Research note · July 9, 2026

Profit Is Not Proof

We ran the luck filter on our own Polymarket leaderboard during the World Cup.

The dollars are real. The skill is the open question.

June was the biggest month prediction markets have had: $44.8 billion in combined trading volume across Kalshi and Polymarket, up about 75% from May's $25.66 billion (The Block, July 1, 2026). Kalshi's FIFA World Cup Winner market alone has taken more than $832 million in bets. A single round-of-16 game, USA vs Belgium on July 6, reportedly drew $122 million on Polymarket and $64 million on Kalshi (Yahoo Finance / CryptoProwl, July 6, 2026). All of these are third-party reported figures, cited to their sources, not Convexly computations.

Bloomberg spent July 6 on the obvious question: who keeps the money (Bloomberg Odd Lots, July 6, 2026). Their answer, sourced from traders on the inside, is that most people lose and a small dedicated group grinding politics and economics wins. That is a profit story. It counts dollars in and dollars out.

Profit is a scoreboard, not a diagnosis. A wallet can be green because the trader is sharp, or because it loaded up on favorites during a lucky stretch, and during a tournament the two look identical: variance is enormous and thousands of new accounts are placing a handful of bets each. So we asked the question the dollar leaderboards skip. Subtract the luck. Who actually beat the prices they paid?

Method, in one paragraph

For every resolved binary position a wallet held, we take two numbers that cannot contaminate each other. The first is the entry price: the volume-weighted price the wallet paid, fixed at trade time before the market resolved, expressed as a probability between 0 and 1. The second is the result: 1 if the outcome came true, 0 if it did not. Realized edge on a position is the result minus the entry price, and a wallet's skill measure is the average of that edge across its positions. Under the assumption that every entry price was fair, each result behaves like a coin flip weighted to its own entry price, which gives an exact per-wallet null distribution and a p-value. Because thousands of wallets are tested at once, we then apply Benjamini-Hochberg false-discovery control at q = 0.10, which caps the share of merely lucky wallets that can slip through as false positives. This is the method Barras, Scaillet and Wermers (2010, Journal of Finance) used to separate skill from luck in mutual funds, pointed at on-chain wallets. It separates skill from luck in-sample; it does not establish forward skill.

Finding 1: platform-wide, real edge is rare but it exists

Of 3,871 Polymarket wallets with at least 30 resolved discretionary positions (378,424 resolved positions after excluding 118,195 sub-hourly crypto up/down micro-market positions held by 2,546 wallets), 178 clear Benjamini-Hochberg false-discovery control at q = 0.10 for positive realized edge over their entry prices. Expected false discoveries inside that cleared set: at most about 17.8 (q times 178). The cleared set is 4.6% of the tested universe (Wilson 95% CI [3.98%, 5.30%]). At the looser raw p < 0.05 screen, 427 wallets clear against about 193.55 expected under the global null, roughly 2.2x chance.

About one wallet in twenty-two, on fully settled history. That is an in-sample ceiling. The prospective out-of-sample holdout that would test whether these wallets keep their edge is filed and running, public receipt pending verification.

Finding 2: our own leaderboard fails the test

We did not exempt our own product. Convexly publishes a top-50 leaderboard cohort ranked by our Edge Score composite, so we pointed the same test at it. Of the 50, 35 had at least 30 resolved positions, enough to read. Zero cleared the corrected bar. One uncorrected hit appeared, which is what chance predicts across 35 tests (about 0.9 expected false positives at a one-sided 2.5% screen); it was not FDR-cleared, it was net-negative on dollars, and it sat at #48 of 50 (Convexly, convexly.app/research/top50-skill-scan; data as of June 9, 2026). The highest-Edge-Score wallet with enough history to read carried a realized-edge CI of [-7.8pp, +11.9pp]. That interval spans zero.

Finding 3: calibration barely helps either

Across 8,656 Polymarket wallets and 582,921 resolved positions, how well-calibrated a wallet is correlates only weakly with how much it profits: Spearman r = +0.148 (95% CI [+0.128, +0.168]), roughly 2% of shared rank variance (Convexly, convexly.app/research/polymarket-10k-wallet-study). Being right about probabilities and making money barely travel together in this data.

Why this matters during the World Cup

More dollars do not make more skill. They make more variance, and variance is when a green balance says the least about the trader behind it. Before you copy a wallet because it is up this month, look at the other number: its edge over the prices it actually paid, with an interval around it. Convexly computes that number free for any Polymarket wallet. Paste a Polymarket wallet address into the analyzer and you get the Brier score, calibration, concentration, realized entry edge with a 95% interval, and Edge Score, interval included.

Skill is findable in this record. A profit leaderboard is the wrong instrument for finding it.

What this is, and is not

This is an in-sample skill-vs-luck separation on public on-chain Polygon data. It is not validated forward skill, not investment advice, and not a claim that the cleared wallets will keep their edge. The forward, out-of-sample holdout is filed and running (public receipt pending verification); a null result there would be a valid, publishable outcome. Every figure above carries its interval or its test, and sub-sample sizes are stated inline: 35 of 50 readable, 178 of 3,871 tested, r computed on 8,656 wallets and 582,921 positions.