The Polymarket book that beat the odds on KPMG's clients

01 · Seven hours before the bank filedTwo accounts, one minute apart

At half past four in the morning, London time, on 14 January 2026, an account called greatfan1983 bought a bet that Wells Fargo would miss its quarterly earnings estimate. It paid about 82 cents for each dollar the bet would return. One minute later an account called Rossatttttt bought the same side of the same market. Seven hours after that, Wells Fargo filed its results with the Securities and Exchange Commission and missed.

Polymarket is a site where people bet on how real events turn out. Elections, football, the price of bitcoin next week. One of its sections covers company earnings: four times a year a listed company reports what it made, analysts publish a consensus estimate beforehand, and the market asks whether the company will come in above it. You buy a share in the answer you believe. If you are right the share pays a dollar, and if you are wrong it pays nothing, so the price of a share is the crowd's own estimate of the odds. A bet at 82 cents is the market saying it is about 82% sure.

That price is the only yardstick that follows. Our blockchain data lake holds every Polygon transaction, and the site settles on Polygon, so we can read the whole trade record of any account and compare what it paid against what happened. In August a report in the Wall Street Journal said federal authorities were looking into an employee of KPMG over a bet on one company's quarterly earnings. A set of accounts on the site had already been going round as a possible match. We took those accounts, pulled every bet they ever placed, and scored each one against the price it paid. There were 19 of them when we started and we found three more on the way, so every figure here counts the 19 unless it says otherwise.

Most of what had been said about them holds up. Two of the auditors had been filed to the wrong firm, our own early attempts got several things wrong of their own, and the question everybody actually wants answered is the one a ledger cannot reach. It records the price and the minute and nothing about who held the keys. Nothing here identifies a person, and we make no claim about how anybody learned anything.

02 · One account, then manyHow the book changed shape in a week

The story starts with a single account, because none of the others existed before it. On 7 October 2025 one called Youaregu was created, the earliest of the 22 on the site's own records, and for three weeks it bet on cricket and on three company earnings markets, none of them audited by KPMG. It finished October up about $3,650, nine tenths of which came from one bet on Coursera. A good month, and nothing about it looks unusual.

Then, over 39 hours on 3 and 4 November, it opened bets on the quarterly earnings of three KPMG audit clients, Hims & Hers, DoorDash and TKO, and it kept adding through the 5th until $19,078 was riding on them. All three came in. It sold part of the Hims & Hers bet 14 minutes after that company filed its results, and part of DoorDash 8 minutes after DoorDash filed. Youaregu made $11,417 on the three and never opened another bet.

The money was already coming out while that last trading went on. Youaregu began withdrawing on the afternoon of the 5th, emptied its wallet in 18 transfers, and finished on the evening of the 6th with the remainder, $78. Twenty-four seconds after the $78 left, that money, less the bridge's fee, landed in an account called lawrence3, which had been created three minutes earlier. Nine accounts were created that day, two the day before and four more on 8 November. Not one of them ever bet more than $2,715 at a time, where Youaregu had staked $10,363 on a single market.

Two weeks in November carry nearly half the earnings work: the week those accounts appeared, and the week a fortnight later. The tall week in March is a single day of bitcoin markets bet in five-minute slices, 30 of them on 8 March, which has nothing to do with company results. The last bet was placed on 26 March 2026, the last sale went through the next evening, and the last dollars left on the 28th. Nothing has moved since.

03 · The book, split three waysThe edge sits in one place

Every bet these accounts took falls into one of three groups. Earnings markets where the company's auditor was KPMG. Earnings markets where it was somebody else, Ernst & Young, PwC or Deloitte. And the rest of what they traded, which turned out to be bitcoin 5-minute markets, cricket and crude oil.

The three groups are the same accounts, betting similar sums in the same weeks. That is what makes the comparison worth anything. The second and third groups are the control on the first.

Read the second row before the first. A win rate on its own says nothing, because a bet bought at 90 cents is supposed to win nine times in ten. The question is always whether a book beat the prices it paid. On other firms' audit clients this book did not beat them. It bought at an average of 75 cents, which predicts 11.3 wins out of 15, and it got 11. About seven in ten trading books would do that well or better. On sport and crypto it came out a shade ahead of the odds and still lost money over five months.

The first row is the outlier. The same accounts, buying at a comparable average price, won 19 of 21.

One market needs explaining, because it is the one place the accounts worked against themselves. On General Mills, greatfan1983 bought the losing side of the consensus at 20 cents on 14 March. Two days later, after the price had moved a long way, Protiyaar bought the other side at 14 cents and lost. We score that market as a loss, which is the harshest of the three reasonable ways to handle it. Scoring it by the side carrying most of the money makes the record 20 and 1, and dropping it altogether makes it 19 and 1. The table below shows what each choice does to the odds.

A note on that figure, because everything above turns on it. It treats each market as one event rather than each account's bet. 7 of these accounts bought the same side of one StoneX market, and that is a single thing happening. Counting bets instead would produce a number hundreds of times smaller, and it would be wrong.

04 · Where the profit actually came fromThree bets, one account

13 of the accounts held a position on a KPMG audit client, and the winnings are not spread across them. Youaregu's three bets were opened on 3 and 4 November, before any of the other accounts existed, and they account for more than half of everything the group made on those markets. The next-biggest earner made less than a tenth.

That concentration is worth being honest about. It weakens any argument built on the money made, since one account betting large enough can carry a whole book. The record survives it. Youaregu held 3 of the 21 markets, and other accounts held 2 of those 3 as well, so taking it out of the set removes a single market. The 20 that remain were won 18 times against prices that predicted 13.7, which happens about one time in 78.

05 · The clockAlmost every bet was placed before the filing

When a listed company reports, it files the numbers with the SEC on a form called an 8-K, and the SEC stamps the moment it accepted the filing. That stamp is a public, independent clock. We took the stamp for every one of these earnings results and compared it against the moment each bet was opened and the moment each dollar went in.

This is the control that matters most, and it is worth spelling out why. Buying a prediction market in the seconds after a result becomes public is a known and legal strategy: the price has not caught up yet and a fast trader takes the difference. A book full of that would look brilliant and mean nothing. Almost none of this one was. Less than half of one percent of the money went in after a filing, and nearly all of that was added to bets that were already open.

The repeats are the part that reads oddly. Two accounts bought Wells Fargo to miss a minute apart, about seven hours before the bank filed. Five accounts bought CarMax to beat within four hours of each other, six to ten hours before CarMax filed. Five bought Home Depot to miss in the 14 hours before Home Depot filed.

06 · Nineteen strangers on PolygonWhat the bridge records show

On Polygon these accounts look like strangers. None ever sent money to another, and none ever took the other side of another's trade. Every address they have in common carries everybody's traffic, because the address that funded all 19 belongs to a bridge, a service that moves money between chains, and it has paid out to more than two million other wallets. That address is 0xf70da978…dbef.

That emptiness is why a search of Polygon alone finds nothing, and it is why our own first pass concluded, wrongly, that the accounts stood apart. The money leaves over the bridge, and the bridge publishes a record of every transfer it carries. Follow one account's money out and another's in, and the two halves meet in the middle.

We traced the second of those chains on our own indexes, with the bridge's records set aside, to be sure the pattern was not an artefact of somebody else's bookkeeping. Four transfers, two chains, 50 seconds, and the only value lost along the way is the bridge's fee. On Polygon that same movement looks like two strangers taking deposits from a service used by millions.

Eleven of these hand-offs exist. Every one runs in the right order, with the wallet in the middle receiving before it pays out, and the first of them is the one that started the second phase of the book. The figures below are the amount measured at that middle wallet, which is why they sit a little under the amount that left the sending account: the bridge takes its fee on the way in and again on the way out.

Those 11 tie 16 of the 19 accounts together. A 17th, KARLSON1970, joins through orisonpro's funding wallet, which merged its withdrawal with one from pops4532 into orisonpro's first deposit. The last two share payout wallets with the others. browjobhyd sent money to 0xabf92618…db9e, which took bridge transfers from eight of them, and kundragame to two wallets that three and four of the others also used. Put the accounts and the wallets between them into one graph and all 19 fall into a single connected group.

That wallet is also where four of the accounts sent their last withdrawals. On 8 March 2026, within 17 minutes of each other, browjobhyd, molgum12, pops4532 and orisonpro each sent money to it over the bridge, $1,328.51 between them, and none of the four withdrew anything again.

Eight of the 19 accounts got their very first money from another account in the network, two of them from accounts we only found later. For seven it was a single withdrawal, landing between 90 seconds and just over six minutes after the account was created. orisonpro's first deposit merged two, and the first of them went out 83 seconds after orisonpro was created. One chain runs two steps deep. Youaregu's last withdrawal became lawrence3's first deposit, and two days later lawrence3's own withdrawal became showredto's.

None of it would mean anything if those middle wallets were services. They are ordinary wallets. Measured on our own indexes each one has a handful of wallets it has ever dealt with, between 1 and 6.

There is a second signal, and it does not depend on money at all. The site records when an account was created, and we hold a census of 2.35 million of them. Five of these accounts were created on the evening of 6 November 2025 inside a single 51-minute window. Four more on 8 November inside 44 minutes. Two on 14 January, 11 minutes apart. Thirteen of the 19 within four days. To see how ordinary that is we drew 20 real Polymarket accounts at random from the 855,280 created over the same months, and asked how often 5 of them land inside any 51-minute window. In 200,000 draws it did not happen once.

07 · Where the money came fromA trail that ends at three exchanges

The money going in was kept carefully apart. Of every payment into these 19 wallets that was not a winning bet being cashed, 57 came over the bridge, and each account was funded from its own separate wallet. There are 20 wallets for 19 accounts, and not one of them fed two. None carries a label of any kind. One step further back, most of those funding wallets turn out to have been topped up from somewhere ordinary.

A single wallet feeding 15 of 17 looks like the common funder the case had been missing, until you see whose wallet it is. An exchange pays out to everyone who withdraws, so sharing one proves nothing at all.

What the funding side does give is a clean path rather than a cluster. For 7 of the accounts, the amount that left the exchange, the amount that sat briefly in the single-use wallet, and the amount that arrived at Polymarket are the same figure within a few cents on the same day. Nothing is pooled or mixed anywhere along the way. For the 17 Solana wallets the trail therefore ends at accounts held at Binance, Bitget and Crypto.com, three sets of identity documents that only those exchanges can read. The three Ethereum wallets were paid in USDT by addresses that carry no label at all.

The last row of that table is the one worth sitting with. Fresh money from exchanges got the book started. After that, nine of the Solana funding wallets were topped up over the bridge with money withdrawn from other accounts in the network, and the two wallets behind gridconwho and Protiyaar were filled that way and no other. Exchange money into the Solana wallets stopped on 18 December 2025. New money still arrived in January, when $6,744 of USDT reached browjobhyd and hulklinkme through Ethereum wallets from two senders with no label.

08 · Why no leaderboard caught thisNothing to find account by account, and three more accounts

The site posts leaderboards and several services rank its traders. None of them flagged this, and the reason is that there was nothing to flag.

We scored each of the 13 accounts on its own record against the prices it paid. The best result in the whole set has a probability of 0.071, and not one clears the usual 0.05 threshold. Ranked against the 72 other wallets that took 3 or more settled bets in the same markets with real money behind them, no outside wallet beats two standard deviations above its expected wins, and these accounts sit in the middle of the table.

The book is only visible if you already know to add the accounts together. Splitting a position across wallets small enough to look like ordinary retail traffic is what kept it off every screen, and any detector built on single-wallet performance will keep missing books like it. The one account that did bet large enough to stand out, Youaregu, stopped trading in the same week the small ones appeared.

That raises the obvious question of how many more accounts belong to this set. We answered it by taking every wallet that held a bet in those 21 markets with real money at stake, 489 of them, and asking the bridge whether any had ever sent money to the same destination wallets. Exactly one had: an account called perlgrow, 0x7d7a4812…e192, created on 18 November 2025, 50 minutes before one of the others. It staked $3,723 across 7 markets for $759 of profit, and it bought two of the same markets in the same minute as two of the published accounts. Two hours after it was created, it also traded directly against one of them, on the other side of a $5.95 bet molgum12 placed on Amer Sports. That makes the set 20 accounts rather than 19.

Running the same kind of check from the funding side found two more, and neither ever bet on these markets. boosthero and bostonguy were created on the evening of 6 November, inside the same 51 minutes as five of the others, and each was funded by a single-use Binance withdrawal. Both bet only on Franklin Resources, an other-auditor company, and on 7 November one seller's order filled the resting buy orders of six accounts in this network in a single transaction, boosthero's and bostonguy's among them. On 8 November the two sent their money into the funding wallets of two accounts created minutes earlier, and 20 seconds later it became greatfan1983's and makertogo's first deposits. That makes 22, and it is still a floor, because an account that moved its money some other way would not show up in either search.

09 · What we can and cannot sayThe limits, and the corrections

The Journal's report of 27 August 2026, summarised elsewhere, described two matters federal prosecutors and the Commodity Futures Trading Commission were pursuing. One was a serviceman who bet on military action, since charged. The other was an employee of KPMG, said to be under inquiry over whether one public company would beat its estimate. No charges against that person have been made public.

We cannot connect these accounts to that inquiry, to KPMG, or to any person, and we do not try. The inquiry as reported covers a single company. The pattern we measured runs across 19 companies. A ledger records what an account paid and when, and it records nothing about who held the keys or how they learned anything. Everything above is a trading record and a set of money moves. What produced them is a question this data cannot answer.

Putting a number like one in 120 in print also owes the reader its working. This one has now been through ten audit passes, and most of them forced a correction. They are listed here because they change how much weight the rest deserves.

Two further things are left open rather than resolved. 2 of the 99 markets have no recorded outcome in our data and are excluded from every count. And 0xedb181ff…e112, a wallet that has dealt with 16 addresses in its whole history, paid $100 into one account's funding wallet a minute before that wallet funded the account, and also paid two addresses these accounts cashed out to. The same wallet sends worthless transfers to families of lookalike addresses, which is the signature of an address poisoning operation. It is either a wallet inside this group or somebody preying on it. We cannot tell which, so we say nothing more about it.

10 · How this was measuredThe method, in full

Every figure comes off Bitquery's own indexes of Polygon, Ethereum and Solana, joined to filings from the SEC and to the public transfer records of the bridge. Anyone can reproduce it: the same data is available through our Polymarket API and the queries run unchanged on our MCP server.

If you want the prices these accounts paid rather than our summary of them, the per-market trade record is queryable directly, and the five-minute markets and the share of Polymarket that is automated are covered in their own pieces. For the mechanics of how a Polymarket result is settled, and who gets paid when the oracle is late, see our work on the oracle race. Polygon's own dependence on this one application is measured here.

11 · What is leftA pattern, and a question a ledger cannot answer

So what were greatfan1983 and Rossatttttt doing at half past four in the morning? What the ledger shows is what the other accounts were doing for five months: buying a company's earnings result before the company filed it, on the clients of one audit firm, at prices that underrated their side often enough to pay. Where that came from is a separate question, and this data has no way to reach it. A ledger records the price and the minute and nothing about who held the keys.

The shape of it is the part that stays with you. A book of this size does not need 22 accounts. Splitting it that way paid a bridge fee on every transfer in and out, and what it bought was a crowd of small traders that every leaderboard read as ordinary retail, which is why nobody had flagged any of them. That same split is what makes the record plain once you add them back together.

Two of the accounts never came back to collect. $1,310.56 is still sitting where they left it.

12 · The recordAddresses and transactions behind this piece