Why Polymarket wallets show 100% win rates (and what the data really says)

Look up the top wallets on Polymarket and you will find people who apparently do not lose. Wallets with a perfect record over hundreds of positions. Winning streaks in the thousands.

Our own pipeline reported exactly that, for months, and it was wrong. What makes it worth writing up is that we can show you both numbers: our wallet table is currently half-migrated, so the same 13,480 wallets are scored by two different methods at the same moment, and the difference is visible in a single query.

Same wallets. Same chain. Same scoring code. The only difference is whether losing positions were counted — and it moves the average win rate by eighteen points and the longest streak by two orders of magnitude.

A genuine 2,025-game winning streak on markets priced near even has a probability of roughly one in 10^609. There are about 10^80 atoms in the observable universe. Whatever produced that number, it was not a run of good bets.

Why the number cannot be true

Start with what a 70.7% average would mean. Prediction markets price outcomes as probabilities: a contract at $0.60 pays $1 if the event happens. If the average participant really won 71% of positions taken at prices near even, the market would be mispriced by roughly twenty percentage points, permanently, across thousands of independent participants. Someone would have arbitraged that away years ago. Markets this liquid are not that wrong.

So either sports prediction markets are the least efficient financial market ever documented, or the measurement is broken. It is the measurement.

The mechanism: losses leave no receipt

Polymarket settles positions through redemption. When a market resolves, holders of the winning outcome submit a transaction to redeem their tokens for USDC. That transaction is on-chain, timestamped, and easy to index.

Holders of the losing outcome do nothing. Their tokens are worth zero, redeeming them would return zero, and paying gas to collect zero is irrational. So they abandon the position, and no transaction is ever written.

The redemption record contains wins and only wins. Any win rate computed from it is 100% by construction — not because the wallet is good, but because losing positions are invisible to the method.

This is why the artifact is worst for the wallets that look best. A wallet with a handful of redemptions and no visible losses reads as flawless. The 2,025 streak is not a trader who won 2,025 times; it is a wallet whose 2,025 redemptions were counted and whose losses were never in the dataset to break the run. Rescored with losses included, the longest streak anywhere in the same population is 32.

What the data shows once losses are counted

The fix is to stop inferring outcomes from redemptions and start grading positions directly: take the position at entry, wait for the market to resolve, and record what happened — win or loss, redeemed or abandoned. Our pipeline does this for every signal it emits.

Between 2026-07-07 and 2026-08-16 that produced 103,421 graded, resolved positions:

Read the last two rows together, because that is where the finding is. Positions were entered at an average implied probability of 50.87% and won 50.71% of the time. The market's price was accurate to within a sixth of a percentage point across a hundred thousand positions.

Two independent routes agree. Rescoring the wallets with losses counted gives 52.1%; grading a hundred thousand positions against resolution gives 50.71% against a 50.87% market price. The redemption record's 70.7% is the outlier, and it is the one every leaderboard is built on.

What this means if you follow other people's positions

Copy-trading products built on Polymarket almost all rank wallets by realised profit or by win rate derived from redemptions. Both inherit this bias, and both inherit it unevenly: the ranking is not merely noisy, it is systematically tilted toward wallets whose losses happen to be least visible.

Three checks worth running on any wallet ranking

1. 1Ask how a loss is detected. If the answer is redemptions, settlements, or 'on-chain PnL', losses are missing and the win rate is inflated. There is no way to patch this after the fact — the data was never collected.
2. 2Look at the extremes. A leaderboard containing 100% win rates over meaningful sample sizes is telling you about its own method, not about its traders. Real edges in liquid markets are small: two or three percentage points is a strong, durable result.
3. 3Compare against the entry price, not against 50%. A 70% win rate on positions entered at $0.72 is slightly below the market. Win rate alone is not a measure of skill; win rate minus implied probability is.

How to reproduce this

Nothing here requires private data. Take any set of wallets, record the positions they open from the public trade feed, wait for the markets to resolve, and grade each position against the resolution — independently of whether a redemption was ever submitted. Then compare that win rate to the average price at entry.

If your numbers come out near the implied probability, the market was efficient and your method is sound. If they come out far above it, check whether your losses are being counted before you conclude you have found an edge.

What we do with this

We grade every signal we publish this way, including the ones that lose, and the record stays up. That is a deliberately uncomfortable choice: measured honestly, no signal source we track — including ours — shows a large edge over the market price, because a market this liquid does not leave one lying around.

What the data does show is where the money and the crowd disagree, when a price moves hard before kickoff, and which wallets are actually behind it. That is information about what is happening in a market, which is a different and more defensible claim than a prediction about what will happen.