Binance's closing-second spike behind Polymarket's Bitcoin bets halves

In June, David Dai, Ruizhe Jia and Shihao Yu published a study of Polymarket’s five-minute Bitcoin contract. Each contract pays on whether Bitcoin ends a five-minute window up or down, and an oracle price at the close decides it. The authors found that trading on Binance jumps in the last ten seconds of each window once the contract went live, that the jump is largest when the bet is close to even, and that a small group of wallets profits in the windows where late trading moved the price.

I wanted to know whether the result can be rebuilt from public data alone, and whether it lasted after the study’s sample ended on 8 April 2026.

The method

I fixed every test and tolerance before I read any data. Binance publishes each day’s one-second trade data with a checksum. Polymarket’s trades come from its public data API, and each wallet’s fills come from the exchange contract’s logs on the Polygon chain, 207 million fills for the eight weeks the paper studies.

The result

Finding	Paper	This run
Binance flow in the last ten seconds, against the rest of the window, after launch	1.50×	1.50×
The same, in windows near even against decided ones	3.92×	3.65×
Windows the paper classes as pushed	1,613	1,613
Pushed windows in Asian overnight hours	55.7%	55.7%
Gain of the wallets the paper calls manipulators, in pushed windows	$8.22M	$8.79M
Share of losses in pushed windows carried by retail wallets	70.5%	72.8%

The closing spike is specific to the close. With the close moved back in ten-second steps, the real close ranks first of thirty. The near-even effect also survives a shuffle of the labels within each day, and it holds at four band widths.

After the sample, from 9 April to 23 October 2026, the spike is still there at 1.24× the rest of the window, about half its size inside the sample. The fall is far larger than its standard error. An earlier quick look at the same months had shown me the direction, so this test was not blind to it. By October the contract also settled on a time-weighted average price, which is harder to move in the last seconds. I did not date the change.

What failed

The wallet split does not reproduce. Market makers match in every cell, with $3.11M in ordinary windows against the paper’s $3.11M, and the whole population is within tolerance at 245,958 wallets against 243,155. My rules flag 904 manipulators against 821, and they gain $1.29M in ordinary windows against the paper’s $0.09M. Counting only fills inside each window gives 800 wallets and $7.97M in pushed windows, but then ordinary windows miss in the other direction. The paper’s figures sit between the two readings of its text, and its code is not public. Both readings agree that these wallets gain in pushed windows and that retail pays.

My first reading of the near-even test classified 13,997 windows against the paper’s 15,223. Counting trades in either outcome token gives 15,252 and 3.94×, close to the paper, but I chose that reading after I saw the miss, so it is not a blind match.

What was not tested

Order-book depth needs a book history that no free source keeps, so the paper’s liquidity claims are not run. No second person has reviewed how I read the claims out of the paper, and a miss can come from my reading as easily as from the study.