What is realized edge?

For each resolved position whose entry price can be recovered from the record, take the outcome (1 if the side the wallet held resolved true, 0 if not) and subtract that entry price. Average those differences. The result, in probability points, is the realized edge: how much more often the priced entries resolved true than the prices paid for them implied. The denominator is the priced count, not the resolved count. A resolved position whose entry price cannot be recovered is dropped from the average, and the n published beside the edge is the priced count that survived.

Is the denominator every resolved position, or only the priced ones?

Only the priced ones, and the gap is large. Measured on 2026-07-29 across the 146 wallet reads that had been reconciled against the chain, the edge sample equalled the priced count on 146 of 146 and equalled the full resolved count on 1 of 146. The mean priced share across those 146 was 63.2% and the lowest was 12.4%. On the 50 wallets published on the board, 34 were reconciled, their mean priced share was 46.9%, and 4 of the 50 cleared 90%. So on a typical board wallet the edge describes roughly half of the resolved record, and the other half carries no recoverable entry price in that read.

How can a 74% win rate come with a negative edge?

Because the entry prices implied more than 74%. A real example from our published board, read fresh on 2026-07-10: 35 wins and 12 losses on the 47 positions that read priced is a 74% win rate, but the realized edge across those same 47 is -2.5 probability points with a 95% interval of [-7.9, +1.9]. The wallet paid prices that implied roughly 77 wins per 100 and delivered about 74. Winning often is not the same as beating the price.

Why measure against the entry price instead of profit?

The entry price is the market's own stated probability at the moment of the trade, fixed before the world resolves the question. Measuring against it asks the only question that separates judgment from luck: did this wallet's chosen sides resolve true more often than the prices said they would? Dollar profit mixes that judgment with position sizing, fat-tailed outcomes, and one-off windfalls.

Entry price and realized edge

The answer first

Realized edge is the average of resolved outcome minus entry price, in probability points, across the positions a read could price. A prediction-market price is a probability: paying 74 cents for a Yes share is the market saying 74%. If your priced entries resolve true more often than the prices you paid implied, your edge is positive. If not, it is negative, no matter how often you won.

The denominator is the priced count, not the resolved count, and the two are not the same number. A resolved position whose entry price cannot be recovered from the record is dropped from the average rather than counted at zero, so the n printed beside every edge on this site is the count of positions that survived pricing. It is the honest denominator, and it is usually smaller than the wallet's full resolved record.

The intuition, with one worked example

Here is a real read from our published top-50 board, run fresh through the analyzer on 2026-07-10. The wallet, Zarvantis, shows 35 wins and 12 losses on the 47 positions that read priced. That is a 74% win rate, and it looks like a sharp record. The realized edge on those same 47 priced positions is -2.5 probability points, 95% CI [-7.9, +1.9]. A 74% win rate is not edge if the entry prices implied more than 74%. This wallet paid prices that implied roughly 77 wins per 100 positions and delivered about 74. The record wins often and still runs slightly behind the prices it paid, and the interval includes zero, so the verdict is not separable from luck.

The actual method

For each resolved binary position the wallet held, the measurement needs three things, and the third is the one that decides the denominator:

The entry price: the volume-weighted average price the wallet paid for the side it held. This is set by the order book strictly before the market resolves.
The outcome: 1 if that side resolved true, 0 if it did not.
Whether that entry price was recoverable at all. Reconstructing what a wallet paid means finding its buy fills for that exact outcome token in the record the read covers. When no candidate token under any collateral yields a buy price, the position is resolved but not priced, and it is excluded from the average rather than guessed at.

The per-position edge is outcome minus entry price. The per-wallet edge is the mean across the priced positions. The design is leakage-free by construction: the predictor (the price) is fixed before the world settles the question, and no profit-derived quantity appears anywhere in the measurement. A guard check confirms the entry-price level is not coupled to the edge statistic (Spearman 0.231 across the 3,871-wallet headline cohort, under the 0.6 screen). The edge is never delivered alone: the 95% interval (lesson 3), the priced count it was computed on, and the resolved count that priced count is a subset of all travel with it in the same record.

How much smaller the priced count is

Not a rounding difference. On 2026-07-29 we measured the 146 wallet reads that had been reconciled against the chain. The edge sample equalled the priced count on 146 of 146 of them, and equalled the full resolved count on 1 of 146. Across those same 146 reads the mean priced share of the resolved record was 63.2%, and the lowest was 12.4%.

On the 50 wallets currently published on the board it is lower still: 34 of the 50 were reconciled, their mean priced share was 46.9%, and 4 of the 50 cleared 90%. Board ranking selects for hyperactive wallets, and a hyperactive wallet accumulates more positions whose entry price this read cannot recover. So on a typical board wallet the edge you are reading describes roughly half of the resolved record, and the published n says which half by size.

That is why the priced count travels in the same sentence as the edge and its interval wherever a read carries the count, and why a read that cannot show what fraction of the record it priced gets no skill verdict: the four-state read returns insufficient instead of calling the edge either way.

That rule is narrower than it sounds, and the board is the case where it bites. None of the 50 published rows carries a priced-share figure at all, so none of them carries a verdict, and the Edge Score you see beside each one is a descriptive composite over the recent-trade window rather than a read of the whole resolved record. A priced share gates the verdict; it does not today gate the composite. The share is a property of what the read recovered, not a claim about what the wallet did: an excluded position is one this read found no entry price for, and nothing more than that.

Where you see this on the site

The free wallet analyzer leads with this number for any Polymarket address. The wallet board carries it per row, the embeddable skill badge quotes it with its interval, and the API returns it as realized_edge with the interval required by schema. The deeper glossary page is /learn/realized-edge.

What this does NOT mean

Realized edge is not profit. A wallet can have a positive edge and a small account, or a negative edge and one huge lucky win. It is not a forecast: the edge describes past priced positions and says nothing binding about the next one. It is not a statement about the whole resolved record either, because the average never saw the unpriced part of it. And a positive point estimate alone is not a skill verdict. The verdict comes from the interval, the sample size, the priced share, and the concentration checks in the next two lessons, which exist precisely so a flattering number cannot outrun its own evidence.

Next lesson: confidence intervals

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: