Methodology · version 2.0 · updated October 27, 2026
How PolyConsensus works
PolyConsensus shows what a few thousand of Polymarket's best-performing traders hold, market by market. This page says whose positions we count, where the data comes from, how a consensus is formed, what our research found when we tested following it, and what the numbers do not tell you.
The short version
- We count the positions of about 2,000 to 2,500 tracked Polymarket wallets, from public on-chain trades and Polymarket's own data.
- A consensus is a head count of those wallets, not a forecast and not a probability.
- Following it did not make money: in a full year of data without hindsight, it lost about 3% a trade after costs. The consensus knows a little more than the price, not enough to pay the spread and the fee.
- The consensus is information, not a money-making signal. Nothing on this site is financial advice.
Whose positions we count
We follow around 2,000 to 2,500 Polymarket wallets at any time. They come from three places:
- Polymarket's monthly top 1,000 by profit. The leaderboard is read again every hour. A wallet that enters the top 1,000 is tracked from the next check; one that drops out stays tracked for 30 days before it is released.
- About 460 wallets picked by hand before October 2026. They are always tracked and carry a "Hand-picked" badge.
- Wallets that just left the monthly list and are still inside their 30 days. The top 1,000 changes a lot from month to month, which is what lifts the total above the two groups above.
A month's profit can be luck, so wallets from the monthly list also have to show a record (the win-rate gate, on since October 3, 2026). If a wallet has settled at least 10 markets in the last 90 days with $50 or more on one side, we keep it only if it won at least 55% of them. A wallet with fewer than 10 such markets has no verdict yet and stays while its history builds up; hand-picked wallets are not gated. A market where a wallet held both sides is a hedge and does not count.
We see wallets, not people: one trader can run several wallets, and we cannot see what anyone holds on other platforms or hedges outside Polymarket. A wallet's trades are indexed from the moment it is tracked; what it already held is taken from Polymarket's own position data.
Where the data comes from
Everything is public. Nothing comes from the traders themselves.
| What | Source |
|---|---|
| Trades of the tracked wallets | The Polygon blockchain: every fill on Polymarket's exchange contracts (OrderFilled events), counted for the wallet whose order was filled |
| Splits, merges, redemptions and conversions | The Polygon blockchain: events of Polymarket's token contracts, including the automatic redemptions Polymarket makes for holders |
| Current and closed positions, profit and loss | Polymarket's Data API, used to correct the positions built from trades |
| Market names, outcomes, end dates and results | Polymarket's Gamma API, including its list of markets that just closed |
| Prices, order books and spreads | Polymarket's order-book (CLOB) API |
| The tracked set | Polymarket's monthly leaderboard, plus the hand-picked list |
Positions are built from the trades and checked against Polymarket's own numbers. Polymarket's position data runs a few minutes behind the chain, so it is applied only when it is newer than the wallet's last trade; after that, the trades stand.
How fresh it is
| What | How often |
|---|---|
| New trades of the tracked wallets | Read from the chain every 30 seconds |
| Holders and consensus of listed markets | Re-counted within seconds of new trades, and every 30 seconds |
| A market reaching 5 tracked holders | Listed within about 2 minutes, usually seconds |
| Prices | Every minute |
| Splits, merges and redemptions | Every 2 minutes |
| Market details and results | Every 5 minutes; Polymarket's list of markets that just closed, every minute |
| Full recount of every market and trader | About every 75 minutes (hourly, never within 30 minutes of the previous one ending) |
| Tracked set and win-rate gate | Every hour |
| Trading-cost estimate (spreads) | Every 2 minutes, for up to 300 markets |
| Research figures | Offline studies, dated where they are shown |
Win rates, records and some figures on trader pages change only with the full recount. A trader who just sold can still show as a holder for a few minutes.
What a consensus is
A market is listed when at least 5 tracked wallets hold a position in it. Four rules decide the rest:
- Who counts as a holder. A tracked wallet holding at least 1 share of an outcome. What is left after a trader sells almost everything does not count.
- Arbitrage is left out. A wallet holding both sides with real money on each (the smaller side at least $20 and a quarter of the larger) is shown but not counted in the head count.
- Which side is the consensus. The outcome with the most money behind it among holders with at least $100 in the market (shares times average price). If nobody has $100 in, the side with more holders.
- The percentage. The share of holders, one vote per wallet, on that side. "7 of 8 hold Yes (88%)" means seven of eight wallets, whatever their size. For bots the API also reports the head count among the holders with $100 or more in ("staked").
You will also see the market's own price for the consensus side, the total value the holders have in the market, and the "smart money entry": the average price the consensus holders paid, weighted by size, against the price now (shown with at least 2 holders and $100 of cost). All of it describes the present: who holds what, not who will be right.
Listed, fading, forming, flips
- Listed (the main list): 5 or more tracked holders now.
- New: first reached 5 holders in the last 48 hours.
- Fading: 5 or more tracked wallets have held it, but only 1 to 4 still do. The smart money is leaving.
- Forming: 2 to 4 hand-picked wallets hold it, whatever the others hold. An early look, before a full consensus.
- Pending: the market has ended and is waiting for Polymarket's result.
- Flip: the consensus side changed, because the money behind the other side became larger. Flips show in "What changed today" within seconds of the trades behind them.
Grades and labels
Each open market gets a research grade, computed from the market as it is now, never from how it later turned out:
- Research candidate: the market matches a rule that looked positive after costs in our October 22 backtests, on small samples: $50k+ or $100k+ behind an 80%+ consensus, a quality net of +3, or 70%+ on an ATP/WTA tour match.
- Neutral: no measured edge either way.
- Avoid: the consensus side is under 30¢, or basketball or culture without a rule in its favour.
- In play, spreads and totals, under 60%, under 5 holders or a half-spread of 2¢ or more: at most Neutral.
Read the grade with the full-year results below. The year-long study tested the same ideas out of sample, and none held up: $50k+ on the consensus side made +3.1% a trade in the first half and lost 2.7% in the second; 3 or more proven winners on the side lost 1.9%. A Research candidate is a pattern that once looked good on a short window, not an edge. Every rule quotes its own evidence in the grade's tooltip.
Other labels you will see:
- Underdog: the consensus side is not the side the market prices higher. Under 40¢ it carries an amber note: longshots lose most.
- Cost ≈ X%: half the spread plus Polymarket's taker fee, as a share of the price, from the order book sampled every 2 minutes.
- Conviction bets: a tracked wallet opened a position at least 3 times its own usual size (its 75th percentile over 90 days) and at least $1,000. Size is not skill: these are mostly bets on favourites.
- Won, Lost, 50-50 and Ended after settlement: whether the consensus side was the winner, or the market was refunded at 50-50.
Trader quality and win rates
Two different records exist, for two different jobs. The 90-day record above decides which leaderboard wallets are tracked. The lifetime record judges a market's holders: among holders with at least $100 in the market, those with at least 20 settled markets and a win rate of 60% or more are counted as quality holders, once on the consensus side and once against it. On a market page the Quality chip shows this count: “N strong traders for · M against”. A net of +3 or more is highlighted. Developers can read it as the quality field of the API.
Win rates on trader pages count only markets that had a consensus (5 or more tracked holders); a win is a settled market where the trader's profit was positive, and a market refunded at 50-50 is neither. They are optimistic. Polymarket's data on closed positions almost never reports a loss; a position driven to zero tends to be dropped instead of listed as a loss. Measured on one top wallet in September 2026: 6,853 winning positions against 2 losing ones. A win rate in the high nineties is normal here, and it is not a measure of skill. Every win rate is shown with its wins, losses and count.
How markets are settled
- Only Polymarket's result counts. A market is resolved when Polymarket closes it with final prices, or its oracle reports the result. We never name a winner from a price on an open market.
- An end date is not an end. Polymarket's end date is an expected date: an election can go to a second round, a game can be postponed. A market stays open until Polymarket closes it.
- Pending means ended and waiting for the oracle, which can take hours, or days when a result is disputed.
- 50-50 refunds. When Polymarket refunds a market at 50¢ a share (a cancelled match, a tie, a walkover), it shows as "Resolved 50-50 (void)" and counts as neither a win nor a loss anywhere on the site.
- A market Polymarket closes leaves the live list within a minute or two; the full record of its holders is written at the next recount.
What the numbers do not mean
- The consensus percentage is not a win probability. 88% of holders on Yes is not an 88% chance of Yes. The market price is a much better estimate, and it already reflects what these traders did.
- The traders' profit is not what following them earns. They are on the list because they already won, which flatters any average. They bought earlier and cheaper than you can, size and exit their own way, and may hedge elsewhere. You buy later, cross the spread and pay Polymarket's fee.
- Accuracy at resolution is not a return. Figures like "majority right at resolution" on the analytics page describe the smart money's own result, at their own prices, with no costs. The research page is the one that measures following.
- A consensus on a favourite is mostly a consensus that the favourite wins. Agreement on a side priced at 85¢ is expected, not news.
What our research found
We tested whether following the consensus makes money. The answer is no. The first two studies below were written down in full before they were run, and each is judged on months it was not tuned on.
1. The site's own rule over a full year, without hindsight (October 26, 2026)
Every trade of every wallet on Polymarket from August 2025 to October 2026 (1.98 billion fills). The tracked set was rebuilt day by day, as the site would have seen it on each day, so nothing in it knows who did well later. The rule: 5 or more tracked holders and 60% or more on one side, buy at the first moment the rule holds, hold to the result. Costs: a 1-cent half-spread plus Polymarket's 2026 fees. Markets ending November 1, 2025 to October 18, 2026; the first half (November to April) is the training half, the second (May to October) the test.
| What was tested | Result per trade | Trades |
|---|---|---|
| The site's rule, test half (May–Oct 2026) | −3.0% (95% range −3.6% to −2.4%) | 86,296 |
| The same, training half (Nov 2025–Apr 2026) | −4.5% | 44,213 |
| Months with a loss, November 2025 to October 2026 | 12 of 12, between −1.8% and −6.9% | full year |
| With no fee at all / with a 2-cent half-spread | −1.4% / −4.8% | 86,296 |
| A strong consensus: 80%+ and 10+ holders | −1.6% | 17,553 |
| Favourites, priced 70 to 95 cents | about −1%, the least bad | — |
| Longshots, under 20 cents | −20% | — |
| $50k+ on the side: training half, then test half | +3.1%, then −2.7% | 641 in the test |
| Switching sides on a flip, against holding | 5.5 points a trade worse | paired |
| Four rules for automated bots (in play, esports), full year | −1.0% to −1.6% | 29,249 to 115,154 |
- The consensus knows a little. Its side wins 0.64 points more often than its price says, which is not luck (p < 0.0001). The spread and the fee cost about 2 points, so following it loses.
- Stronger consensus loses less, but still loses. All nine combinations of 3/5/10+ holders and 60/70/80%+ agreement were negative, every one with its whole 95% range below zero.
- No narrower idea survived. 86 filters were chosen on the first half (price bands, categories and sports, favourites and underdogs, time to resolution, holder count, money on the side, proven winners, exit rules, six other ways of picking the wallets). 7 were positive in the first half; none was positive and significant in the second.
- Hold, do not trade the flips. Selling when the consensus is lost, or switching sides on a flip, was worse than holding in every test.
- The earlier "break-even" year was mostly hindsight. An earlier study that used today's top traders showed +0.5% a trade on the same rule and window; with the wallets the site would actually have tracked at the time, it is −3.5%. About 4 points came from picking wallets after we knew they had won.
2. A model that weights each trader by skill (October 27, 2026)
We built a model that weighs every holder by their own track record up to that hour, drops dust, hedgers and market-maker-like wallets, merges wallets that trade as one, and turns the holders into one fair probability. It was retrained every month on the past only and tested on May to October 2026 (95,861 consensus formations).
- It forecasts no better than the market price. The best version was level with the price; the fuller ones were slightly worse. The holders added nothing measurable.
- No trading rule built on it made money. 16 versions were tested; none passed (the best p-value was 0.385). The best-looking one made +3.7% a trade on only 552 trades, with a 95% range of −10.8% to +22.3%: a coin flip, not an edge.
- Why. Traders with a strong record really do win, by 6.6 points against the price they paid. By the time enough of them hold a market for the site to notice, the price has moved (from 63¢ to 69¢ on average), and against the price a follower pays their edge is 0.1 points. The skill is real, but it is speed. (This last finding was looked at after the main results.)
3. Copying single traders (October 26, 2026)
27,345 wallets that were ever in the tracked set as it stood at the time, with every order-book fill each one made from December 2022 to October 25, 2026. A copier mirrors every buy and sell at a fixed fraction of the trader's size, pays the spread and fee on both, and holds the rest to the result.
| What | Result |
|---|---|
| Copying every buy and sell, at the traders' own prices | −6.3% per dollar over 42.7 million settled markets |
| The same, copied an hour later | −15.4% per dollar |
| Wallets clearly profitable at their own prices (95% range above zero) | 1,235 of 19,954 with 30+ settled markets (about 500 would pass by luck alone) |
| Still clearly profitable when copied an hour later | 159 of 17,773 (with this many wallets, about 440 would pass by luck alone) |
A wallet's figures cover its whole history, including the run that got it on the leaderboard, so even the at-their-prices count flatters them. Each trader's own figures are on their page and in the copy-trader research.
Earlier studies
An October 22, 2026 backtest on two short windows found the consensus lost 4.8% a trade after costs (10,051 trades), with a few slices that looked positive; those slices are the ones the full-year study then failed to confirm. An October 25 study of today's top 1,937 traders over a year showed about break-even, which the no-hindsight rebuild above traced to hindsight in choosing the wallets.
What we watch from here
The live signal ledger, started October 25, 2026, freezes every consensus signal when it forms and settles it when Polymarket resolves the market, after costs. It cannot be fitted after the fact, so it is the fresh test. The research lab lets you replay your own rule with the same costs and checks.
The fair reading: the consensus is a useful way to see what skilled traders hold and where they disagree with the crowd. It is information, not a money-making signal. If you trade, start small and remember that Polymarket charges fees (see the fee calculator). Nothing on this site is financial advice.
Known limits
- Lag. Trades reach the site within seconds to a couple of minutes, Polymarket's position data a few minutes later, and records only with the full recount (about every 75 minutes). A position can be briefly stale.
- Wallets we cannot see. Only tracked wallets count: a few thousand out of everyone on Polymarket. A trader's other wallets, positions moved by transfer, and anything off Polymarket are invisible.
- New markets. A market Polymarket has only just created can take a minute or two to be named; trades in it are counted once it is.
- Resolution delays. A result waits for Polymarket and its oracle; disputed markets can sit in Pending for days.
- Paper bots are paper. The bots on account dashboards, the live ledger and the research lab trade on paper: prices from the order book or the last trade, no real orders, so real fills at size would cost more.
- The research rebuilds the past. Our day-by-day tracked set is a reconstruction (it matches Polymarket's real top 1,000 for 68.5% of wallets; every other way we tried of picking the wallets gave about the same result), prices are the last trade rather than the order book, and a 1-cent half-spread is optimistic for thin markets. Those last two flatter the results; real trading would do worse.
Quick answers
Which traders does PolyConsensus follow?
Around 2,000 to 2,500 Polymarket wallets at any time: the monthly top 1,000 traders by profit (re-read every hour), about 460 wallets picked by hand, and wallets that dropped off the monthly list in the last 30 days. A wallet from the monthly list that has settled at least 10 markets in the last 90 days with $50 or more on one side is only kept if it won at least 55% of them.
What does "consensus" mean?
A market is listed when at least 5 tracked wallets hold at least 1 share of it. The consensus side is the outcome with the most money behind it among holders with $100 or more in the market, and the consensus percentage is the share of holders, one vote per wallet, on that side. "7 of 8 hold Yes (88%)" is a head count, not a forecast.
Is the consensus percentage the chance that the side wins?
No. 88% means 88% of the tracked holders hold that side. It says nothing by itself about the probability of the outcome, and the market price already contains nearly all of what these traders know.
Does following the smart money make money?
Not in our research. In a full-year study without hindsight (October 26, 2026), buying the consensus side as soon as the site's rule held and keeping it to the result lost 3.0% a trade after a 1-cent spread and Polymarket's fees in the test half of the year (May to October 2026, 86,296 trades), and every month of the year was negative. None of 86 narrower ideas held up out of sample.
Can I copy a single top trader instead?
Rarely with success. Over the full history of 27,345 wallets that were ever tracked, copying every buy and sell after costs lost 6.3% per dollar, and 15.4% when copied an hour later. Of 19,954 wallets with 30 or more settled markets, 1,235 were clearly profitable at their own prices; copied an hour later, only 159 of 17,773 were, fewer than luck alone would produce.
What do Research candidate, Neutral and Avoid mean?
A label from our October 22 backtests, computed from the market as it is now. Research candidate: the market matches a rule that looked positive after costs on small samples. Neutral: no measured edge. Avoid: a pattern that lost clearly more than the costs, such as a consensus side under 30 cents. The full-year study later tested the same ideas out of sample and none held up, so a Research candidate is a pattern that once looked good, not an edge.
Why do some traders show a win rate near 100%?
Polymarket's data on closed positions almost never reports a loss: a position driven to zero tends to be dropped instead of listed as a loss. Our win rates count only what that data shows, so they are optimistic. Treat a very high win rate as a feature of the data, not as proof of skill.
How fresh is the data?
New trades of the tracked wallets are read from the Polygon chain every 30 seconds, and the holders of listed markets are re-counted within seconds of them and every 30 seconds. Prices refresh once a minute, a market Polymarket closes leaves the live list within a minute or two, and a full recount of every market runs about every 75 minutes.
Does PolyConsensus charge a fee or give financial advice?
The site adds no fee of its own; trading on Polymarket carries Polymarket's fees (see the fee calculator) and the spread. Nothing here is financial advice, and PolyConsensus is not affiliated with Polymarket.
Version history
Version 2.0, last updated October 27, 2026. When a rule or a figure on this page changes, the version goes up and a row is added here. Site changes in general are on What's new.
| Version | Date | What changed |
|---|---|---|
| 2.0 | Moved from /how-it-works to /methodology. The full-year, no-hindsight study of October 26 replaces the October 22 backtest as the headline result. Added the skill-weighted model and the copy-trader research, where the data comes from, how fresh it is, the market lists, how markets are settled, and the known limits. | |
| 1.1 | The research grade (Research candidate, Neutral, Avoid) replaced the 0 to 100 Strong / Medium / Weak score, and quality holders are judged by their lifetime record. | |
| 1.0 | First version, published as How it works: the tracked traders, the consensus rules, the signal badges and the October 22 backtest. |
PolyConsensus is not affiliated with Polymarket. Explore: markets, 2026 midterm markets, the trader leaderboard, research, large trades, analytics, the API, terms and privacy.