Shelfware Labs
SW-05·A forensic decompositionLive

The Wallet That Fooled Me Longest

The single most convincing account on the board — human, diligent, diversified, compounding a bankroll from nothing. I was sure this one was the real thing. So I took it apart: the P&L, the luck-o-meter month by month, the sizing, the timing, and the one early streak the whole thing rests on. Every piece that made it look like skill turned out to be something else.

In the last story I audited the whole Polymarket leaderboard looking for one account worth copying, and didn't find one. This is the account that made me look hardest before it fell.

It wasn't the flashiest name on the board — not the biggest bankroll, not the gaudiest win rate. It was better than that: it looked professional. Where the whales looked like lottery winners and the consistent accounts turned out to be bots, this one looked like a person who had quietly figured something out. It held up against every quick test I threw at it, and then a few slow ones. I had it flagged as my single best candidate — the one I was actually going to follow with real money. Here is everything that made it convincing, and everything that took it apart.

Why it was so convincing

Start with the parts that were real, because they were genuinely impressive. This account did all the things a sharp is supposed to do:

The green flags — all real
🕒 late76% of its game bets were placed in the final hours before kickoff — reacting to real news (who's hurt, who's actually starting), not just betting on a schedule days ahead.
🧠 its own calls78% of its bets were picks the rest of the crowd wasn't making — it wasn't copying anyone. And on the bets where it went against the other top accounts, it turned out to be right 58% of the time. That's what someone who actually knows better looks like — not someone getting lucky.
🌍 spread outA record across six different sports and 440 bets over four months — not one lucky night on one game.
📈 pressing its winsIts bets got bigger as it won — which looks like exactly what a person with a real edge does: put more behind it as the money comes in.
😴 human hoursIt went quiet from about 4 to 7 a.m. every night and bet a few times a day. A real person, asleep on a schedule — not a bot grinding around the clock.

Put that together and you have a diligent, independent, diversified human, compounding a self-made bankroll on late information. That is exactly what a real edge is supposed to look like from the outside. I believed it for a long time. Then I stopped looking at what it looked like and started looking at what it did.

The P&L curve that sold me

This is the chart that did most of the convincing — and, it turns out, the one that hid the most. On this scale it looks like it barely moves for two months, then goes nearly vertical through May and June: the look of an edge that has finally found its gear. That vertical ramp is exactly what sold me.

cumulative profit, daily — the real feed
$360k$180k$0 MarAprMayJunJul $341k · Jul 4 then −38%
The real daily curve. It looks like it barely moves for two months, then ramps hard through May and June (+$173k in June alone) to a peak of $341,538 on July 4, before handing back 38% of the whole pile in the ten days after. This is the chart that sold me — but "barely moves" is a lie the scale is telling. There's a whole first act buried in that flat line.

The dollars are so much bigger at the end than the beginning that an ordinary chart crushes the entire first act into that flat stretch. So here is the exact same profit on a logarithmic scale — where every tenfold move (a few hundred dollars to a few thousand, a few thousand to tens of thousands) takes the same vertical step, giving the early days room to breathe:

the same profit, log scale — the first act, revealed
$1k$10k$100k MarAprMayJunJul $341k · Jul 4 $18k · wk 1 then gave it all back →
The identical numbers, on a log scale. Now the opening is visible: a spike to ~$18,000 in the first week (the invisible bump from the chart above), then it gave almost all of it back and spent April slightly underwater — before the real climb, three orders of magnitude from a few hundred dollars to $341k. That early spike-and-collapse isn't noise around the story. As the next chart shows, it is the story.

If you stop here, you copy this account. I almost did. The problem is that a rising P&L answers the wrong question. It tells you the account made money. It does not tell you whether the account was right — and those two things come apart exactly when someone is betting bigger and bigger into a hot streak.

Win rate is a decoy

Dollars are the wrong question. So is the one everyone reaches for next: win rate. A win rate on its own carries no information whatsoever, because the only thing that gives it meaning is what you paid. Winning 59% of your bets sounds strong — but if you bought them at prices implying 65%, you lost. Winning 24% sounds like a catastrophe — but if the market said 20%, you're printing money. The number is empty until you put the price next to it.

That's not a hypothetical. It's this exact account, split by the kind of price it was buying:

the same account, three win rates — only one is any good
What it was buyingbetsthe line impliedit actually wonvs the line
Favorites (over 60¢)7965%59%−5.0 pts
Coin flips (40–60¢)34053%58%+5.2 pts
Longshots (under 40¢)2530%24%−5.6 pts

Three win rates — 59%, 24%, 58% — and the only good one is the unremarkable-looking middle one. On favorites it won 59% and was five points worse than the market. On longshots it won 24% and was nearly six points worse. Its 58% on coin flips is the best thing it did all year. So when you see a bettor's win rate quoted with no price attached — on a leaderboard, in a screenshot, in a pitch — it isn't evidence. It's decoration.

The luck-o-meter was leaking the whole time

Which leaves exactly one honest ruler: the gap between what they won and what they paid to win. That's the luck-o-meter — a plain z-score measuring how far the wins ran above what the prices implied, weighted by how many bets you're judging. (A reading near 1 is what luck alone throws off; near 3 is real skill; below 1 is noise. The full explainer is in the leaderboard piece.) So I ran it month by month. Here is the same account, same period, graded on the only thing that counts:

luck-o-meter (beat-the-line z), by month
+2+10−1 MarAprMayJunJul luck floor
The same account, graded on skill instead of money. Only its lucky first month clears the luck floor — the level luck alone throws off. After that it falls every single month1.95 → 0.97 → 0.72 → 0.57 — and goes negative in July (−0.92): it won less often than the prices implied. The June that added $173k? Its skill score that month was 0.57 — noise. It wasn't getting better. It was betting more into a run that was already cooling.

Two charts, one account, opposite stories. The dollars went up and to the right; the skill went down and through the floor. That gap is the whole thing — it's the fingerprint of a lucky streak being pressed, not an edge being compounded.

So how did it make $341,000?

Fair question — and if I can't answer it cleanly, you shouldn't believe the rest. If the edge was fading to noise, how did the money keep climbing, so steadily, to a third of a million dollars? Because past a point it stopped being about the edge at all and became about the size.

Watch the bets themselves. For its first three months the account staked a fairly steady couple thousand dollars a game. Then, as fresh money poured in, they ballooned:

the typical bet, by month
$10k$5k$0 MarAprMayJunJul $2.3k $8,800 · Jul
The median bet climbed from about $2,300 a game to $8,800, and its biggest single bets from $6,100 to $63,700. That growth wasn't compounded out of winnings — the account was fed capital, topped up from about $600 to $128,000 of outside deposits by June. More money in, bigger bets out.

Now set that beside the only number that means anything — how far it beat the line — and the account gives itself away:

the edge collapses while the money explodes
Monththe line impliedit wonvs the linestaked that month
March53%69%+16.1 pts$108k
April54%67%+12.4 pts$72k
May55%57%+2.7 pts$389k
June54%56%+2.2 pts$975k
July50%44%−5.8 pts$813k

Read the two ends of that table against each other. The prices it paid implied about 54% every single month — it was buying coin flips from the first day to the last. In lucky March it won 69% against a 53% line: +16 points, the spike. But in May and June — the months that actually made the money — it beat the line by 2.7 points, then 2.2 points. That is the entire edge behind a third of a million dollars: winning about two games in a hundred more than the market said it would. Two points isn't a skill. It's the size of the gap luck throws off all day long. What changed wasn't the edge — the edge was shrinking. What changed was the money: staked per month went from $108,000 to $975,000. Nearly a million dollars a month riding a two-point maybe. That's not an edge compounding; it's size × variance — and a few big bets carried it, the ten largest wins (mostly five-figure World Cup over-unders) making up more than a quarter of all the winnings.

Put it in terms of a game everyone reads the same way. Four months and 440 bets sounds like a real track record — until you time it against a blackjack table, which deals around a hundred hands an hour. The account's entire career is about one evening at the table. The World Cup run that made half its money and earned its spot on the leaderboard? Roughly an hour of that evening. Now picture the player: he sits down, and across one session runs a small stack into a fortune — by betting bigger and bigger, shoving a quarter, a third, sometimes the whole stack onto a single hand, and topping up from the cashier whenever it dips. Genius? Or do you already know how the night ends if he keeps his seat? The card counters casinos actually fear are the opposite of this: their edge is about one percent, and they bet a small, fixed, disciplined fraction of their roll precisely so an ordinary cold streak can't bust them. Anyone shoving a third of the stack onto a coin flip isn't a sharp — that's a gambler on a heater, and every heater ends. The only question is when the coin turns.

It turned in July. The bets were as big as ever — an $8,800 median, singles up to $63,700 — but now it won 44% on prices implying 50%: nearly six points worse than the line. The account handed back $128,000 in ten days. A real edge doesn't invert that fast; a big bankroll on a coin that stopped landing does. That symmetry — up fast, then down just as fast at the very same stakes — is the signature of luck, not skill.

The diversification was a mirage

"Six sports" was one of the things that most convinced me — a one-tournament fluke can't hide inside a diversified book. So I split the profit by sport. It does hide; you just have to look.

where the "edge" actually lived — profit and skill, by sport
Sportbetswin%luck-o-meterprofit
World Cup11055%0.70+$114.7k
Basketball8658%0.77+$47.5k
Club soccer4569%1.67+$29.2k
NHL1275%1.33+$25.7k
MLB6548%−0.69+$11.7k
Tennis7662%1.14+$2.3k
Everything else4650%−0.30−$13.4k

Watch the win% column do its decoy work one more time: 75% on hockey, 48% on baseball. Read those two alone and you'd copy the wrong one — only the luck-o-meter beside them tells you the hockey was twelve bets of noise and the baseball was a genuine loss to the market. Now look at what's actually holding the record up. Half the profit is one World Cup — a tournament that expires, taking its markets and its edge with it. Take that away and the "diversified sharp" is a good basketball run, a strong twelve-bet NHL sample, and then trouble. Look at MLB: sixty-five bets, and against the closing line the luck-o-meter is −0.69. It lost to the market on baseball — it only shows a profit because it happened to win a few favorites at good prices. An earlier, cruder pass had graded its baseball at "+33%" and called it a category strength. Measured honestly, baseball was a category it was bad at, bailed out by variance. The diversification wasn't insurance against a fluke. It was camouflage for one.

And stay on that World Cup a beat longer, because it's the load-bearing wall — and it can't hold the weight. A World Cup is one month of games; this was 110 bets. That is nowhere near enough to tell a real edge from a lucky run — at that sample size the luck-o-meter simply can't separate skill from variance, and here it reads 0.70, squarely in the noise. Yet that single hot tournament is exactly what put the account on the leaderboard in the first place: the winnings that made it visible — that made me notice it — came from a sample far too small to mean anything. The thing that got it on my radar is the thing least able to prove it was real.

It all came down to one five-day streak

Here's the part that finally cracked it open — and it's the thing I should have looked at first. I stopped reading the account's betting history and went to look at its bank statement. Prediction markets like this one settle on a public blockchain, which means every dollar that has ever moved in or out of the account is out there for anyone to read. So I read it. And the money told a far simpler story than 440 bets across six sports.

The account opened on March 14 with a single deposit: $606.60. Six hundred dollars. Over the following five days it topped up by barely another twenty. And by March 19 — five days later — the balance had crossed $24,000. It turned six hundred dollars into twenty-four thousand in under a week, and it did it on winnings alone: total fresh deposits across that whole stretch came to $621. This is the same run you half-saw at the very top of the page — the little spike-and-crash the log chart pulled out of that flat line is this.

account balance, first five days — straight off the blockchain
$24k$12k$0 141516171819 March $24,181 · day 5 $607 seed →
The account's balance over its first five days, read straight off the public ledger. One $607 deposit on March 14; $621 in deposits all week. By March 19 it was worth $24,000 — a 40× run, all of it winnings. This five-day streak is the foundation the entire rest of the record is stacked on top of.

A forty-fold run in five days is not a skill you can copy. It's a coin coming up heads a lot while you keep letting it ride — and that is precisely how the account was betting. Look at the size of those early bets against the money it actually had:

each of the first 20 bets, as a share of the whole bankroll
100%50%0% bet 11020 a sane max (~5%)
How much of the entire bankroll each of the first twenty bets put at risk. Careful bankroll management risks a percent or two per bet (the green line), so that a few wrong calls in a row can't wipe you out. This account's opening three bets each staked essentially the whole account on one game, and even after it built a cushion it routinely risked a quarter to a half of everything on a single result. While the coin kept landing heads, that looked like conviction.

That five-day run is the record. Remember the skill chart, and how March scored higher than any month that followed? This is that March. Everything the account did afterward — the six sports, the four patient months, the late-news discipline that so impressed me — was played on top of a bankroll a lucky opening week had already inflated forty-fold. And you can prove that one week is doing all the work in a single move: score the account's entire history on the luck-o-meter and it reads 1.4, barely above a coin flip. Delete just that opening streak and score everything after it, and it falls to 0.6 — pure noise. Take away the lucky week and there is no skill left in the rest of the record at all.

The streak wasn't the account warming up for a great season. The streak was the account. Everything after it was a coin-flipper with a big bankroll and a very good story — one that I, holding a spreadsheet full of green flags, very nearly believed.

The verdict

And then the forward test — the whole reason to keep a watchlist — rendered its opinion in real time. It peaked on July 4 and handed back $128,000 in the ten days after — a 38% collapse — winning 44% on prices that implied 50%: six points worse than the line, its skill score finally negative. Its first real losing stretch, and the account I'd graded my top pick got quietly re-graded, by its own results, from "confirmed sharp" to "cooling."

The single most genuine specimen on the whole board — a real, diligent, independent human — and it still wasn't worth copying.

That's the lesson this one taught, and why it earns its own file. Everything about it that looked like skill was something else wearing skill's clothes: the rising P&L was a pressed streak, the falling-but-still-positive edge was noise, the diversification was one expiring tournament, and the "compounding" was one lucky opening run it kept betting into. When the most convincing account on the largest prediction market on earth decomposes into a hot streak and a good World Cup — that's not a story about one wallet. It's the clearest evidence I found that there was no wallet to copy at all.

The plain-English version of the whole leaderboard audit is I Tried to Copy Every Winner on the Leaderboard; the mechanics of why records mislead are in How You Actually Win a Prediction Market. All wallet identities here are aliased. The daily P&L curve is the authoritative user-P&L feed; the by-sport and monthly-skill splits are from the trade-tape reconstruction, which runs a little under the authoritative totals. The opening deposit, the five-day balance ramp, and the bet-size figures are read directly from the account's public on-chain ledger.

SHELVEDcause: the record was the streak
"The most convincing wallet on the board was a lucky opening in a very good suit."