Short answer: Dai, Jia and Yu (2026) found Bitcoin spot orders spiking in the last seconds of Polymarket's 5-minute windows, then reversing. A push against a side priced at 90¢ or more overturned it 34.2% of the time, against 1.0% with no push, and 821 wallets gained $8.2 million in those windows, 93% of it from retail traders. A push pays because a fixed window settles every position on one print. On a rolling horizon, expiries are spread across time, so a 10-second push reaches only the fills from a 10-second stretch: with the same position and the same push, it's worth about 30× less. A fixed 2× also caps the return at +100¢ per $1. A rolling horizon doesn't stop a holder pushing at their own expiry, and the study's point about length still applies: the shorter the horizon, the less a push has to overcome.
A market that settles on a price you can trade has a weak point: anyone holding the contract can trade that price at the moment it counts. This article sets out what the research found on Polymarket's 5-minute Bitcoin market, works through why the attack paid, and shows which parts a rolling horizon contract removes and which it doesn't.
What the study found
Settlement Manipulation in Prediction Markets, by David Dai, Ruizhe Jia and Shihao Yu, covers Polymarket's 5-minute Bitcoin contract from its launch on Feb 12 to Apr 8, 2026. At the time, one Chainlink price settled each window: the contract paid $1 if the close was above the open. Chainlink's price tracks Binance to within a few basis points, so a large order on Binance in the last seconds could carry the settlement across the strike. The price then drifted back.
In this article, a "push" means that order. It has nothing to do with the push that refunds a tie in a horizon contract.
The authors took the top tenth of windows by last-10-second order flow, about 1,600 windows, and compared them with the rest:
| Per window | Windows with a push | Other windows |
|---|---|---|
| Net Binance order flow in the last 10 s ($M) | 1.7 | 0.068 |
| PnL of the wallets identified as manipulators ($) | +5,096 | +6 |
| PnL of retail traders ($) | −4,715 | −221 |
| PnL of market makers ($) | −381 | +215 |
| Windows where retail lost overall (%) | 65 | 48 |
And a push changed who won. This is how often the cheaper side won, by what the dearer side cost 10 seconds before the close:
| Price of the dearer side (¢) | Cheaper side won, push against the dearer side (%) | Cheaper side won, no push (%) |
|---|---|---|
| 50–60 | 65.4 | 41.4 |
| 60–70 | 68.3 | 29.1 |
| 70–80 | 54.3 | 17.7 |
| 80–90 | 54.4 | 8.7 |
| 90–100 | 34.2 | 1.0 |
The findings have limits. The authors identify pushes from order flow and manipulators from realised profits, so they don't observe intent, and they don't link any wallet to a spot order. As CoinDesk reported, they didn't prove that the same people held the contracts and placed the orders. They do rule out hedging: the flow arrives in a burst at the very end, and it still appears when the contract is nearly decided and there's almost nothing to hedge.
Their remedy is a longer horizon. In their model, a longer window gives the price more time to wander from the strike, so a push of a given size can change the result less often. The 15-minute contract, settled by the same oracle and heavily traded, supports it: the order-flow jump at its close was 0.068 log points against 0.406 for the 5-minute contract, with no reversal. They also list averaging or randomising the close, capping positions and taxing spot trades at settlement, and they call it "a settlement-design problem".
They didn't study rolling contracts. Both contracts in their data are fixed windows, and no real-money order book lists rolling ones, so there's nothing to measure yet. What follows is our arithmetic on their mechanism, not their finding.
Why a push pays
The gain from a push comes down to one line:
gain = position that settles on the push × how far the push moves its win chance − cost of the push
The cost is set by the spot market. The first two terms are set by the contract, and a fixed window priced in cents is generous on both. We tested them on 30 days of Bitcoin prices at 1-second resolution (Sep 27 to Oct 26, 2026, Binance BTCUSDT, 8,638 five-minute windows), with a push of 5 basis points, about $42 at that month's price of $83,800.
1. How much of a position one push reaches
This is the largest difference. In a fixed window, every position settles on the same print, however early it was built. A rolling contract has its own expiry, a fixed time after its own fill. Build the same position in both, evenly over 5 minutes, and push for the last 10 seconds of a fixed window or for any 10 seconds of a rolling book:
| Fixed 5-minute window | Rolling 5-minute contracts | |
|---|---|---|
| Position built over 5 minutes | $3,000 | $3,000 |
| Share of it settling during a 10 s push | All of it | 1/30 of it |
| Position settling during the push | $3,000 | $100 |
| Win chance without the push (%) | 49.4 | 49.4 |
| Win chance with a 5 bp push (%) | 76.8 | 77.0 |
| Expected gain from the push | $1,644 | $55 |
The push moves the win chance by the same amount, but it reaches 30 times less of the position. To reach the rest, you'd have to hold the price up for 5 minutes rather than 10 seconds, and the study's pushes are bursts that start reversing within 10 seconds.
A trader can still aim for a slice by buying a lot in one burst. Only fills in the same second share an expiry on Rizon, so a position that settles within a 10-second push has to be filled within a 10-second stretch, one horizon earlier. That means taking whatever the book offers in those seconds, at a strike that gets worse with each level, rather than building patiently over the window. The study's wallets earned about $5,096 a window. At +54¢ per $1, that takes about $9,400 of size filled inside those few seconds.
2. What a win pays
In a market priced in cents, the side behind gets cheap as the close nears. Over our sample, 10 seconds before the close, it was worth:
| Side behind by (bp) | Fair price with no push (¢) |
|---|---|
| 1 | 19.8 |
| 2 | 14.1 |
| 3 | 7.9 |
| 5 | 1.5 |
Holding a contract at those prices when a push carries it pays many times over. Buying the side behind by up to 5 basis points, at those prices, and pushing 5 basis points returned +567¢ per $1 in our sample, with the win chance going from 14.3% to 95.3%. The study's most striking number sits at this end of the table: sides priced under 10¢ won a third of the time when pushed.
A horizon contract has no such price. Every contract costs half its payout, so a win returns +100¢ per $1 and no more, on any horizon and however far the number is from the strike.
Treat the +567¢ as a ceiling. It assumes you can buy at the fair price, and the study finds market makers pulling quotes near the close. We also can't tell what those wallets paid. And the 15-minute window offers the same cheap prices in our data, yet the study finds little there, so this can't be the whole story.
3. Buying after you've looked
A fixed window stays on sale until the close, at a strike set minutes ago. So a trader can watch the price, see that it's near the strike with seconds left, and only then buy. A rolling market has nothing like that on sale. Every fill starts a new contract with a full horizon ahead and a strike near spot, so whoever plans to push has to commit a full horizon before they can know whether the price will be anywhere near their strike.
What a rolling horizon doesn't change
A push still moves the win chance. A holder can still trade spot at their own expiry. A rolling 5-minute contract has the same spread of outcomes as a fixed 5-minute window bought at the open: +54¢ per $1 for a 5 bp push on both. What the rolling market removes is the reach, the cheap price and the look first, not the push.
Length matters as much as ever. The study's central result carries over: the shorter the horizon, the less the price wanders before expiry, and the more a push decides.
| Rolling horizon | Win chance without the push (%) | Win chance with a 5 bp push (%) | Expected PnL per $1 of size with the push (¢) |
|---|---|---|---|
| 5 minutes | 49.4 | 77.0 | +53.9 |
| 1 minute | 47.3 | 91.6 | +83.1 |
| 15 seconds | 41.4 | 98.2 | +96.4 |
(The win chances without a push fall short of 50% because of ties, which refund both sides.) Rizon's 15-second and 1-minute books are shorter than the contract in the study. A 5 bp push nearly decides a 15-second contract, so on those horizons the protection is the cap and the reach, not the difficulty of the push.
The settlement rule still matters. Rizon's rolling markets settle on the last live price at or before expiry: one print, from Coinbase, with Kraken as backup. That's fine for paper money, where there's nothing to gain. A real-money rolling market should settle on an average across venues.
It's untested. This is reasoning and arithmetic. Nobody has measured manipulation on a real-money rolling order book, and paper money gives nobody a reason to try.
What each fix does
These aren't rival fixes. Each removes a different part of the attack, and they'd be stronger together.
| Fix | What it removes | Who uses it |
|---|---|---|
| A longer horizon | The share of windows where a push can change the result. Within 5 bp of the strike 10 s before the close: 54.3% of 5-minute windows, 37.8% of 15-minute | The study's main remedy |
| Averaging the close | Most of a short push's effect. A 5 bp push over the last 10 s moves a 60 s average by 0.83 bp, and only 11.8% of windows sit that close | Polymarket since August 2026, Kalshi since January 2026 |
| A rolling expiry | The reach of one push: fills from seconds, not a whole window | Rizon |
| A fixed 2× price | The cheap side: a win returns +100¢ per $1 at most | Rizon |
| Position caps, randomised closes, a tax on settlement-window spot trades | The size of the prize, or the certainty of its timing | Proposed in the study |
FAQ
What is settlement manipulation?
Holding a contract that settles on a price, then trading the underlying at settlement to move that price in your favour. The price usually drifts back afterwards, because the order carried no information.
Were Polymarket's 5-minute markets manipulated?
A 2026 study found the pattern you'd expect from manipulation: large Binance orders in the last 10 seconds, price reversals afterwards, and $8.2 million of gains to 821 wallets in those windows. It doesn't prove intent, or that the same people held the contracts and placed the orders.
Did Polymarket fix it?
It changed the rule. In August 2026 Polymarket moved its crypto up/down markets from a single price to a time-weighted average, which makes a short push much weaker. We haven't seen a published measurement of how much that reduced it.
Are rolling horizon contracts immune to manipulation?
No. A holder can still trade spot at their own expiry, and a short horizon makes that easier. A rolling market limits how much one push can reach, and a fixed 2× limits what a win pays: the same push on the same position is worth about 30× less over 5 minutes, but it isn't worth nothing.
Why was the 15-minute contract much less affected?
Because the price has more time to wander, it's less often near the strike at the close, so a push of a given size changes the result less often. That's the study's explanation, and our data shows the same thing: 37.8% of 15-minute windows are within 5 bp of the strike 10 seconds out, against 54.3% of 5-minute windows.
Can Rizon's markets be manipulated?
Rizon is paper money, so there's nothing to gain today. If it listed real money, its short horizons and its single-print settlement would need the fixes above.
Originally published at rizon.markets. Rizon is a paper-money exchange for horizon contracts.
Top comments (0)