Picture the last three seconds of a five-minute market. A contract on Polymarket has asked a simple question - will Bitcoin be above or below this line when the clock hits zero - and thousands of dollars are sitting on either side of it. For four minutes the price has barely moved. Then, in the final breaths, Bitcoin twitches: a sudden run of buying on Binance lifts the spot price just enough to flip the outcome, the contract settles, and a moment later the price drifts back to where it was. Nobody rang a bell. To the retail trader who was on the wrong side, it simply looked like bad luck.
A new working paper says it was not luck. Three researchers - David Dai, Ruizhe Jia and Shihao Yu, from Stanford University and Singapore Management University - went through roughly two months of Polymarket's five-minute Bitcoin contracts and found the same twitch happening again and again, always in the last seconds, always reverting once the money had changed hands. They gave the paper a plain title: "Settlement Manipulation in Prediction Markets."
The scale they describe is not small. About 821 wallets carry the fingerprints of the tactic, and together they captured an estimated $8.2 million in profit over the window studied. Of that, roughly $1.28 million came straight out of the pockets of ordinary retail traders in the sample - the people on the other side of the last-second move, who thought they were reading a fair market.
Why five minutes is the weak point
The flaw is not Bitcoin, and it is not prediction markets in general. It is the length of the window. Polymarket's shortest Bitcoin contracts settle on a Chainlink price feed that reads the spot price at the exact instant the timer ends. Over five minutes, moving that price for a few seconds is cheap: a trader with enough size can push Binance's order book, hold it just long enough for the settlement to photograph it, then let it snap back. Over a longer window the same push costs far more and lasts far less, because the market has time to correct before the shutter closes. The researchers found almost none of the same behaviour in Polymarket's 15-minute Bitcoin contracts. Their fix is exactly what that implies: longer settlement windows, or a price that averages the final stretch of trading (a time-weighted average) rather than freezing a single instant.
What the number is for
This is the part worth being clear about, because it is the whole reason this desk keeps repeating it. A price on a prediction market is a probability - the crowd's best running guess at whether something will happen. Read that way, it is genuinely useful. But a probability is only as honest as the way it is measured, and a five-minute Bitcoin contract is measured at the one moment that is easiest to bend. The shorter the window, the less the number is a forecast and the more it is a race - and in a race the fastest, best-funded player, not the most correct one, tends to win.
None of this makes the study a reason to fear the tool. It makes it a reason to understand it. A prediction market is at its best over horizons long enough that no single trader can move the answer - an election months away, a tournament that runs for weeks, a policy decision that lands on a known date. It is at its most fragile in the last three seconds of a five-minute chart. Polymarket has not publicly responded to the paper. But the lesson for anyone reading a price does not need its permission: the number tells you what the crowd believes, not what to chase, and the honest ones are the ones nobody can nudge at the buzzer.
