We Went Looking for Sharp Money on Polymarket. Here's What $80M of On-Chain Tape Told Us.
Thursday, June 11, 2026
Our system reads one signal: the gap between how many bets land on a side and how much money lands on it. That data comes from sportsbook aggregators, behind a paywall. Polymarket is the opposite — an exchange where every fill is public, permanent, and attributable to a wallet. This is the full record of the day we spent finding out whether the chain could replace the paywall: our thesis, our methods, and two beautiful ideas executed by the data.
Experiment 1 — Can the crowd's tape replicate the sharp-money split?
The construct maps cleanly: a sportsbook's bet% is the share of tickets on a side; on an exchange that's the share of trade count. Money% is the share of handle; on an exchange, the share of notional. We built it twice: a live logger snapshotting every MLB market's tape each half hour, and — because the chain never forgets — a historical backfill through Polymarket's on-chain orderbook index, reconstructing the pre-first-pitch tape for 418 games (Mar 27 – Apr 27) and cutting each game's tape at any timestamp we wanted, retroactively.
Joined to the sportsbook splits we'd logged on the same 413 games:
| Measure | Correlation (Pearson) | Verdict |
|---|---|---|
| Tickets% vs sportsbook bet% | +0.51 | crowds lean alike |
| Notional% vs sportsbook money% | +0.32 | weakly related |
| The sharp signal (money − tickets) | −0.01 | zero. unrelated. |
The divergence construct — the thing that actually carries the signal — does not transfer. And the exchange version has no predictive value of its own: games where the exchange's "sharp side" hit our trigger threshold covered 57.0% vs a 55.6% base rate. Noise. Why? An exchange price absorbs big money instantly; there is no stale sportsbook number being pushed against. The phenomenon our signal measures structurally cannot exist there.
Experiment 2 — Forget the crowd. Follow the winners.
But the chain offers something no sportsbook ever will: names. Every fill belongs to a wallet, and every wallet's P&L is computable. So we computed all of it — netted every wallet's pre-pitch moneyline position across those 418 games and settled them against final scores. 13,202 wallets. After filtering to serious, directional bettors (≥8 games, ≥$500 staked, not market-making both sides): 1,219 accounts.
The leaderboard was breathtaking. The top 18 made a combined +$3.88M in one month. The styles looked like skill: "dog snipers" (70% underdogs at depressed prices, +33% ROI), "favorite grinders" (steady +12–18% on chalk), and one account that won 75% of its games while paying an average entry implying 48%.
A month ago we might have published that list. Instead we did the thing this site exists to do.
Experiment 3 — The holdout. Three times.
The leaderboard was selected on March 27 – April 27 data only. So we pulled the same 18 wallets' trades for May 1 – June 10 — six weeks they couldn't have been selected on — and settled those too:
| Test | Result |
|---|---|
| Top-18 cohort, holdout window | −$662,470. Six wallets went silent entirely (incl. the #1 and #2 earners). Only 2 of 18 stayed profitable — fewer than coin-flipping predicts. |
| The most "skilled-looking" account (92 games, 62% wins, +$343k) | bet $1.2M in the holdout and lost $338k |
| April, split in half: do wallets hot in H1 stay hot in H2? (383 accounts) | ROI rank-correlation +0.006 |
| May–June, split in half (190 accounts): top-20 of the first half (+$1.06M)… | …made +$10k on $8.5M wagered (+0.1%) in the second half. Rank-corr +0.11. |
| Fade-the-losers (bottom-20) | also dead — losers regress to the house take, same as winners |
Three separate eras. Zero persistence, every time. The May–June "who was hot" leaderboard we built next looked exactly as impressive as April's — +$486k on nine games at the top — and the split test says it is exactly as meaningless. Pre-game MLB leaderboards on this venue are hindsight, structurally. Roughly 1,300 near-coin-flippers, a fat lucky tail, an unlucky one, and a house take in the middle. (It didn't help that the venue's pre-game MLB volume collapsed ~10× between April and June — whatever real skill was ever in the pool largely left the table.)
What survived
Two things. First, the boring one: real prices. The same infrastructure that ran these experiments now stamps a true, executable exchange price on every play we log — which keeps our own ROI honest in a way screenshot "best odds" never could. Second, the lesson, which we'd rather learn on $80M of other people's fills than on our bankroll: every in-sample result on this data died out-of-sample, the same day, twice. The sharp-money construct we actually bet — sportsbook splits, locked rules, forward-only record — remains the one signal that has survived its own holdout. That's not a victory lap; it's a bar. Everything we test next has to clear it before it touches a dollar.
Methods, briefly: live tape via Polymarket's public data API (30-min snapshots); historical tape and per-wallet ledgers via the public Goldsky orderbook subgraph (indexed through Apr 28), order-fill events netted maker-centric per wallet per game, settled at $1/winner against final scores; holdout trades via per-user activity queries and full per-market tape where reachable. Wallet addresses are public on-chain; we truncate them here out of courtesy. Nothing in this article is betting advice — it's the opposite: a record of ideas we tested and declined to bet.
← All research