Skilled or Just Lucky? Polymarket's Top Winners, Tested

How many Polymarket winners survive a luck test?

Far fewer than the leaderboards suggest. The whole skill vs luck Polymarket question turns on one test: would a winner's edge survive a re-run? As reported by CoinDesk in April 2026, the LBS/Yale study found only about 3.14% of prediction-market traders are genuinely skilled, with most lucky winners regressing — methodologies differ, but the method is reproducible per wallet via calibration, alpha-vs-mid, and resolved-market sample size.

Every few months a study lands that says the quiet part out loud: most of the people winning at this are lucky, not good. The latest is a London Business School / Yale paper, reported by CoinDesk in April 2026, that ran prediction-market traders through a luck test and found that only about 3.14% were genuinely skilled — and that roughly 60% of the lucky winners regressed when their bets were effectively re-run.

That is a brutal number, and it is easy to misuse in both directions. Doomers read it as "prediction markets are a casino." Promoters ignore it and keep screenshotting whoever is up the most this month. Both are wrong in the same way: they treat a leaderboard as a skill ranking, when a leaderboard is mostly a luck ranking with skill mixed in somewhere.

This post is about the somewhere. The coin-flip study describes a method for separating the 3% from the 97% — and that method maps almost perfectly onto data we already reconstruct for every Polymarket wallet. So we will explain the test in plain English, show why a profit leaderboard can't run it, and then walk through exactly how to run a version of it yourself, on any wallet, in a few minutes.

We can't restate those exact third-party figures as our own measurement — methodologies differ — but the test behind them is reproducible per wallet using calibration, alpha-vs-mid, and a sample size counted in resolved markets, not trades.

The LBS/Yale coin-flip method, in plain English

The study's logic is older than crypto, and it is the same logic that exposes a lucky fund manager. Imagine a room of 10,000 people each flipping a coin ten times. By pure chance, around ten of them will flip ten heads in a row. If you only show up after the flips and rank everyone by heads, those ten look superhuman. They are not. They are the tail of a distribution.

The fix is not to admire the streak. It is to ask a different question: if these people flipped again, would the same ones win? A skilled forecaster's edge shows up again. A lucky one's "edge" evaporates, because there was never a mechanism producing it — just variance that happened to land their way.

The coin-flip method does three things, roughly:

1. Separate the signal from the streak. It models how much of a trader's result could have come from luck alone, given how many bets they placed and at what odds.
2. Re-run the world. It asks whether the winners' edge survives when you account for that luck — a bootstrap or out-of-sample check that re-tests the same skill on data the streak can't have rigged.
3. Count the survivors. Whoever still beats chance after that is the skilled slice. In the study, as reported, that slice was about 3%.

The honest caveat: that is their model on their data, and the assumptions matter. A different luck model, a different market, a different time window gives a different headline number. So we are not adopting "3.14%" as a Polymarket fact. We are adopting the question — does this edge survive a re-run? — and operationalizing it on-chain.

Why a P&L leaderboard can't tell skill from luck

Here is the core problem with ranking by profit. On a prediction market, a single large win can come from being right or from being early-and-lucky on a binary event, and the P&L column looks identical either way. Profit is the output of skill plus luck plus stake size. It does not decompose itself for you.

Consider two wallets that both finished a quarter up +$40,000.

Ranked by profit, they tie. Ranked by whether the result would repeat, they are not remotely the same wallet. The Survivor has an edge a re-run would reproduce. The Regressor has one heads-streak that a re-run would, on the study's logic, wash out about 60% of the time.

This is the same trap as win rate, which we covered in Why win rate is the most misleading number on Polymarket: a number that feels like skill but is mostly a record of price preference and variance. A profit leaderboard is one rung better than a win-rate leaderboard, and still nowhere near a skill test.

What "skilled" means here: calibration and alpha-vs-mid over resolved markets

To run a coin-flip test on a wallet, you need metrics that behave like skill — stable for the same wallet over time, hard to manufacture with stake size or price preference, and gradeable against a baseline the trader doesn't control. Two do most of the work.

Calibration (Brier score), benchmarked against the market. Every entry price is an implied probability: buy YES at 64¢ and you are asserting the event is more than 64% likely. Calibration asks whether your 64% claims actually happen ~64% of the time, across every resolved market. The Brier score squares the gap between what you implied and what occurred, then averages it — lower is better. The reason it resists luck is that it grades you relative to the difficulty of the claim. Being right about a 96¢ favorite proves almost nothing; the market already said 96%. Being repeatedly right about 60¢ propositions that resolve closer to 75% is judgment, and it is the kind of thing that recurs. On Polyrank, a wallet's calibration is benchmarked against the market's own implied probabilities on the same markets — beating the crowd's calibration is the bar, not beating a coin flip.

Alpha-vs-mid. Calibration grades judgment at resolution; alpha-vs-mid grades execution now, fill by fill. For every fill we reconstruct from the raw Polygon transaction, we ask: did this trader get a better price than the market's midpoint at that moment, and did the price subsequently move their way? One fill beating the mid is luck. Hundreds of fills systematically beating the mid is an edge you can watch forming before any market resolves — which is exactly the kind of signal a coin-flip test is looking for, because variance doesn't reliably beat the mid hundreds of times in a row.

Sample size, measured in resolved markets. This is the part everyone skips. A coin-flip test is only meaningful if there were enough independent flips. Twenty resolved markets is an anecdote; a thousand fills concentrated in two markets is one opinion, repeated. Counting resolved markets — not trades, not fills, not volume — is how you avoid crowning someone whose "track record" is three correlated bets on the same event.

The fingerprint of a survivor vs a regressor

You can usually tell which side of the line a wallet sits on by its shape, before any formal re-run. The profiles below are illustrative archetypes — built from patterns we see repeatedly across the dataset, not specific real accounts, and the numbers are illustrative, not measured.

The Survivor. ~300 resolved markets across several categories. Brier score that beats the market baseline on the same markets. Alpha-vs-mid positive and stable quarter over quarter — the edge shows up in Q1, then again in Q2, on entirely new markets. Drawdowns exist but never threaten the account. Its P&L is built from many medium edges, not one miracle. Re-run its world and the same edge reappears, because there is a mechanism producing it.

The Regressor. A huge number on the P&L column and a tiny number in the resolved-markets column. The profit traces to one or two binary events that happened to break right. Calibration is roughly indistinguishable from the market's implied prices — meaning it forecasted no better than the odds already did. Alpha-vs-mid is noisy and near zero. This is the wallet a profit leaderboard loves and a coin-flip test deletes.

The tell is persistence under partition. Split a wallet's history in half by time and score each half independently. A survivor's calibration and alpha-vs-mid look similar in both halves. A regressor's "edge" lives almost entirely in one slice — the slice with the lucky event — and collapses in the other. That split-half check is a poor-man's version of the study's re-run, and you can do it by eye on a wallet's history.

Why luck regresses and skill persists

The deep reason the coin-flip test works is regression to the mean. A result with no underlying mechanism is, by definition, not reproducible — the next sample pulls back toward the population average. A result with a mechanism (you genuinely price events better than the market) reproduces, because the mechanism comes along to the next market.

Sample size is the lever that converts this from philosophy into a usable filter. The more independent resolved markets a wallet has, the harder it is for luck alone to fake a strong calibration score across all of them. Ten markets can be ten heads-streaks. Three hundred markets with a Brier that beats the baseline is not a streak you can flip your way into — the probability of doing that by chance falls off a cliff as the count rises.

This is why we treat resolved-market count as a gate, not a vanity stat. Below a threshold, a wallet's metrics are flagged as low-sample and we decline to crown it, no matter how green the P&L looks. Concretely, the smart-money screen won't tag a wallet "sharp" until it has at least 100 resolved markets, and the copyability verdict needs at least 30 before it will render an assessment at all. It is also why a wallet up +$40k across three markets is, statistically, closer to the coin-flippers than to the forecasters — and why the same +$40k across hundreds of markets is the opposite.

The third-party base rates are consistent with this picture, with the usual caution that the methodologies differ. As reported by Yahoo Finance, roughly 70% of Polymarket traders lose money and the top ~0.04% of wallets captured most of the profits; a separate ~12.7% profitable figure also circulates in press coverage. None of those are our measurements, and none of them, on their own, separate skill from luck — but a world where a tiny tail holds the profit is exactly the world the coin-flip study is describing, where most of that tail is variance and a sliver of it is real.

How to run the same test on any wallet yourself

You do not need the study's code to apply its logic. Here is the checklist, in order, runnable on any wallet's free Polyrank profile — no login, no wallet connection.

1. Read the resolved-market count first. Not trades, not volume. If it is small, stop treating the wallet as evidence. Small sample = unrun test.
2. Check calibration against the market baseline. Did this wallet forecast better than the prices it traded against? If its calibration just tracks the odds, it has no demonstrated edge to re-run.
3. Check alpha-vs-mid. Are the fills systematically beating the midpoint, across many entries — or is the P&L one timing windfall wearing a track record?
4. Split the history in half by time. Do calibration and alpha-vs-mid hold up in both halves? Persistence across the split is your re-run. Edge in only one half is a heads-streak.
5. Put drawdown next to P&L. A result achieved while routinely risking the account is a different result — and a fragile one a re-run can break.

Each of these is one of the 77 weighted skill metrics on a wallet's profile, drawn from families like profitability, calibration, risk, alpha, conviction, execution, and category specialization. The point of the screen is not to find the biggest number; it is to find the numbers that would survive being re-run.

To make this concrete, the table below is the live result of a skill screen — wallets that clear the smart-money preset right now. It is not a profit ranking; net-losers are filtered out, and the ordering is driven by the metrics that resist luck (calibration, alpha-vs-mid, risk-adjusted return over a real resolved-market base), not by who happened to win biggest last week.

#	Trader	Score	P&L	Win%	Volume
1	ga…22	100.0	+$851.5K	51%	$70.87M
2	03…33	100.0	+$272.1K	98%	$150.57M
3	va…gh	100.0	+$409.7K	50%	$17.96M
4	0x…5b	100.0	+$330.4K	50%	$21.56M
5	Ag…ng	100.0	+$425.5K	50%	$15.33M
6	Lu…ow	100.0	+$305.9K	95%	$13.25M
7	po…er	100.0	+$206.3K	50%	$9.66M
8	xu…08	100.0	+$173.1K	50%	$8.66M
9	la…ow	100.0	+$201.3K	52%	$4.43M
10	UU…LR	100.0	+$110.7K	50%	$4.69M

That is the difference between a leaderboard and a screen. A leaderboard answers "who is up the most?" A screen answers "whose edge would survive a re-run?" — which is the only question the coin-flip study says is worth asking.

You can also build your own version of this test. The rankings builder lets you weight the 77 metrics yourself — crank up calibration and sample-size requirements, downweight raw P&L, and watch the leaderboard reshuffle from "luckiest" toward "most likely to repeat." For a weekly read on which screened wallets are actually moving, the smart-money report does the running for you.

FAQ

Are Polymarket's top traders skilled or just lucky?+

Some of each, and the leaderboard can't tell you which. As reported by CoinDesk, the LBS/Yale study estimated only about 3% of prediction-market traders are genuinely skilled, with most winners regressing on a re-run; methodologies differ, so treat the exact figure with caution. The practical answer is to test each wallet individually — calibration vs the market, alpha-vs-mid, and a large resolved-market sample — rather than trusting its profit rank.

What does "regression to the mean" mean for traders?+

A result that has no underlying mechanism tends to pull back toward average over time, because there was nothing producing it but variance. A lucky winner regresses; a skilled one does not, because the skill comes along to the next market. The coin-flip test is essentially a check for which traders' edges don't regress.

Why isn't P&L enough to identify skill?+

Profit is skill plus luck plus stake size, fused into one number. A wallet up $40k across three markets and a wallet up $40k across three hundred markets look identical on the P&L column but are statistically very different — one is mostly a heads-streak, the other is mostly a mechanism. You need metrics that decompose the result, which P&L alone never does.

How many resolved markets are "enough" to trust a wallet?+

There is no single magic number, but the logic is clear: the more independent resolved markets, the harder luck can fake a strong calibration score across all of them. A handful of markets can be a streak; hundreds with a market-beating Brier score cannot. Polyrank treats resolved-market count as a gate and flags low-sample wallets rather than ranking them as proven — the smart-money screen requires at least 100 resolved markets.

Does Polyrank claim it found the "3% skilled" figure?+

No. That figure is from the LBS/Yale study as reported by CoinDesk, on their data and their model — we cite it, we don't reproduce it. What Polyrank does is operationalize the method behind it: per-wallet calibration, alpha-vs-mid, and sample size, all reconstructed from public Polygon transactions, so you can apply the same skill-vs-luck question to any individual wallet.

Methodology & disclosures

Polyrank reconstructs every metric from raw Polygon blockchain data — roughly 2.9 million wallets and ~35 million deduplicated trades spanning 3+ years of Polymarket history, rebuilt from fills, splits, merges, redemptions, and resolutions. Every number on a wallet's profile traces back to a public Polygon transaction; nothing is self-reported. The "Survivor" and "Regressor" profiles in this post are labeled illustrative archetypes, not specific accounts. The 3.14% / 60% figures are from the LBS/Yale study as reported by CoinDesk (April 2026); the ~70% / ~0.04% / ~12.7% figures are reported by Yahoo Finance and other press. All third-party statistics are cited as reported, and their methodologies differ from ours — we do not restate them as Polyrank measurements.

Polyrank is an independent analytics product and is not affiliated with Polymarket. Everything here is read-only, non-custodial analysis of public blockchain data. None of it is investment advice, and no metric — including the ones that survive a luck test — promises anything about future results. We measure skill; we don't sell certainty.

---

Run the test yourself. Paste any address into the free wallet lookup — no login, no wallet connection — and check its calibration, alpha-vs-mid, resolved-market sample, and all 77 skill metrics in seconds. Then build your own skill screen and see who survives it.

Founding Trader: the first 250 seats lock in 40% off Pro, for life.