50% higher relative to the rest of the cycle than it was before launch (0.406 log points), with absolute return 15% higher and near-mid depth 5% higher. About a tenth of that final-ten-second spot move reverses in the next ten seconds.

The jump is larger where a small push can still change the result. In the roughly 6% of cycles whose contract price still implied a near-even outcome just before the close, the near-settlement order-flow jump is about 3.9 times that in the rest. The post-close reversal is about a quarter in those near-even cycles and a tenth in the others.

They then label cycles. For each cycle they compare the final-ten-second Binance flow with that cycle’s own typical flow. The top decile — about 1,600 cycles — they classify as manipulated. That is a statistical cut, not a court finding.

They identify “manipulators” as the wallets that gain in those classified cycles: “an identification based on realized profits rather than a direct observation of intent.” CoinDesk, reporting the paper on 7 August 2026, writes the same limit: the paper did not prove intent, and did not directly establish that the spot-market orders were placed by the same people holding Polymarket positions.

So the headline number is a profit-based classification on that tape. It is not “the study found fraud.”

The number on the page is $8.2 million. Read the identification.

821 wallets fit their profit rule — about one in three hundred of the roughly 243,000 that traded the contract. Those wallets take $8.2 million in the pushed cycles and break even in the rest.

Setting market makers aside (the authors say they quote passively and end each cycle flat), 93% of the remaining loss in those cycles falls on retail. Those retail accounts are on the losing side in 65% of manipulated cycles, against 48% of normal ones.

Inside the classified cycles, a push against the favorite flipped a near-even book 65% of the time, against 41% in ordinary trading. When the market gave one side a 90-to-100% chance just before the close, a push against it reversed the outcome 34% of the time, against 1% with no push. The authors’ sentence: “A bet the market treated as near-certain was overturned one time in three.”

They reject a pure-hedging story for that last group: a binary that is already almost certain barely still moves with spot, and “essentially all of the signed trading falls in the final fifty seconds.” That is their argument on the page. It is not a live 15 August frequency table.

The opened HTML extract of the paper ends during the later empirical sections. Later tables, and whether the 11 August version evaluates the live TWAP change, are UNKNOWN from this retrieval.

What the venue said it changed — and what the docs actually show

CoinDesk reported on 7 August 2026 that Polymarket had replaced the single-price snapshot on short-dated crypto contracts with a time-weighted average price, delivered through Chainlink Data Streams, and had announced $1 million in liquidity rewards across the affected markets through August. CoinDesk quotes a Polymarket statement to that effect from an X post. This note did not open that post. The wording is CoinDesk-reported.

CoinDesk, and a CryptoBriefing piece dated the same day, report the announced windows: 30 seconds for five-minute markets, 60 seconds for fifteen-minute and four-hour markets. CryptoBriefing dates the cutover to 7 August 2026, 00:00 UTC, and lists seven assets: Bitcoin, Ethereum, Solana, XRP, HYPE, BNB, and Dogecoin.

Official docs retrieved 15 August 2026 confirm that Chainlink-computed 30-second and 60-second TWAP feeds exist, and that they are lookback windows, not publication cadences. The opened TWAP page does not say which market duration uses which window. Chainlink, the same page says, does not currently publish the custom feed’s sampling boundaries, weighting, rounding, or missing-input behavior.

The official liquidity-rewards page states that Polymarket is adding $1 million across impacted markets through August, and that the allocation “applies only to crypto 5-minute, 15-minute, and 4-hour markets that settle on TWAP.” Those figures are configured caps, not measured payouts. The opened table: five-minute $550k (Bitcoin $300k; SOL, ETH, HYPE, and XRP $200k split evenly; BNB and DOGE $50k split evenly); fifteen-minute $350k; four-hour $100k.

The authors had already named both levers on the page: lengthen the horizon (they say the five-minute footprint is sharp and the fifteen-minute footprint is attenuated), and replace a single instant with an average. They did not measure the 7 August change. Their sample ends 8 April 2026.

What this paper does not say

It does not say the same windows are still there on 15 August 2026. UNKNOWN.

It does not say the TWAP change removed the transfer they measured. UNKNOWN.

It does not say which lookback window a named live five-minute Bitcoin market used at retrieval on 15 August 2026. Official docs describe both 30-second and 60-second feeds. The announcement, as reported, assigned 30 seconds to five-minute books. A live market’s own resolution-source line was not retrieved this run. UNKNOWN.

It does not say the 821 wallets committed fraud, or that they placed the Binance orders. Identification is realized profit in classified cycles. UNKNOWN.

It does not say how to take a settlement window, how to trade Binance into a close, or how to fade a five-minute book.

If you came here because a thread offered a five-second fraud story, the useful move is smaller than picking up the story or picking up a recipe: keep the dated measurement, keep the identification rule on the same line as the dollar figure, and refuse the folklore until another primary replaces the 8 April 2026 end date.