CrowdIntelThe prediction market became a casino. The customers became machines.
Six months, $53B, 1.4 billion trades — 78% of them in markets that live less than a day, most of them placed by 15,752 machine accounts. The full macro picture, from every fill ever placed on Polymarket.
The record year nobody described right
The common story says Polymarket peaked with the 2024 US election. The data says the opposite: the first half of 2026 alone did $53.0B — 64% of all volume in the platform’s history, seven and a half times the entire election year. 82% of every fill ever placed happened in 2026, and 89% of every market that has ever traded was created in it — about one new market every eight seconds.
| Year | Volume | Fills | Wallets | Markets |
|---|---|---|---|---|
| 2024 | $8.82B | 58M | 645,427 | 19,497 |
| 2025 | $20.95B | 241M | 1,519,537 | 237,787 |
| 2026 (H1) | $53.03B | 1,402M | 1,886,578 | 1,918,043 |
The casino pivot
What changed wasn’t appetite for forecasting elections. It was the product. Markets that live less than 24 hours — five-minute Bitcoin candles, hourly up-or-down contracts — went from 0.3% of trades in 2024 to 42.6% in 2025 to 78.4% in 2026.
Over the platform’s whole life, 70% of resolved markets lasted under a day — and a quarter of every trade ever made happened in a market that lived under an hour. The “will-X-happen-by-December” market is the brand; the sub-hour candle bet is the business.
The machines run the floor
Who trades a market that lives 5 minutes? Mostly software. 15,752 machine-class accounts — 0.5% of all wallets — have placed 70% of all fills and 53% of lifetime volume. The one period machines lost share was the election quarter, when humans flooded in — then left.
The same signature shows in trade size: since the election quarter, volume grew 4× while trade count grew 23×. The average fill collapsed from $193 to $32. Growth that arrives as smaller, faster, more numerous trades is automation, not adoption.
It also resolves a paradox. Concentration looks like it collapsed — the top-100 accounts’ share of yearly volume fell from 73.3% in 2023 to 19.4% in 2026 as the user base exploded. But the liquidity didn’t go to the crowd. It went to machines. The oligarchy didn’t democratize — it automated.
The price of hope
Are Polymarket’s prices honest? Across every fill on every resolved market — the largest calibration measurement ever published for a prediction market — the answer is: impressively yes, except where hope is involved. A 50¢ contract wins 50.1% of the time. A 99.7¢ contract wins 99.8%. But sub-1¢ lottery tickets, priced at 0.29¢ on average, win 0.07% of the time — buyers of long shots overpay roughly 4×. Coin-flip underdogs at 35–45¢ run 2–4 points below their price; mild favorites at 55–66¢ quietly win more than they cost.
This is the classic favorite-longshot bias, measured on ~30 billion shares. The crowd buys hope; the machines sell it to them.
Who wins — and who stays
Per Polymarket’s own figures, 71.2% of accounts are net losers (95% CI 67.9–74.3). The median account is down −$2.62; the median winner is up just +$14.38. Meanwhile the top 50 winners have taken $291M.
And the crowd doesn’t stay to win it back. Of the 640,110 wallets that arrived for the 2024 election, roughly 90% are gone — the median 2025 joiner lasted six weeks from first trade to last.
The American clock
The “global wisdom of crowds” has a circadian fingerprint. Human fills climb through the US morning, plateau across US market hours, peak in the New York afternoon and bottom out while America sleeps — a 1.6× swing this week. The machines trace the same curve: liquidity lives where its human counterparties do.
The median winner redeems the same day a market resolves; 90% of redemptions land within 24 hours and only 1.7% wait a month. Professional plumbing — not tourists forgetting tickets.
How we computed this
Full on-chain reconstruction, reproducible queries, and what we do not claim.
The dataset is the whole exchange
Every figure derives from a full reconstruction of Polymarket’s on-chain order flow — 1.70 billion fills across 2.17 million markets, September 2020 through July 15, 2026. Nothing here is sampled or estimated unless labeled as such. Our monthly volume series reconciles to Polymarket’s own reported figures within 98%.
“Machine-class” is arithmetic, not accusation
An account is machine-class when its fill cadence and scale are unreachable by a human at a screen — thousands of fills per day, sustained for months. It is a floor, not a ceiling: attribution slightly undercounts the busiest market-makers. Market-making is a legitimate, necessary role; machine ≠ misconduct.
Profit figures are Polymarket’s own
Win/loss claims come from Polymarket’s public profit API (an 800-account random-sample survey, 95% CI reported) — never from our internal ledger. The calibration curve needs no profit accounting at all: it is pure counting — price paid versus how the market resolved, share-weighted across all resolved markets.
What we do not claim
No fraud, no manipulation, no identity claims — trade data cannot prove intent, and we tested for self-trading and found none. Addresses are pseudonymous. A concentration or calibration fact is a market-structure observation, not a verdict on any participant.