Two point nine million wallets walked into Polymarket expecting to out-think the crowd. Galaxy Research just ran the tape on every single one of them โ and the verdict landed like a cold tick.
69% of Polymarket's retail accounts are underwater. Combined retail losses: $338.9 million.
Now the buried part. The median losing account bled exactly $3. Three dollars. That's not a massacre โ that's a mosquito bite. But the mosquito bit 69% of the room, and a thin, vicious tail of accounts swallowed the entire bill.
The clock stops, but the chain doesn't.
I've been scraping Polymarket's on-chain settlement history since the 2024 election blowout. What Galaxy just formalized is the thing I felt in the data months earlier: this was never a democracy of information. It's a liquidity machine with a very narrow, very fast professional core. And this week, three separate jurisdictions decided that machine looks a lot like a casino.
Context: why this report lands like a subpoena
For the uninitiated โ Polymarket is the crypto-native prediction market where you buy YES or NO shares on real-world outcomes. Elections. Fed decisions. Sports results. Science milestones. Prices float between $0 and $1 and settle on whatever actually happens. The international book runs on-chain, with results resolved through an optimistic oracle (UMA is the industry-standard assumption, though the report never names it). A separate, independently structured US entity handles American users.
Galaxy pulled the entire on-chain history โ every position, entry price, holding period, and payout. First time anyone has run a full behavioral census at this scale. 2.9 million accounts trading on a human rhythm. Then, after filtering out 125,429 accounts firing more than 50 orders per active day, they published numbers the bull case never wanted printed.
I ran a version of this playbook during the ETF pre-approval leak in early 2024 โ cross-referencing Coinbase options volume against historical IPO patterns and publishing before institutional consensus formed. Galaxy just did the institutional-grade version. They let the chain speak before the press release.
Timing is everything. The report drops as Baltimore sued both Polymarket and Kalshi in August over unlicensed sports betting, New York hit Polymarket's US arm in September for unlicensed gambling and letting 18-to-20-year-olds trade, and South Korea indicted 26 users โ 18 referred to prosecutors, $12.7 million on the table. This isn't a price story. It's a narrative detonation.
Core: the 4.1% and the long tail
Here's the structural truth nobody's leading with: 4.1% of accounts generate 80.8% of all orders.
Read that again. Less than one in twenty wallets supplies four-fifths of the activity. That's not a community. That's a handful of automated and semi-automated operators renting the retail crowd as exit liquidity.
Zoom out and the plumbing is obvious. Those automation accounts are the market makers โ they quote, they arb, they harvest. Retail is the consumer, paying the spread on every conviction bet. The platform's liquidity doesn't come from a benevolent crowd. It comes from a rent-seeking engine, and 4.1% of wallets hold the keys.
I've watched this pattern before. During the Ethereum merge sprint in 2022, I scraped validator data and caught a 15% slashing-rate deviation hours before the outlets did โ the tell was never in the headline, it was in who was moving. Same instinct here. The distribution of activity is the story.
Now the methodology caveat Galaxy admits and everyone else skips: the 50-orders-per-day threshold is a heuristic, not a science. Set it higher and you miss the quant-leaning semi-pros farming edges just below the line. Set it lower and you flag a manic human who genuinely loves politics. The report publishes no sensitivity analysis. Treat the 125,429 automation figure as directional, not exact.
Then the wallet-versus-human problem โ the one the researchers themselves flag. A wallet is not a person. One operator spins up fifty addresses; one person runs a dozen positions across separate keys. That inflates the 2.9 million "retail" count and muddies the specialist split with duplicate identities. I've done this scraping work. Attribution is a nightmare.
But the behavioral signals cut through. About 44% of traders concentrate more than 60% of their activity on a single topic โ specialists, not tourists. And here's the counterintuitive gut-punch: specialists win 28% of the time. Generalists win 30.4%. The focused player is worse.
It sharpens by theme. Sports specialists โ the largest bloc โ post the lowest win rate. Tech and science specialists post the highest: 41.2%.
Then the churn. After a losing streak, 15.2% of users vanish within 30 days. After a win, only 6.1% do. That's survivorship selection running live โ the platform keeps its winners and bleeds its losers.
Most losers don't leave all at once, though. The median loss is $3, so the majority shrug and reload. It's the long tail โ the few accounts eating catastrophic drawdowns โ that supplies the $338.9 million headline. A median of $3 against a sum of $338.9M is a textbook extreme right-skew. The average tells you nothing. The tail tells you everything.
Liquidity flows where trust is liquid. And right now, the trust is thinning.
Contrarian: the double vulnerability nobody mapped
Here's what the coverage missed. Everyone reports "retail loses." Nobody connects the two maps.
Look at the overlap. Sports is the largest specialist cohort โ and the least profitable. Now look at the regulators. Baltimore, New York, Korea โ every action frames these contracts as gambling, specifically sports betting. The exact user base driving Polymarket's traffic is the exact user base losing the most money and the exact user base regulators are most eager to criminalize.
That's a double vulnerability stacked on one population. The traffic engine and the legal liability are the same people.
There's a second trap. The report notes winners "bet bigger and trade more frequently." Careful โ that's an outcome, not a cause. Winners have more capital and more conviction to size up; they didn't win because they sized up. Read it backward and you'll size into a losing book. I've watched retail make that exact mistake on my own platform's dashboards.
And the thing the report never touches: the settlement layer. Prediction markets live and die on their oracle. If the resolver ever disagrees with reality, every position above it is contested. No peer review. No contract audit. No consensus detail. Just the behavioral census. The infrastructure risk is the blind spot the whole analysis leaves open.
Whispers before the ticker opens. And the whispers say the "democratized information market" was always a professional's playground.
Takeaway
So where does the tick land next? Watch the New York and Baltimore rulings. If a court formally stamps these contracts as gambling, that's the precedent that re-prices the entire sector โ Kalshi's CFTC license didn't shield it, and that should terrify every compliant operator.
Watch Korea's user-side prosecutions, too. Going after traders instead of platforms is a demand-side chill that spreads faster than any civil suit.
Prediction markets aren't dead. The tech and science cohort at 41.2% proves a real information edge exists. But the story isn't "retail can win." The story is who wins, how fast they move, and how long before the regulators price the fun out of it.
Trust no one, verify everything, move fast. The chain already told us the ending.