How PolyTrace measures whale evidence

The threshold applies to one trade, not to a position built up over several. It is a fixed dollar amount rather than a percentile of current activity, which means what qualifies today qualifies next quarter — the bar does not move when the market gets busier, and a quiet week does not promote small trades into the tape.

Trades are read from Polymarket’s live feed and verified on-chain. The data layer captures every category, including sports and esports. Individual views and alert channels may apply their own category filters, so a filtered feed is not a claim that those trades were absent from the underlying dataset.

PolyTrace organizes the public Polymarket fills of $1K+that its live collector received and retained. This tracked subset is not Polymarket’s complete trade history or a wallet’s complete account activity.

Direction is normalized to the economic Outcome, not copied from BUY or SELL. Buying No and selling Yes can express the same view. Eligible fills are grouped in a bounded window, and a signal event appears only when a published rule passes: independent-wallet alignment, or unusual covered volume. Trading flow on its own does not establish that a wallet opened, added to, closed, or reversed a position.

Each signal retains its ranking reason, independent-wallet count, concentration, opposing evidence, baseline, coverage, freshness, and links to the recorded fills. Market-making, arbitrage, high-frequency, unknown-direction, or otherwise ineligible activity can remain in raw evidence without counting toward ordinary consensus. A signal is an investigation queue, not a recommendation, forecast, or statement that the same execution is still available.

• Partial publishes only observed facts while required source coverage is missing.
• Stale retains an aged last-good snapshot and labels it as delayed, not current.
• Unavailable means the source or publishing gate cannot support a value; it is not zero.
• Unmapped keeps a recorded fact outside Category or entity conclusions when the required link is absent.

Current positions, Category history, prices, and signal baselines may therefore have different states on the same page. The page does not use global wallet metrics or a different source to make an incomplete dataset look complete.

BUY/SELL, trade size and execution time are recorded trade facts. Open, Add and Reduce are inferred holding actions, not an exact reconstruction of every wallet action. The explanation icon beside each label shows the official Polymarket snapshot basis, token holding evidence, collection time and age at the trade. If evidence cannot support an action, only the icon is shown with the reason.

Frozen snapshot-only records use token presence or quantity in the official snapshot. When a record identifies the flow-adjusted method, its explanation distinguishes that official baseline from net recorded large-trade changes for the same wallet and token after snapshot collection completed and strictly before the current trade. It reports the frozen estimated pre-trade quantity; the current trade and trades at the same timestamp are excluded. An Open label can therefore mean a positive baseline was offset by negative net flow, not that the original snapshot had no holding.

Snapshot age is measured from collection start; the explanation shows that timestamp or the collection interval. Missing, late or stale evidence does not become a zero balance. A later snapshot with no holding is only a later observation: it does not prove that a specific SELL closed the position.

A position is decided when it resolved one way or the other inside the window. Positions still open in unsettled markets are undecided and excluded. A position that came out exactly even is dropped from the sample rather than counted as a loss — the rate is meant to answer “of the positions that went one way or the other, how many went the trader’s way”.

A losing position counts whether or not the trader redeemed it. What decides a position here is the market settling, not the wallet claiming anything afterwards. That distinction is the whole ballgame: a losing share is worth $0, so there is nothing to collect and no reason to send the transaction, and a rate built from redemption events therefore sees every win and almost none of the losses. Records like “217 wins, 0 losses” for a wallet that in fact lost 218 times are what that mistake produces. PolyTrace reads settled positions, so the unredeemed loser is in the denominator.

The window is a fixed 90 days rather than a fixed number of trades. A trade count sounds equivalent but is not: it covers all of history for a wallet with 300 trades and a few months for one with 50,000, so putting the two on one leaderboard compares nothing. That version of the code scored the same wallet 1.91% one way and 13.42% another, depending on which path computed it.

The window is 90 days because that is how long a settled result is kept. Category results are retained for 90 days, and the win rate is computed from those same retained positions, so the measurement window is the retention window rather than a separate choice. Whether to restore a longer window by retaining more is an open decision, not a settled one; until it is made, this page states the window the numbers actually come from. It said 180 here until September 2026, which was the retention period before the two were reconciled and not the period any published rate ever covered.

Not every wallet’s rate covers 90 days. A wallet whose results have never been published through the category pipeline still carries the rate an earlier measurement left behind, over a trailing 180 days, and nothing has rewritten those numbers since August 2026. Rather than relabel them, PolyTrace publishes the window each rate was actually measured over and the trader page prints it beside the rate — so a “180d” on a profile is that older measurement, not a second methodology.

A position that settled before the window opened is not in the sample. This is the limit worth carrying away, because it shows up on screen as an apparent contradiction: a wallet can read 94.9% over its recent settled positions while the same page lists a −$61,848 position that settled four months ago, or one still open and far underwater. Both are accurate and neither corrects the other. A win rate is a statement about one bounded sample, never about a wallet’s whole book, which is why the open-exposure line below sits beside it.

Below 20 decided positions there is no win rate. It is withheld, not shown as zero and not shown as a raw fraction. One wallet with 3,606 positions had 12 that could be decided and reported 91.67% — on screen that was indistinguishable from a rate backed by hundreds. A wallet whose history was too long to fetch across the full window is withheld for the same reason: a partial fetch biases the sample toward recent activity.

The count is the wallet’s open positions in the latest position snapshot. The share beside it is the absolute value of the unrealised losses, divided by the absolute value of every unrealised gain and loss. A wallet whose open book is +$40,000 on some positions and −$10,000 on others is 20% underwater by this measure. The denominator is the total unrealised movement, not the capital at stake, so the figure answers “which way is the open book leaning” rather than “how much is at risk”. When nothing open has moved in either direction the denominator is zero, and the share is withheld instead of printed as 0%.

It is read from one snapshot, and that snapshot’s time is printed with it. A number with no time attached would be the more useful-looking version and the less true one: these positions are marked at whatever the book showed when the snapshot was collected, and prices have moved since.

This is a description, not a prediction. An underwater position can settle a winner and a position in profit can settle at zero — the share says where the open book sits right now, and nothing about where it lands. It is shown beside the settled sample because the two questions are routinely confused: a rate backed by 40 settled positions is silent about the 400 still open, and a profile that shows only the settled half invites the reader to assume the rest resembles it.

Win rate ignores what a position cost. Edge does not. Being right 90% of the time on 80¢ favourites is close to break-even; being right 70% of the time at 35¢ is a large, real gap. Edge is the metric that separates those two, and it is stated here in full so any wallet’s number can be recomputed by hand.

Step 1 — score each settled fill. Take the fill’s entry price p as a probability between 0 and 1 (a 35¢ fill is p = 0.35). Then:

• BUY that won: (1 − p) / p — paid p per share, received $1.
• BUY that lost: −1 — the whole stake.
• SELL that won: −1 — the side sold went on to pay, so the complement held was worthless.
• SELL that lost: p / (1 − p) — the complement won, and it cost 1 − p.

“Won” and “lost” here mean the outcome the fill was on, after the market settled. A BUY at 35¢ that won scores (1 − 0.35) / 0.35 = +1.857; the same fill losing scores −1.

Step 2 — average, with the thin samples pulled toward zero. Sum every fill’s score and divide by the number of fills plus 15:

edge = Σ edge_i / (n + 15)

The 15 is 15 phantom fills that scored exactly zero. It is deliberately conservative: an unknown wallet is assumed to have no edge, so a 21-fill hot streak is discounted toward zero while a 400-fill record is barely touched. Without it the top of any edge ranking is entirely small samples. The published number is this shrunk figure — not the raw average — which is why a wallet’s hand-computed raw mean will always be larger in absolute terms than what the page shows.

Step 3 — the interval. The standard error published beside the number is the sampling error of that same shrunk average. The 95% interval is the number ± 1.96 standard errors. A band that spans zero means the sample cannot distinguish this wallet from a wallet with no edge at all.

How the Edge leaderboard is ordered. By the low end of that interval — edge − 1.96 standard errors — and not by the edge itself. Ranked on the point estimate, the top of the board is a list of wallets that went 21 for 21: the shrinkage in Step 2 pulls a thin sample toward zero but cannot pull a perfect one below a large, well-measured record, and a reader scanning the first screen would be reading noise in rank order. Ordering by the lower bound asks a different question — how good is this wallet at least — and a wallet answers it well only by having both a real edge and enough settled fills to have established it.

This changes the order and nothing else. Every wallet still shows its own measured edge, the same number it would show under any other sort, and the row names the bound it was ranked by beside its sample so the two are never confused. A wallet whose interval spans zero has a negative lower bound and sorts below wallets that look worse on the headline figure; that is the intended reading, not a defect.

Gates. Below 10 settled fills in the window there is no edge: it is withheld, not shown as zero. A category gets its own line only at 10 settled fills or more; below that the shrinkage pins the number near zero anyway and the line would read as measurement where there is none. Fills under the gate still count toward the pooled number and the sample size — they are not dropped, only unreported per category.

What is excluded. Fills in markets that settled VOID carry no win or loss and are removed rather than scored as either. Fills recorded after a market settled are removed as well: their price is no longer a forecast. Open positions in unsettled markets are not in the sample at all, so an edge never contains an unrealised mark.

The price band, and why it is nearly the whole range. A fill is scored when its entry price is above 0.001 and below 0.999. That band exists for one reason: the arithmetic in Step 1 divides by p on a BUY and by 1 − p on a SELL, and at a literal 0 or 1 there is nothing to divide by. It used to be 0.01 to 0.99, which quietly did a second job it was never measured for — deciding that a fill at 99½¢ is not a real bet.

It is one. Paying 99.9¢ to win $1 is a position taken at odds the trader accepted, and the premise of this whole metric is that the price is the odds. Scored, such a fill returns an edge of roughly nothing, which is the honest reading of a near-certain bet; excluded, it flattered every wallet whose sample happened to be built out of long-odds fills and made high-price sweeping look like an absence of evidence rather than an edge near zero. The defect the old band was really catching — cheap fills on markets that had already been decided — is removed directly by the post-settlement rule above, which is the filter that belongs to that problem. Widening the band brought 159 of the 258 wallets on the head-whale board into measurement, against 119 under the old one.

One thing the old band was quietly doing, kept separately. An admitted fill is scored at its entry price clamped to 0.01–0.99, which caps what a single fill can contribute at +99 per dollar risked (and at −1 when it loses). Nothing is dropped: a fill at 0.0011 is measured, at odds of 0.01. The cap exists because the Step 1 arithmetic is unbounded — at a tenth of a cent, one winning ticket is worth +999 and one wallet's error bar becomes that ticket's error bar rather than its own record's. The clamp can only touch fills between 0.001 and 0.01, or between 0.99 and 0.999 — the range the old band excluded outright.

A limit worth stating plainly: there is no pre-match / in-play split. Event start times are not in the dataset, so a position taken the day before a match and one taken at half-time score identically. For sports — the largest recorded category — that matters: entering in-play against a price that has already absorbed the first half is a different skill from entering beforehand, and this number cannot tell them apart. Only post-settlement fills are excluded, which is a weaker filter than an event-start filter. This is a known gap, not a rounding detail.

Why it is the wrong read for a market maker. A market-making or arbitrage wallet is not forecasting; it quotes both sides and earns the spread. Its settled fills therefore include large numbers of intentionally losing legs, and the resulting edge describes the payoff structure of that strategy rather than any judgement about outcomes. A high figure on such a wallet is not evidence of foresight and a low one is not evidence of its absence. Wallets whose classification is anything other than directional carry that caveat on their own page.

Edge is descriptive, not predictive. It reports how a settled sample paid relative to what it cost, over one window. It is not used as an alert threshold and it is not combined into a single 0–100 rating — a composite would put this number back behind a black box, which is the thing the page exists to avoid.

Recent position ROI divides total PnL across decided positions in the current 90-day source window by those positions’ total cost basis. It answers how the observed position set performed relative to capital put at risk; it does not estimate what the next trade will return.

PolyTrace shows the metric only when at least 10 positions are decided, the source window is current and fresh, every included position has a positive cost basis, and the total denominator is positive. If one row fails those checks, the entire metric is hidden rather than calculated from a flattering subset.

Four related terms on this site share a family resemblance and must not be read as one thing:

• A whale trade is an event: a single fill worth $1K+. It says the money moved; it says nothing about who moved it or how good they are.
• The whale tag describes a wallet’s size: more than $1M in lifetime Polymarket volume. Size only — a whale can be losing money, and some of the largest wallets tracked here are heavily unprofitable.
• The smart money tag describes a wallet’s results: more than $100,000 in volume, a win rate above 55%, and profitable overall. All three at once — volume alone, or profit on a thin sample, does not qualify.
• A signal event describes an evidence cluster: eligible fills in a bounded window passed a transparent ranking rule. Its Outcome direction, reason, opposing evidence, concentration, coverage and freshness remain attached; passing a rule does not make the trade predictive or copyable.

The practical consequence: a whale feed is dominated by size, and size is mostly noise. The interesting set is the overlap — large trades from wallets whose measured record earns attention. That is why every trade on this site links to the record of the wallet behind it, and why the guides read wallets rather than trades: how to track Polymarket whales covers the feed, and how to evaluate a Polymarket wallet covers the record.

• Recent position ROI is backward-looking. It describes one current source window and does not predict the next fill, market, or settlement result.
• Copyability is still a shadow metric. Detection delay, price movement, wallet type, market state, and the order book — bid/ask spread and executable slippage, frozen at discovery time — all feed the tier, but the tier is not shown to users yet. It needs a prospective validation window against a size-only baseline before it can support a trading claim.
• Win rate is the wrong question for some wallets. Market-making and arbitrage strategies take many tiny losses by design. A low win rate on such a wallet describes its strategy, not its skill.
• Profiles key on proxy wallets. A Polymarket proxy wallet is the address Polymarket exposes publicly, which is not necessarily the owner’s main wallet. One person may hold several; PolyTrace does not link them.
• PnL includes open positions. Unrealised profit moves with the market and is not money taken off the table.