“The bigger money” is between related markets, says an arbitrage guide. The study it cites found 0.24% of the $40 million there
A widely read X article teaches Polymarket arbitrage from a 2025 study that counted about $40 million of it. We checked the article against the study, and checked every multi-outcome event on Polymarket for the trade it teaches.
The claim“In a year of Polymarket data, researchers measured roughly $40 million in real, realized arbitrage profit, extracted by people who did nothing but notice that the numbers didn't sum right.” “The bigger money, and the harder game, is between two different markets that are secretly about the same thing.” “This article will give you EVERYTHING you need to start making first money on Polymarket arb.” — X (@bored2boar)
The short version
The article, “They made $40 million on Polymarket arb because prices didn't add up to a dollar”, was posted on X on August 1, 2026 and has been viewed about 248,500 times. It explains two kinds of arbitrage. In the first, the prices of all outcomes in one market add up to less than $1, so buying one of each locks in a profit, or to more than $1, so the No side is cheap. The second spans two markets about the same event, such as a winner market and a winning-margin market. It calls this second kind “the bigger money” and the place “where the real inefficiency lives”. It bases this on a 2025 study from IMDEA Networks and the Oxford Internet Institute, says the author's own bot is “sitting at 6 figs clean PnL”, and ends with an invitation to the author's Telegram group.
We read the study, “Unravelling the Probabilistic Forest: Arbitrage in Prediction Markets” (arXiv 2508.03474, the only version). Most of the article's details match it: the year of data (markets that resolved from April 2024 to April 2025), the 8,659 single markets and 1,578 multi-outcome markets totalling 17,218 conditions, the 92% topic-matching accuracy, the language model that judged whether two markets were linked, the O(2n+m) search cost, and the warning that the two legs of a trade do not fill at the same moment. The $40 million is real as a figure: the study gives $39,587,585.02. But it is an estimate the authors rebuilt from wallets that held matching positions bought for less than $1, not a count of arbitrage trades, and it ignores fees “because Polymarket currently does not charge per trade executed”, which is no longer true of many markets. The “$3.7 billion” is the election's total trading volume, cited from another source, not a peak.
On the article's main point the study says the opposite. Arbitrage between related markets made about $95,157 of the $39.6 million, 0.24%. The rest came from single markets whose prices did not add up to $1: about $10.6 million from two-outcome markets and about $29.0 million from multi-outcome markets, most of it from buying No on every outcome ($17.3 million) or Yes on every outcome ($11.1 million). The study found that the related-market gaps appeared mostly when trading was thin and averaged around $100 at most. The money was also concentrated: one account made $2.0 million, and the top 10 made $8.18 million, about a fifth of the total.
Is the simpler trade still there? On October 28 at 18:30 UTC we checked every open multi-outcome event on Polymarket at its live order book: 9,150 events. In 6,269, every outcome still in play had a seller, so the full set could be bought. Nine full sets cost less than $1; after Polymarket's fees, five still showed a profit, worth $4.66 in total at the sizes on offer. $4.43 of that is one market, the 2030 World Cup final location, where the money would be tied up until the host is settled, possibly not before 2028. The reverse trade, selling every outcome for more than $1, cleared fees in none. Most low totals were not real gaps: the event listed outcomes nobody could buy, such as placeholder candidates. Our own scanners, which watch for these gaps continuously, tell the same story. They flagged 279 ladder markets (“above $90k” priced below “above $100k”) at quoted prices since July, and none survived at prices you could actually trade. Of 1,065 pairs of related markets a language model linked for us, none showed a profit at tradable prices.
The article does not link the author's bot or wallet, so the “6 figs clean PnL” cannot be checked. The same author's September guide to Bitcoin arbitrage did link a wallet: it never made the trade the guide taught, and lost money after the guide was posted. Verdict: the article describes the study's methods accurately but turns its main finding around. Nearly all the money it cites came from single markets whose prices did not add up, not from related markets. The total is a fee-free estimate from 2024–25, and on the day we checked, the trade it teaches was worth a few dollars across all of Polymarket.
Technical appendix
Exhibit A Where the study's $40 million came from
From “Unravelling the Probabilistic Forest: Arbitrage in Prediction Markets” by Oriol Saguillo, Vahid Ghafouri, Lucianna Kiffer and Guillermo Suarez-Tangil (arXiv 2508.03474, v1, August 5, 2025), sections 7.2–7.4. The study estimates profit from wallets that held positions covering every outcome, bought for less than $1 in total. It ignores fees.
| Kind of arbitrage | Profit | Share |
|---|---|---|
| Two-outcome market, Yes + No bought below $1 | $5,899,287 | 14.9% |
| Two-outcome market, Yes + No sold above $1 | $4,682,075 | 11.8% |
| Multi-outcome market, buying No on every outcome | $17,307,114 | 43.6% |
| Multi-outcome market, buying Yes on every outcome | $11,092,286 | 27.9% |
| Multi-outcome market, selling Yes / selling No | $616,453 | 1.6% |
| Between related markets (combinatorial), 4 pairs | $95,157 | 0.24% |
| Sum of the parts | $39,692,372 | 100% |
The study states the total as $39,587,585.02. Its parts add up to about $105,000 more; it does not say why. Shares are of the sum of the parts. The study's top account made $2,009,632 and its top 10 made $8,184,861 together.
Exhibit B The article's statements about the study
| Article says | Study says |
|---|---|
| “roughly $40 million in real, realized arbitrage profit” | A “realized estimate” of $39,587,585.02, rebuilt from wallets' trades, ignoring fees. |
| Combinatorial arbitrage is “the bigger money” | About $95,157, 0.24% of the total. |
| April 2024 to April 2025; 8,659 single and 1,578 multi-outcome markets; 17,218 conditions | Matches (markets that resolved in that year). |
| Election volume “over $3.7 billion of it at the peak” | “over $3.7 billion in total trading volume” during the 2024 elections, cited from another source. |
| Topics matched with text embeddings, about 92% accurate | Matches (100 hand-checked samples). |
| A language model decided whether two markets were linked | Matches, with human review; on election day 13 of 46,360 pairs were confirmed as linked. |
| Brute-force search costs O(2n+m) | Matches. |
| The trades are non-atomic: one leg can fill without the other | Matches. |
Exhibit C Every multi-outcome event on Polymarket, October 28
Scanned 18:22–18:30 UTC on October 28, 2026: every open event Polymarket runs as a set of outcomes where exactly one wins. For each, we took the best price a seller was asking for Yes on every outcome still in play, and the best price a buyer was bidding. A full set pays exactly $1. Fees are each market's own taker fee. “Sets on offer” is the smallest amount for sale at the best price on any one outcome.
| Events | |
|---|---|
| Open multi-outcome events | 9,150 |
| Full set could be bought (every outcome in play had a seller) | 6,269 |
| Full set cost less than $1 | 9 |
| …and still less than $1 after fees | 5 |
| Full set could be sold (every outcome had a buyer) | 4,434 |
| Sold for more than $1, after fees | 0 |
| Event | Cost of a set | Profit per set after fees | Sets on offer | Locked profit |
|---|---|---|---|---|
| World Cup 2030: location of the final | 94.4¢ | 2.6¢ | 171.6 | $4.43 |
| How many Republican House incumbents will not win their primary | 98.3¢ | 1.7¢ | 5.0 | $0.09 |
| How many cities will Waymo operate in by December 31 | 99.1¢ | 0.9¢ | 4.7 | $0.04 |
| Central Bank of Colombia decision in December 2026 | 99.4¢ | 0.6¢ | 0.7 | $0.00 |
| Lowest Great Salt Lake elevation in 2026 | 99.8¢ | 0.2¢ | 48.0 | $0.10 |
| Total | $4.66 |
A single snapshot cannot see gaps that open and close within seconds; the study itself judged prices from recent trades, carried forward for up to 2.5 hours, not from live order books. In an earlier scan the same afternoon, the lowest totals came from events that list placeholder outcomes (“Candidate A” to “Candidate Z”, “Other”) that nobody is selling, so a full set could not be bought at any price.
Exhibit D What our own scanners have found
Our trading system watches for these gaps continuously and re-prices every flag at the live order book before counting it.
- Ladders (“above $100k” priced higher than “above $90k” on the same asset and date, the article's step two): 279 flags at quoted prices across 32 events from July 8 to October 27, 2026, an average gap of 5.1¢. At tradable prices, none survived; the best was 0.8¢ the wrong way.
- Related markets (a language model proposed pairs where one outcome logically forces another, the study's combinatorial arbitrage): 1,065 pairs from July 10 to 19, 2026. 80 looked mispriced at midpoint prices; none showed a profit at tradable prices. Of the 876 that have resolved, the linked outcomes agreed in 849.
- Two-outcome Bitcoin markets (Up + Down for less than $1): 0 hits in 4,195 scans; see our check of the same author's September guide.
Exhibit E Links in the article
The article ends with an invitation to the author's Telegram group, which is also the website on the author's X profile, and promises a step-by-step build of an arbitrage bot in a later part. The profile describes the author as having five-plus years in arbitrage and “6 figs on memes”. No wallet or bot is linked, so the “6 figs clean PnL” claim cannot be checked.