Why is win rate not the skill measure?

Because win rate mostly reflects which prices a wallet buys. Across the 3,871 wallets in the frozen discretionary cohort, mean entry price and win rate move together at Spearman rank correlation 0.951: wallets that buy favorites win more often, mechanically. The realized edge subtracts the entry price from each outcome, so it nets that effect out. Entry price and realized edge correlate at only 0.231 on the same cohort.

Can a wallet win 90% of the time and have no edge?

Yes, easily. Buy ten positions at 90 cents each and win nine: the win rate is 90%, and the realized edge is exactly zero, because the prices already implied nine wins in ten. The wallet did precisely as well as the market said it would, which is the definition of no edge.

Is a low win rate evidence of a bad trader?

No. A wallet that buys longshots at 10 cents can lose 85% of its positions and still hold a positive realized edge, because the prices implied it would lose 90%. Win rate without the prices paid is a style description, not a quality measure, in either direction.

Win rate vs edge

The answer first

A high win rate is mostly a description of taste, not talent: wallets that buy favorites win often because favorites win often. The realized edge subtracts each entry price from each outcome, so it removes that mechanical effect and asks the question that matters: did the wallet do better than the prices it paid implied?

The intuition, with one worked example

Buy ten positions at 90 cents each. Nine resolve true. Your win rate is 90%, and your realized edge is exactly zero: the prices already implied nine wins in ten, and that is what you delivered. Nothing was beaten. The same logic runs the other way for the wallet in lesson 2: a 74% win rate on 47 resolved positions with a realized edge of -2.5 points, 95% CI [-7.9, +1.9], because the entry prices implied more than 74 wins per 100. Win rate reads the scoreboard. Edge reads the scoreboard against the point spread.

The actual method

The coupling is measurable, and we publish it. In the frozen discretionary cohort behind the skilled-wallet registry (3,871 Polymarket wallets, each with at least 30 resolved positions):

Mean entry price and win rate move together at Spearman rank correlation 0.951. Buying favorites and winning often are nearly the same fact. This is the favorite-longshot structure of the market, not evidence of skill.
Mean entry price and realized edge correlate at only 0.231 on the same wallets. Once the price is netted out, almost none of the favorites effect survives, which is exactly why the edge, not the win rate, is the skill measure.

This near-0.95 coupling is also why a leaderboard sorted by win rate is a list of favorite-buyers, and why Convexly never ranks wallets that way.

Where you see this on the site

The analyzer shows both numbers for every wallet: the win rate as a descriptive stat and the realized edge with its interval as the measure that carries the verdict. The PnL vs skill study makes the same point one level up, ranking the same cohort by profit and by the skill composite side by side.

What this does NOT mean

It does not mean win rate is useless: it describes a wallet's style, and the calibration work in the next lesson uses win rates inside price bands to test whether entries match outcomes. It also does not mean favorites buyers are unskilled; some of the registry-cleared wallets buy favorites and still beat their prices. The claim is narrower: a win rate quoted without the prices paid carries almost no information about skill, in either direction.

Next lesson: calibration and the Brier score

Convexly publishes new methodology research roughly every 6-8 weeks plus the /learn series on a rolling cadence. Get the next paper in your inbox when it ships: