Polymarket Insider Trading: The Charged Cases, and What a Screen Catches

Correction, 8 October 2026 (Tallinn): the tight screen used raw token notional instead of directional acquisition cost for its concentration gate. The corrected screen has 1,316 bets, and a matched gap of −5.40 cents per share (market-bootstrap 95% interval −10.53 to −0.07 cents). Read the follow-up article and downloadable paper for the recomputed comparisons. Earlier sensitivity tables below remain historical and have not all been recomputed.

The platform-wide long-shot baseline and the corrected tight-screen calculation are separate analyses. Do not combine the older sensitivity tables with the corrected tight-screen denominator.

The tight-screen figures in this section are the original September results, superseded where recomputed by the correction above. They are preserved to document the earlier publication; use the follow-up paper for the corrected strict and topic/time comparisons.

US authorities have charged two traders with insider trading on Polymarket, and the long-shot screen used to flag insider wallets holds bets from both of the wallets that match the complaints. Narrowed to what it is meant to catch, a heavy long-shot bet within 48 hours of the outcome from a wallet under two weeks old with almost no other bets, it leaves 1,331 of 113,863 qualifying bets. They won 20.1% of the time at an average price of 22.5¢, and did 5.2¢ a share worse than matched bets from established wallets (95% interval −10.7¢ to +0.1¢). The screen catches some episodes that made the news, inside many more crowds that lost, and it flagged neither charged case.

The bet that an insider screen on Polymarket is built from has a public definition: more than $2,500 wagered, cumulatively and within an hour, on an outcome priced at 35 cents or less. The published screen then keeps the wallets that won more than 75% of theirs while concentrating on a few markets and topics, and those wallets are the ones that get written about. What has not been published beside them is the denominator — what that bet type does for everybody who places it, across the whole platform.

Vultax ran the bet rule on the historical fill dataset published by Boka Qin and Rui Yang as TimeSeventeen/Polymarket-v1. Across 714,691,643 Polymarket fills from 1 January 2025 to 28 April 2026, settled markets only, it catches 113,863 bets placed by 17,858 wallets, $1,238,620,067 staked. 24.37% of them won, against an average entry price of 23.01¢, and the stake-weighted return was −3.67%. As a population, the bet type loses money.

Then apply the wallet filter as it is written, with no minimum number of bets. It returns 2,070 wallets holding 2,325 qualifying bets and $58.67 million of gross profit — and 1,888 of those 2,070 wallets placed exactly one qualifying bet, 2,026 of them one or two. A success rate measured over a single bet is not a success rate. Requiring three bets or more leaves 4,966 wallets, of which 33 won every one; if each bet had simply won at its own market-implied price, 18.5 wallets would have done the same. The observed count is 1.79 times chance, and 32 of the 33 placed only three or four bets.

Where the bet type did pay is politics: 11,049 bets across 5,111 wallets, $140.9M staked, won 23.61% of the time at an average of 19.66¢, for +16.87%. Of the four categories with more than 4,000 flagged bets it is the only one that made money. Sports, which is 61.7% of every flagged bet, returned −1.77%. Long shots on Fed rate decisions returned −96.44% on 604 bets at a 2.15% win rate, and global elections −86.94% on 1,006.

A separate signature reported alongside it is wallet age, and it points the same way. 4,860 wallets whose first qualifying bet came within 24 hours of their first-ever fill placed 27,910 of these bets, staked $381.3 million and lost $19.58 million: −5.14%, the worst of five age bands. None of this shows that nobody trades Polymarket on private information. It shows that this screen, on its own, does not separate the people who do from the crowd, and that a win rate produced by it means nothing without the numbers on this page beside it.

US authorities have twice charged a trader with insider trading on Polymarket: a US service member, over the operation to capture Nicolás Maduro, and a Google employee, over Google's Year in Search list. Both are allegations. In each case one wallet in the fill record matches the quantities the complaint describes, and both wallets are among the 113,863 bets of the wide screen.

The tight version of the screen, set out below, flagged neither. The first wallet was under six days old and bet 33 hours before the outcome, but it had already traded in 3 other markets and held 76% of its dollars in this one, so it fails the almost-no-other-bets rule. The second placed its long-shot bets 3 to 49 days before the results and had traded in 20 or more other markets. The extra conditions were meant to narrow the wide screen to the suspicious few; on the two cases where a prosecutor has set out the facts, they narrow them away.

Both cases are in the prediction-market lawsuit tracker, with the other court cases over event contracts.

Download table figures:CSVJSON with sources

The rule as published: more than $2,500 wagered at 35 cents or less, cumulatively and inside an hour, on one side of one market. Vultax implemented that directly. One bet is one wallet-market-side episode in which a rolling one-hour sum of stake over the wallet's long-shot fills reaches $2,500, where the side actually held was priced at 35 cents or less. Profit and loss is measured over that episode's long-shot fills only, not over whatever else the wallet held in the market.

The historical fill input is the TimeSeventeen/Polymarket-v1 dataset by Boka Qin and Rui Yang (CC BY 4.0): 714,691,643 fills between 1 January 2025 and 28 April 2026, covering both standard binary markets and the multi-outcome negative-risk markets that a binary-only tape leaves out. Market truth comes from the 662,215 markets in that window that resolved to one of their two labelled outcomes; unsettled markets are excluded, because a screen that cannot say who won cannot say who was right. For each fill the side actually held is resolved from the tape's signed exposure field, so a wallet that sold the favourite is counted as holding the long shot in the same way as a wallet that bought it.

The reason to run this platform-wide rather than on a shortlist is that a screen with no denominator cannot be read. Pick any rule, apply it to a population, keep the wallets that won, and you have a list of wallets with excellent records; whether that list means anything depends entirely on how often the rule wins for everybody. 113,863 bets is the denominator. Every win rate quoted against this bet type should be read against 24.37%, and every return against −3.67%.

The tight-screen figures in this section are the original September results, superseded where recomputed by the correction above. They are preserved to document the earlier publication; use the follow-up paper for the corrected strict and topic/time comparisons.

The published screen asks only for a heavy long-shot bet. A tighter screen adds the traits that reporting on suspected insiders points to: the event resolves within 48 hours, the wallet is under two weeks old, and it has almost no other bets. Vultax wrote each condition as a rule, fixed the rules and the tests in a protocol on 27 September 2026 before reading any outcome, and ran them over every qualifying bet.

1,331 bets meet all four, from 1,321 wallets in 720 markets, with $9.0 million staked, and they returned −9.9% on stake. The table sets them against established wallets making the same kind of bet in the same 48 hours, matched on category, price, time to the outcome and half-year. The gap is negative in every version: the whole group, the bets whose age was confirmed against public profiles, and a second implementation written from the protocol alone. None supports the idea that this screen finds bets that beat their price.

Politics and geopolitics, where the cases that motivate the screen sit, hold 448 of the tight bets. They won 17.9% at 18.8¢, 0.9¢ below the price, with an interval of about ten cents either side: no sign of an edge, and too few bets to rule out a moderate one.

Download table figures:CSVJSON with sources

The tight-screen figures in this section are the original September results, superseded where recomputed by the correction above. They are preserved to document the earlier publication; use the follow-up paper for the corrected strict and topic/time comparisons.

Four in five qualifying bets, 89,217 of 113,863, already fall within 48 hours of the outcome; age and breadth do the cutting. Within 48 hours uses the market's own clock: the start of its final run to 95¢ or more, defined the same way whichever side wins, and bets below 5¢ are left out because the market is already inside that run. New is under 14 days since the wallet's first fill on either side of any trade, claimed only for wallets first seen from February 2025, because earlier ones may predate the record. Almost no other bets is at most one other market before, with at least 90% of the wallet's dollars in this one. Everything is measured at the moment of the bet; later trading never decides who qualifies.

Age in a fill record can make a wallet look younger than it is, so every tight bet was also checked against the wallet's public profile and its earliest public activity: 1,036 of the 1,331 are under 14 days old on both, 241 had an older profile and 104 had activity from before the two weeks. The venue documents its profile date as profile creation, which is still not proof of when a person first opened an account.

In practice the screen mostly finds first trades: 998 of the 1,331 bets came from wallets less than an hour old. They also cluster, with 780 of them in 169 markets that hold two or more, which is why every interval here resamples whole markets.

Download chart figures:CSVJSON with sourcesSVG chart

The tight-screen figures in this section are the original September results, superseded where recomputed by the correction above. They are preserved to document the earlier publication; use the follow-up paper for the corrected strict and topic/time comparisons.

The tight bets arrive in clusters: several new wallets buying the same side of the same market in its last two days. Of the 40 clusters with five or more tight bets, 5 won and 35 lost; at the prices paid, 8 would have been expected to win.

The winners include episodes that made the news. All 23 tight bets on a US strike on Iran by 28 February 2026 won, as did all 19 on Axiom being named in an insider-trading investigation. Many more clusters were simply wrong: 33 bets on Jake Paul to beat Anthony Joshua, 30 against the removal of South Korea's president, 27 on the Patriots to beat the Seahawks. In January 2026 the screen flagged 15 bets on a US strike on Iran by the 14th, and all of them lost. Nothing the screen measures tells the clusters apart before the outcome.

Download table figures:CSVJSON with sources

Across every category the screen's own population lost money: −$45.5M gross on $1.24bn staked. The median flagged bet staked $5,460.

The $2,500 gate is doing work, and it is worth knowing which way. Drop it and keep only the price condition — any position held at 35 cents or less in a settled market — and the same window holds 62,148,514 positions across 1,883,778 wallets, $3.33bn staked, a 15.32% win rate at an average of 14.53¢, and a −5.35% return. The gate selects for higher prices, higher win rates and less-bad returns than long-shot betting in general. A screened cohort's win rate is not remarkable against the unscreened market; it is remarkable only against the screen's own base rate.

The flagged population is also not a population of repeat offenders. 10,270 of the 17,858 wallets, 57.5%, placed exactly one qualifying bet in sixteen months, and 2,622 placed two. 4,966 placed three or more, and those 4,966 are the only wallets any claim about a record can be about.

Download table figures:CSVJSON with sources

Read this by category or not at all. Sports is 61.7% of the flagged bets and behaves nothing like a rate decision, and the platform-wide number is mostly a sports number wearing a platform label.

Politics is where the bet type paid: $140.9M over 11,049 bets by 5,111 wallets, won 23.61% of the time against an average price of 19.66¢, for +16.87% gross. Culture returned −14.49%, Sci-Tech −26.29% and Finance −57.21%. The two subjects the market prices most confidently lost almost everything staked against them: 2.15% of 604 Fed-rate long shots won at an average price of 12.07¢, and 5.67% of 1,006 global-election long shots at 17.66¢.

Below a few hundred bets a category's return is a handful of resolutions rather than a rate: the 713 bets tagged World returned +105.67% and the 626 tagged Romania +69.39%, both on a small number of large winners. Categories are the market's own tag as the tape records it and each bet carries exactly one, so the Soccer and Games rows below sit beside Sports rather than inside it. The table covers every category with at least 600 flagged bets; all 219 are in the data file.

Download table figures:CSVJSON with sources

Download chart figures:CSVJSON with sourcesSVG chart

Across the whole screen the win rate, 24.37%, is slightly above the average price paid, 23.01¢ — 1.36 points of apparent edge — and the money still went backwards. Both are true at once because they are weighted differently: the win rate counts each bet once, and the return weights each bet by its stake. The large bets did worse than the small ones.

By category the gap runs both ways. Sports won 25.54% at 24.48¢, about a point of edge, and lost 1.77% of its stake. Politics won 23.61% at 19.66¢, four points of edge, and returned +16.87%. Finance won 15.68% at 21.89¢ and lost 57.21%. Culture is the clearest case of the two weightings disagreeing: it won 21.19% of its bets at an average of 17.62¢, three and a half points of edge per bet, and still lost 14.49% of the $13.3M staked. The extreme case is Fed rate decisions, where an average price of 12.07¢ was itself generous: 2.15% of those bets won.

That is the shape of a market that prices long shots roughly right and occasionally, in one category, does not. It is also why a win rate quoted on its own says nothing: 24% is excellent at 12 cents and poor at 30.

Download chart figures:CSVJSON with sources

33 of the 4,966 wallets with three or more qualifying bets won every one of them. The question a base rate lets you ask is how many you would expect if nothing but the market's own prices were at work.

Take each wallet's qualifying bets and assume each wins with the probability the market gave it, which is its own average entry price. Multiply those probabilities together and you have the chance that wallet goes perfect; add the result across all 4,966 wallets and you have the number of perfect records chance produces. That number is 18.5. Observed is 33: 1.79 times chance. Held instead against the flat 24.37% platform win rate, the expectation is 20.4.

Where the records sit matters more than the ratio. 32 of the 33 come from wallets that placed three or four qualifying bets, the band where chance manufactures perfect records; 18.02 were expected there. One came from the 5-to-9 band against 0.45 expected. In the 10-to-24 band, 890 wallets, none is perfect and none was expected to be, and the best record there is 70%. Among the 753 wallets with 25 or more, the best is 48%. A run that reflects something real should survive a longer sample. None of these reaches one.

The 33 are not nothing: 106 bets between them, $3.53 million of gross profit, a median wallet age of 7.3 days at the first qualifying bet and a median of 18.5 days still trading after the last one. Three of the 33 have an age that cannot be trusted, for the reason set out further down. They are a reasonable place to start asking questions. They are not, by themselves, an answer.

Download table figures:CSVJSON with sources

The bet rule is only the first half of the published screen. The second half is a wallet filter: keep the wallets that won more than 75% of their qualifying bets and concentrated on a small number of markets and topics. Run that filter on this reproduction exactly as written, with no minimum number of bets, and 2,070 wallets come back, holding 2,325 qualifying bets, $19.06 million staked and $58.67 million of gross profit. Those are excellent-looking numbers.

They are also almost entirely an artefact of how few bets each wallet placed. 1,888 of the 2,070 wallets — 91.2% — placed exactly one qualifying bet and won it, which scores as a 100% success rate. 2,026 of them, 97.9%, placed one or two. Only 44 wallets in the whole platform cleared 75% over three or more qualifying bets, and only 12 over five or more. Adding a concentration condition changes almost nothing: requiring three or fewer distinct categories still leaves 2,068 of the 2,070.

This is the arithmetic every outcome-conditioned screen has to answer. At a 24.37% base rate, one qualifying bet in four wins; among 17,858 wallets that is thousands of one-for-one records before anybody has done anything unusual. The filter is not wrong to start there — it is a way of narrowing a very large tape to something a person can look at. It is a starting list, and its size tells you how much work is left rather than how much has been found.

Download table figures:CSVJSON with sources

A separate signature, reported alongside the screen, is a wallet created shortly before the trade. Vultax holds a first-seen date for all 2,404,822 wallets that appear on either side of a fill in the window, and 14,733 of the 17,858 flagged wallets have one that can be used; the other 3,125 are covered in the limits below.

Sorted by how old the wallet was when it placed its first qualifying bet, the youngest band is the worst. 4,860 wallets that qualified within 24 hours of their first-ever fill placed 27,910 bets, staked $381.3 million and lost $19.58 million: −5.14%. 23.50% of those bets won, and the average wallet in that band won 18.31% of its own. The 24-to-48-hour band lost 7.12%. The only band that finished ahead was two to seven days old, and it returned +0.26% on $130.8 million.

So the "created hours before trading" filter, applied on its own across the whole platform, selects a population that loses more than the average flagged bettor does. It is a real property of some wallets that turned out to know something. It is also the ordinary signature of a new account taking one big swing.

Download table figures:CSVJSON with sources

A first-seen date is where a record starts watching a wallet, which is not always where the wallet started. Vultax checked 500 wallets from this record against the oldest trade Polymarket's public activity API returns for them. Of 250 first seen in January 2025, the month the record begins, 134 (53.6%) had an older public trade, by a median of 43 days. Of 250 first seen from February 2025 onward, 17 (6.8%) had.

The pattern is exact. All 151 older trades date from before 1 January 2025, and the other 349 wallets matched the record to the second. Inside the window a first-seen date is reliable; at its start it is wrong more often than right. The 17 later wallets are returning traders: each traded in 2024, was absent from the first month or more of the record, and came back, a median of 146 days after its oldest trade.

For the bands above, that means a wallet counted as hours old can be a 2024 account returning, at a rate of about one in fifteen away from the start of the record. It makes the youngest band look newer than it is; it does not rescue its return. An alert that calls a wallet fresh can settle the question with one request: the activity endpoint filtered to trades, oldest first, limit one.

Download table figures:CSVJSON with sources

Download chart figures:CSVJSON with sourcesSVG chart

A screen is a filter, not a finding. This one selects on price, size and speed — three things visible to anyone reading the tape — and then the outcome does the rest of the work. Sorting the result by win rate guarantees a profitable top group whatever the population: among the 4,966 wallets with three or more qualifying bets, every band above a 25% win rate made $126.7 million between them, and the band below it lost $143.2 million. The wallets at the top of that sort are the wallets that won. That is what the sort is for.

Wallet age measured from the record is a lower bound, not a birth date. Vultax's record starts on 1 January 2025, so a wallet that traded in 2023 and again in 2025 looks new. A stratified sample of 500 wallets checked against Polymarket's own public activity endpoint, 500 of 500 resolved, found older trades for 134 of 250 wallets (53.6%) first seen in January 2025 and 17 of 250 (6.8%) first seen from February onward. These were separate samples restricted to wallets with at least five fills; no validated eligible-population weights were applied, so they do not yield a platform-wide error rate. 3,125 of the 17,858 flagged wallets, 17.5%, fall in the unreliable cohort and are left out of every age figure on this page.

Two things the screen cannot see at all. It counts a position the same way whether the wallet crossed the spread to get it or was sitting on the other side of somebody else's order, so a market maker filled passively at 30 cents looks identical to a buyer who chased the price. And a wallet is a Polymarket proxy address, not a person: one operator running ten wallets counts as ten wallets, and the address clustering that would fix that is not part of this measurement. Every claim of the form "N wallets did X" inherits that.

The appear, win, withdraw pattern cannot be completed from this record either. It is a record of fills — maker, taker, price, size, fee — with no deposits, withdrawals or redemptions in it. What is visible is dormancy: whether a wallet stopped trading after its last flagged bet. And the plainest limit of all is that a public screen is not the instrument that matters most. The venue runs its own surveillance and refers accounts, and the CFTC's February 2026 advisory names insider trading on event contracts as conduct it polices on any designated contract market. Those parties can see identity, funding and intent. A screen on the public tape can see three numbers.

Download table figures:CSVJSON with sources

The screen reproduced here is published work. In August 2026 the Anti-Corruption Data Collective released “Classifying Insider Trading Risk: Analyzing the Longshot Betting Ecosystem on Polymarket”, which examined settled Polymarket markets through 5 May 2026, defined a long-shot bet as more than $2,500 wagered at 35 cents or less, and sorted the platform’s long-shot wallets into four groups. 556 of them — the group it called Orcas — met a profile of betting in a small number of markets and on very few topics with a success rate above 75%. Reuters, which reported the work first, described 152 of those wallets as concentrated on military and defence markets, winning about $8 million at an average 97.2% win rate against roughly 52% for long-shot bets in those markets generally.

CNN reported the following day that a senior Polymarket official said the company had referred dozens of accounts showing signs of military insider trading to the Department of Justice, referrals that predated the report, and that its internal surveillance monitors around 150 further signals and trading patterns. Polymarket publishes a market-integrity policy that prohibits trading on confidential information and describes real-time surveillance and referral to law enforcement. Polymarket itself has not been accused of wrongdoing.

Bloomberg Businessweek covered the same question in July 2026, reviewing about 34,000 Polymarket transactions flagged by the analytics platform Polysights between August 2025 and June 2026. It reported that 57% of the flagged wallets had been created less than 24 hours before their flagged trades, and that 38 connected wallets betting across 90 markets tied to Iran and Venezuela won 98% of them for about $1.6 million. It noted that a flag does not prove wrongdoing.

In September 2026 the Wall Street Journal described a different shape of the same problem. Wallet clustering by Bubblemaps linked 19 Polymarket accounts that went 41 for 42 on earnings bets covering roughly 18 companies audited by KPMG, beginning in November 2025, for about $22,000 of gains. KPMG said it has zero tolerance for the misuse of non-public client information, and Bubblemaps said clustering on its own does not prove insider trading. That case is the argument for address clustering as a technique, and clustering is exactly the piece this measurement does not have.

The academic version is “From Iran to Taylor Swift: Informed Trading in Prediction Markets” by Joshua Mitts of Columbia Law School and Moran Ofir, an SSRN working paper from March 2026. Screening every Polymarket contract from February 2024 to February 2026 with a composite score built from bet size, profitability, pre-event timing and directional concentration, it reports 210,718 suspicious wallet-market pairs, a 69.9% win rate more than 60 standard deviations above a permutation null, and roughly $143 million of anomalous profit. A later working paper sets out the limitation that this study meets from the other side: a score with profitability in it can only be computed once the market has resolved, and it does not separate skill from luck by itself.

Window: 1 January 2025 to 28 April 2026, 714,691,643 Polymarket fills, settled markets only, 662,215 of them. A bet is one wallet-market-side episode whose rolling one-hour stake reached $2,500 over fills where the side held was priced at 35 cents or less; profit and loss covers that episode's long-shot fills only. Stake is shares multiplied by entry price, and gross profit is winning shares minus stake. Groups whose total long-shot stake never reaches $2,500 cannot contain a qualifying window and are removed before the rolling sum is taken.

Two differences from the published version are worth stating, because they move the count. That version says more than $2,500; this reproduction uses $2,500 or more, which is how the first report of it rendered the rule. And this record runs from 1 January 2025 to 28 April 2026 and includes the multi-outcome negative-risk markets that a binary-only tape leaves out, where the published version covered settled markets through 5 May 2026. The two populations are not the same set of bets and their counts should not be expected to match; what carries across is the bet type and the base rate for it.

Returns are gross of fees: Polymarket charges takers a fee on some categories, set by the category's rate and the price as contracts × rate × p × (1 − p). Categories are the market's own tag as the tape records it, one per bet, so Soccer and Games are rows beside Sports rather than parts of it. Returns are stake-weighted; win rates count each bet once, which is why a category can win more often than its price implies and still lose money.

The tight screen's rules and tests were fixed in a protocol at 20:45 UTC on 27 September 2026, before any outcome was read. A second implementation, written from the protocol alone, reproduced the 113,863 qualifying bets and the 89,217 within 48 hours. Entry price is the average cost of the shares bought in the qualifying hour; established wallets are held to the same February 2025 start as the tight group. The cluster view and the two charged cases were added after the outcomes were read and are descriptive: the charged wallets were matched to the complaints by quantity, date and profit, a strong match but not an identification, and no wallet is named.

The window closes on 28 April 2026 and says nothing about the months since. Wallet identity is an address, not a person, and no address is published here. Vultax's calibration study measures the same favourite-longshot effect from the price side, and the trading-cadence study describes the population these wallets sit inside.

Nothing here is a recommendation to trade, and no wallet selected by this screen is being accused of anything. Challenge a figure at [email protected]; material changes are logged in the changelog.

The wallet-age check fixed its sample on 16 September 2026: wallets in each group ordered by hash, the first 250 with at least five fills taken, and each compared with its oldest public trade on 19 September. The two groups are reported side by side and are not weighted up to a platform-wide rate. Download all 500 rows, the summary figures, the method notes and the reproduction bundle.