Does following the smart money pay?
What buying the smart money consensus would have returned, after a 1¢ spread and Polymarket's taker fee: over a full year rebuilt without hindsight, then live, signal by signal, since the ledger started. Past results, not advice.
The short answer
No. Over a full year of Polymarket data, with the tracked wallets rebuilt day by day as the site would have seen them (no hindsight), buying the consensus side as soon as the site's rule held and keeping it to the result lost 3.0 % a trade after costs in the half of the year the study was not tuned on (May–October 2026, 86,296 trades, 95 % CI −3.6 to −2.4 %). All 12 months were negative. The consensus does know something, 0.4 points more than its price (hit rate minus average price), but the spread and the fee cost about 2 points. None of 86 narrower ideas held up out of sample. The earlier "break-even" result came mostly from choosing the wallets with hindsight. The consensus is information, not a money-making signal.
The 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 for each day from that day's data alone: the monthly top 1,000 by profit, the 30-day grace, the win-rate gate and the hand-picked wallets. The rule is the site's: 5+ tracked holders and 60 %+ on one side, entered at the first moment it holds (at least 1 h before the end date, price 0.05–0.95), held to the result, $1 a trade, a 1¢ half-spread plus Polymarket's 2026 taker fee. Markets ending November 1, 2025 to October 18, 2026. The study was written down before it was run: November–April chooses, May–October tests.
| 5+ holders, 60 %+, held, 1¢ + taker fee | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Test half, May–Oct 2026 | 86,296 | −3.0 % | −3.6 to −2.4 % |
| Training half, Nov 2025–Apr 2026 | 44,213 | −4.5 % | −5.5 to −3.5 % |
| Full year (descriptive) | 132,881 | −3.4 % | −4.0 to −2.9 % |
| Test half, no fee at all | 86,296 | −1.4 % | −2.0 to −0.7 % |
| Test half, 2¢ half-spread | 86,296 | −4.8 % | −5.3 to −4.2 % |
| Test half, entered 1 h late | −3.5 % | −4.2 to −2.8 % | |
| Buy the other side instead (test half) | 86,268 | −11.2 % | −12.2 to −10.1 % |
Every month from November 2025 to October 2026 was negative, between −1.8 % and −6.9 % a trade. The consensus side won 64.2 % of the test-half trades at an average price of 0.638: it beats its price by 0.4 points (64.2 % minus 63.8 %; not luck, p < 0.0001), and pays about 2 points to trade.
Stronger consensus loses less, but still loses
| Agreement | 3+ holders | 5+ holders | 10+ holders |
|---|---|---|---|
| 60 %+ | −3.1 % (165,467) | −3.0 % (86,296) | −2.1 % (30,790) |
| 70 %+ | −2.4 % (115,875) | −2.7 % (65,373) | −1.6 % (24,335) |
| 80 %+ | −2.0 % (91,983) | −2.6 % (56,164) | −1.6 % (17,553) |
All nine are negative with their whole 95 % interval below zero.
By price
| Price of the side | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| 0.05 – 0.20 (longshots) | 3,615 | −20.5 % | −28.2 to −13.2 % |
| 0.20 – 0.30 | 3,266 | −3.3 % | −9.1 to +2.9 % |
| 0.30 – 0.50 | 16,830 | −4.8 % | −6.6 to −3.1 % |
| 0.50 – 0.70 | 23,546 | −2.9 % | −3.9 to −1.9 % |
| 0.70 – 0.90 (favourites) | 26,050 | −0.9 % | −1.5 to −0.3 % |
| 0.90 – 0.95 | 12,989 | −0.2 % | −0.6 to +0.3 % |
The leads, chosen in the first half and tested in the second
86 filters were tried in the training 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, and the 15 candidates of the October 22 backtest. Only the 7 that were positive in training went on to the test, which needed p < 0.0071 (corrected for 7 tries). None passed; none was even promising.
| Lead | Training half | Test half | Test trades |
|---|---|---|---|
| Geopolitics | +3.7 % | +1.6 % (CI −4.1 to +7.1) | 928 |
| Finance | +2.1 % | −1.4 % | 716 |
| Economy | +1.5 % | +2.5 % (p 0.65) | 287 |
| $50k+ on the consensus side | +3.1 % | −2.7 % (CI −7.6 to +2.9) | 641 |
| 80 %+, side worth $100k+ | +2.9 % | −0.3 % | 835 |
| 80 %+, $100k+ more on the side than against | +8.2 % | −0.9 % | 680 |
| 60 %+, $100k+ staked, sell 6 h after the consensus is lost | +0.7 % | −2.5 % | 1,937 |
- Proven winners on the side: 3+ in the market with two thirds on the consensus side lost −1.9 % a trade (49,161, CI −2.6 to −1.1 %).
- Categories and sports: none significantly positive. Tennis −3.4 % (12,026), soccer −3.1 %, esports −2.5 %, crypto −3.4 %, basketball −4.8 %; politics +1.3 % and MMA +5.8 % (337 trades) positive but not significant.
- Exits: holding beat every exit. Selling 1 h after the consensus is lost cost 1.2 points a trade; switching sides when the money flips cost 5.5 points (−8.5 % a trade).
- Other ways of picking the wallets (no win-rate gate, top 500, top 2,000, realized profit, volume): all about −3 %. The hand-picked wallets alone: −0.6 % (2,939 trades, CI −4.6 to +3.7 %).
- Automated rules: four in-play and esports bot rules lost −1.0 % to −1.6 % a trade over the full year (29,249 to 115,154 trades), −2.2 % to −3.1 % at a 2¢ half-spread.
- Descriptive only, it uses the future: test trades whose money side later flipped lost −28.2 % (30,071); those that never flipped made +10.5 % (56,225). Nobody knows at entry which is which, so it cannot be traded.
Why the earlier year looked break-even
| Wallets | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Today's 1,937 leaderboard wallets, all year (hindsight) | 44,597 | +0.5 % | −0.3 to +1.4 % |
| The wallets the site would have tracked each day | 129,209 | −3.5 % | −4.1 to −3.0 % |
| The difference on the 34,886 markets both traded | −2.4 points | −3.7 to −1.1 |
About 4 points of the old result came from picking wallets after we knew they had won. Today's top traders held only 22 % of the tracked set's member-days during the year: most of the time, the site would have been following wallets that later fell away.
Caveats, none of which changes the answer: our day-by-day tracked set is a reconstruction from order-book profit (it matches Polymarket's real top 1,000 for 68.5 % of wallets, and every other way of picking the set gives about −3 %); prices are the last trade, not the order book, so a 1¢ half-spread flatters thin markets; neg-risk conversions are not applied. Source: the pre-registered phase 2 report of October 26, 2026.
What holds in every study
Four datasets (the hand-picked set in spring, every tracked wallet in October, today's top traders over a year with hindsight, and the full year without hindsight). Exact figures differ; these patterns do not:
- Following every consensus loses after costs. −3.0 % a trade out of sample over the full year; −4.8 % in the October 22 windows.
- Longshots lose most. Under 0.20 the consensus side lost −20.5 % a trade over the full year.
- Favourites lose least, but still lose. 0.70–0.90 lost −0.9 % a trade over the full year (the earlier hindsight figures of +0.4 % to +1.4 % did not survive).
- Hold to the result. Selling when the consensus fades, switching on a flip or taking profit at 0.95 cost money in every study.
- Costs decide. Every cent of half-spread costs about 2 points a trade.
- "Elite against the crowd" loses. Following the top wallets when they disagree with the rest lost badly.
Can a skill-weighted model beat the price?
No. We built a model that weighs every holder by their own track record up to that hour (shrunk when the record is short, recent results counting more), drops dust, hedgers, arbitrageurs and market-maker-like wallets, merges wallets that trade as one, adds bigger-than-usual bets and "proven" holders, and turns the holders into one fair probability. It was pre-registered, retrained every month on the past only and tested on the full year without hindsight: May–October 2026, 95,861 consensus formations (October 27, 2026).
- As a forecaster it is no better than the price. The best version was level with the price (log loss −0.11 thousandths a decision, 95 % CI −0.34 to +0.08); the fuller ones were slightly worse. The holders added nothing measurable.
- As a trading rule, every version loses or is noise. 16 versions were pre-registered; none passed the correction for 16 tries (the smallest p was 0.385).
| Rule | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Logistic model, any edge | 27,832 | −1.5 % | −2.3 to −0.6 % |
| XGBoost, any edge | 22,561 | −4.3 % | −5.9 to −2.7 % |
| XGBoost, 2-point edge | 2,798 | −6.8 % | −12.7 to −0.6 % |
| Small logistic, 2-point edge (the best-looking cell) | 552 | +3.7 % | −10.8 to +22.3 % |
| Control: the site's plain consensus | 83,500 | −3.0 % | −3.6 to −2.3 % |
The best-looking cell is a coin flip, not an edge: 552 trades with an interval from −10.8 % to +22.3 %. Why it fails (looked at after the main results): holders with a strong record really do win, by 6.6 points against the price they paid (0.63 on average). By the time the consensus forms the price is 0.69, and against the price a follower pays their edge is 0.1 points. The skill is real, but it is speed; costs of about 2 points then turn a zero edge into a loss.
Copying one trader
The copy-trader research asks what following a single wallet would have returned: mirror every buy and sell at a fixed fraction of their size, pay the spread and fee on both, hold the rest to the result. On the full history (27,345 wallets that were ever in the tracked set as it stood at the time, every order-book fill from December 2022 to October 25, 2026; snapshot of October 26, 2026):
| What | Result |
|---|---|
| All wallets, at their own prices | −6.3 % per $1, over 42.7 million settled markets |
| All wallets, copied an hour later | −15.4 % per $1 |
| Clearly profitable at their own prices (95 % interval above zero) | 1,235 of 19,954 wallets with 30+ settled markets (about 500 would pass by luck alone) |
| Still clearly profitable when copied an hour later | 159 of 17,773 (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. Copying within minutes of a trader's own entry, before the price moves, has not been tested yet.
What we know, and what we do not
Measured, out of sample
- Following the consensus loses after costs: −3.0 % a trade, every month of the year.
- No narrower rule tested beats the costs: 0 of 86 leads, 0 of 16 model rules.
- Longshots lose most; favourites lose least; holding beats every exit.
- Skilled traders exist, but their edge is in the price they get, not in what a follower can buy later.
Not known yet
- Whether copying a skilled trader within minutes of their trade can keep their edge (not yet tested).
- What the live ledger below will show over months: the one record that cannot be fitted after the fact.
- How much worse real fills at size are than the 1¢ half-spread assumed.
Updated 2026-10-27. Sources: the pre-registered full-year reports of October 26 (the site's rule) and October 27, 2026 (the skill-weighted model), the copy-trader snapshot of October 26, and, as history below, the October 22 backtest and the October 25 one-year study. How the consensus is counted and what every label means is on the methodology page.
Live ledger
Every signal since the ledger started, frozen when the consensus formed and settled when Polymarket resolved the market. The figures say what following the consensus at formation would have returned, after a 1¢ spread and Polymarket's taker fee. Past results, not advice.
Recording since 2026-10-25 06:50 UTC. 36,398 signals so far: 32,217 settled and 4,181 still open. Figures updated 2026-11-01 13:38 UTC.
By price at formation
| Price band | Settled | Won · lost · 50-50 | Hit rate | Break-even | Avg price | Return a trade, after costs |
|---|---|---|---|---|---|---|
| Under 0.05 | 290 | 22 · 268 · 0 | 7.6 % 5.1 % to 11.2 % | 3.0 % | 0.02 | +98.7 % +9.1 % to +188.4 % |
| 0.05 – 0.30 (longshots) | 2,204 | 708 · 1494 · 2 | 32.2 % 30.2 % to 34.1 % | 21.6 % | 0.20 | +61.8 % +50.4 % to +73.1 % |
| 0.30 – 0.50 | 9,787 | 5100 · 4680 · 7 | 52.1 % 51.2 % to 53.1 % | 48.2 % | 0.46 | +7.9 % +5.8 % to +10.0 % |
| 0.50 – 0.70 | 12,899 | 7771 · 5114 · 14 | 60.3 % 59.5 % to 61.2 % | 56.9 % | 0.54 | +6.6 % +5.1 % to +8.1 % |
| 0.70 – 0.95 (favourites) | 5,878 | 4658 · 1214 · 6 | 79.3 % 78.3 % to 80.3 % | 83.4 % | 0.82 | −5.0 % −6.3 % to −3.7 % |
| Over 0.95 | 1,159 | 1106 · 53 · 0 | 95.4 % 94.1 % to 96.5 % | 98.8 % | 0.98 | −3.4 % −4.6 % to −2.2 % |
By topic (the 10 with most settled signals)
| Topic | Settled | Won · lost · 50-50 | Hit rate | Break-even | Avg price | Return a trade, after costs |
|---|---|---|---|---|---|---|
| Crypto | 13,642 | 8378 · 5264 · 0 | 61.4 % 60.6 % to 62.2 % | 56.6 % | 0.54 | +10.9 % +8.9 % to +12.9 % |
| Soccer | 9,897 | 6095 · 3802 · 0 | 61.6 % 60.6 % to 62.5 % | 59.8 % | 0.58 | +9.0 % +6.2 % to +11.8 % |
| NBA | 2,379 | 1232 · 1147 · 0 | 51.8 % 49.8 % to 53.8 % | 53.9 % | 0.52 | +10.1 % +4.2 % to +15.9 % |
| Weather | 1,334 | 797 · 537 · 0 | 59.7 % 57.1 % to 62.3 % | 60.8 % | 0.59 | +1.7 % −7.1 % to +10.5 % |
| Tennis | 1,249 | 723 · 504 · 22 | 58.9 % 56.1 % to 61.6 % | 58.5 % | 0.56 | +3.1 % −3.1 % to +9.4 % |
| Esports | 1,102 | 639 · 457 · 6 | 58.3 % 55.4 % to 61.2 % | 58.4 % | 0.56 | +0.6 % −4.9 % to +6.2 % |
| College Football | 702 | 384 · 318 · 0 | 54.7 % 51.0 % to 58.3 % | 54.2 % | 0.52 | +8.4 % −0.5 % to +17.2 % |
| Other Sports | 442 | 263 · 179 · 0 | 59.5 % 54.9 % to 64.0 % | 53.4 % | 0.51 | +18.5 % +2.9 % to +34.1 % |
| NFL | 430 | 226 · 204 · 0 | 52.6 % 47.8 % to 57.2 % | 55.3 % | 0.53 | +14.5 % +0.2 % to +28.7 % |
| NHL | 399 | 202 · 197 · 0 | 50.6 % 45.7 % to 55.5 % | 53.0 % | 0.51 | +9.4 % −6.7 % to +25.5 % |
By quality net at formation
Quality net = holders with a strong record (60 %+ wins over 20+ settled markets, $100+ in the market) on the side, minus those against it.
| Quality net | Settled | Won · lost · 50-50 | Hit rate | Break-even | Avg price | Return a trade, after costs |
|---|---|---|---|---|---|---|
| Net 0 or less | 27,499 | 15945 · 11534 · 20 | 58.0 % 57.4 % to 58.6 % | 56.3 % | 0.54 | +7.3 % +5.7 % to +8.9 % |
| Net +1 to +2 | 4,550 | 3277 · 1264 · 9 | 72.2 % 70.8 % to 73.4 % | 65.3 % | 0.63 | +20.1 % +16.5 % to +23.6 % |
| Net +3 or more | 168 | 143 · 25 · 0 | 85.1 % 79.0 % to 89.7 % | 82.0 % | 0.81 | +12.4 % −2.0 % to +26.7 % |
| No quality data | 0 | — | — | — | — | — |
By share of holders on the side
| Head count | Settled | Won · lost · 50-50 | Hit rate | Break-even | Avg price | Return a trade, after costs |
|---|---|---|---|---|---|---|
| Under 60 % of holders | 11,597 | 6770 · 4815 · 12 | 58.4 % 57.5 % to 59.3 % | 58.8 % | 0.57 | +4.5 % +2.0 % to +7.0 % |
| 60 – 79 % | 12,085 | 7165 · 4910 · 10 | 59.3 % 58.5 % to 60.2 % | 57.0 % | 0.55 | +10.0 % +7.4 % to +12.6 % |
| 80 % or more | 8,535 | 5430 · 3098 · 7 | 63.7 % 62.6 % to 64.7 % | 57.3 % | 0.55 | +14.2 % +11.9 % to +16.5 % |
| Unknown | 0 | — | — | — | — | — |
Formed or flipped, in play or not
| Kind | Settled | Won · lost · 50-50 | Hit rate | Break-even | Avg price | Return a trade, after costs |
|---|---|---|---|---|---|---|
| Consensus formed | 22,968 | 13839 · 9106 · 23 | 60.3 % 59.7 % to 60.9 % | 56.9 % | 0.55 | +11.7 % +10.1 % to +13.3 % |
| Consensus flipped to this side | 9,249 | 5526 · 3717 · 6 | 59.8 % 58.8 % to 60.8 % | 59.8 % | 0.58 | +2.8 % −0.3 % to +5.9 % |
| Not in play when it formed | 18,984 | 11412 · 7557 · 15 | 60.2 % 59.5 % to 60.9 % | 57.2 % | 0.55 | +8.2 % +6.4 % to +10.0 % |
| Match in play when it formed | 5,677 | 3425 · 2244 · 8 | 60.4 % 59.1 % to 61.7 % | 59.0 % | 0.57 | +14.4 % +10.3 % to +18.6 % |
| Unknown | 7,556 | 4528 · 3022 · 6 | 60.0 % 58.9 % to 61.1 % | 58.0 % | 0.56 | +7.5 % +4.8 % to +10.3 % |
By how the signal was seen
| Recorder | Settled | Won · lost · 50-50 | Hit rate | Break-even | Avg price | Return a trade, after costs |
|---|---|---|---|---|---|---|
| Seen by the hourly recount | 286 | 154 · 132 · 0 | 53.8 % 48.1 % to 59.5 % | 53.7 % | 0.52 | −14.5 % −25.4 % to −3.7 % |
| Seen live | 31,931 | 19211 · 12691 · 29 | 60.2 % 59.7 % to 60.8 % | 57.8 % | 0.56 | +9.4 % +7.9 % to +10.8 % |
Newest settled signals
- Ethereum Up or Down - November 1, 8:30AM-8:35AM ETUp at 0.90 · flipped 2026-11-01 · CryptoWon +10 %
- Ethereum Up or Down - November 1, 8:30AM-8:35AM ETDown at 0.51 · formed 2026-11-01 · CryptoLost −100 %
- Taranto: Andrea Colombo vs Adrian AndreevAdrian Andreev at 0.69 · formed 2026-11-01 · TennisWon +40 %
- Set 1 Winner: Juan Manuel Cerundolo vs Miomir KecmanovicJuan Manuel Cerundolo at 0.50 · formed 2026-11-01 · TennisWon +91 %
- Bratislava 2: Max Houkes vs Dominik MacejMax Houkes at 0.97 · flipped 2026-11-01 · TennisWon +2 %
- Bratislava 2: Max Houkes vs Dominik MacejDominik Macej at 0.03 · formed 2026-11-01 · TennisLost −100 %
- Bratislava 2: Oleg Prihodko vs Yannik KelmOleg Prihodko at 0.84 · flipped 2026-11-01 · TennisLost −100 %
- Counter-Strike: The Huns Esports vs 5star (BO3) - Moscow Cyber Games PlayoffsThe Huns Esports at 0.49 · formed 2026-11-01 · EsportsWon +93 %
- Rolex Paris Masters, Qualification: Thiago Agustin Tirante vs Mattia BellucciMattia Bellucci at 0.64 · flipped 2026-11-01 · TennisWon +51 %
- Taranto: Leonardo Rossi vs Alessio BalestrieriAlessio Balestrieri at 0.48 · formed 2026-11-01 · Tennis50-50 −1 %
- Rolex Paris Masters, Qualification: Rinky Hijikata vs Vit KoprivaRinky Hijikata at 0.69 · formed 2026-11-01 · TennisWon +42 %
- Taranto: Manas Dhamne vs Loann MassardLoann Massard at 0.58 · flipped 2026-11-01 · TennisWon +66 %
- Taipei: Alexander Donski vs Omar JasikaAlexander Donski at 0.42 · formed 2026-11-01 · TennisLost −100 %
- Ryukyu Golden Kings vs. SAN-EN NeoPhoenixSAN-EN NeoPhoenix at 0.50 · formed 2026-11-01 · Other SportsWon +91 %
- Taranto: Manas Dhamne vs Loann MassardManas Dhamne at 0.69 · formed 2026-11-01 · TennisLost −100 %
Newest open signals
- Odense BK vs. Silkeborg IF: O/U 1.5Over at 0.82 · formed 2026-11-01 · Socceropen
- Exact Score: Falkirk FC 2 - 0 Celtic FC?Yes at 0.50 · formed 2026-11-01 · Socceropen
- Ethereum Up or Down - November 1, 8:35AM-8:40AM ETUp at 0.14 · flipped 2026-11-01 · Cryptoopen
- Will SBV Excelsior win on 2026-11-01?No at 0.95 · flipped 2026-11-01 · Socceropen
- SBV Excelsior vs. FC Twente: O/U 3.5Under at 0.56 · formed 2026-11-01 · Socceropen
- Telstar 1963 vs. FC Utrecht: O/U 3.5Under at 0.64 · formed 2026-11-01 · Socceropen
- Viborg FF vs. AC Horsens: 1st Half O/U 0.5Under at 0.26 · formed 2026-11-01 · Socceropen
- Solana Up or Down - November 1, 8:35AM-8:40AM ETDown at 0.45 · formed 2026-11-01 · Cryptoopen
- XRP Up or Down - November 1, 8:35AM-8:40AM ETDown at 0.62 · formed 2026-11-01 · Cryptoopen
- Will Ascoli Calcio 1898 FC win on 2026-11-01?Yes at 0.35 · formed 2026-11-01 · Socceropen
- Djurgardens IF vs. Degerfors IF: Both Teams to ScoreNo at 0.47 · formed 2026-11-01 · Socceropen
- Will SBV Excelsior vs. FC Twente end in a draw?No at 0.79 · formed 2026-11-01 · Socceropen
- Viborg FF vs. AC Horsens: 1st Half O/U 1.5Under at 0.52 · formed 2026-11-01 · Socceropen
- Will Degerfors IF win on 2026-11-01?No at 0.92 · formed 2026-11-01 · Socceropen
- Paris FC vs. AS Monaco FC: O/U 2.5Under at 0.47 · flipped 2026-11-01 · Socceropen
How a signal is recorded
- One signal per market and side. A signal is opened the first time a market on the consensus list (5+ tracked holders, still open, before its end date) has its consensus on a side, with the price, holders, money and quality of that moment frozen. If the consensus flips, the new side gets a signal of its own and the old one stays: a follower holds to resolution (holding beat every exit in the backtest). A flip back opens nothing.
- When. Most signals are recorded from the live change log within seconds of the consensus forming, at the price shown then. The rest are recorded at the hourly recount of every market, up to about an hour later, at the price of that moment.
- Costs. The entry is the quoted price plus a 1¢ half-spread, plus Polymarket's taker fee for the market's category at that price (fee schedule). A share pays $1 if the side wins, nothing if it loses, $0.50 if Polymarket resolves the market 50-50. A 50-50 counts in the return, never as a win or a loss.
- Settlement. Only Polymarket's own result counts, never a price. A result Polymarket changes within three days is corrected here.
- Intervals. 95 % intervals: Wilson for hit rates, a normal approximation for mean returns; none under 20 settled signals. Several markets of one event move together, so the true uncertainty is larger than these intervals.
- Left out. 1,368 markets already had a consensus when the ledger started; they were not seen forming and are not in these figures (335 have settled, −8.5 % a trade after costs if bought at the start).
For history before the ledger started, the market research page replays the site's rule over the exported fills, hour by hour (reconstructed, kept apart from the live ledger): how similar signals turned out by topic, price, money and quality, and what prices did after a consensus formed, flipped or lost its holders.
Test it yourself
The research lab replays your own rule (who to follow, topic, price, money, quality, timing, exit) over the same history, with the same costs and checks as the backtest; every run is public in the library. The signal explorer tests eight more ideas, pre-registered before looking at the results: nothing passed. The copy-trader research shows what copying each tracked trader would have returned after costs, at their entry and an hour later, and whether their unusually large bets did better.
Earlier studies, kept as history
Before the full year was rebuilt without hindsight, we ran shorter or hindsight-flattered studies. Their figures are kept here, dated, because other pages and old links quote them. The full-year study above replaces them. Every positive rule below was tested there out of sample, and none held up.
October 22, 2026: two short windows (superseded)
A point-in-time replay of the site's own rules: a trade enters the consensus side when a market reaches 5+ tracked holders, at the price of that moment plus a 1¢ half-spread and Polymarket's 2026 taker fee, $1 a trade, held to resolution. Two windows: A, March 10 – April 28 2026 with the hand-picked wallets (clean data), and B, October 2 – 22 2026 with every tracked wallet.
Window B is provisional: part of October's fill times were stamped early (a block-time error, being repaired), so some trades saw fills before they happened. The windows also disagree with each other: across 752 strategies tested in both, their rankings were negatively correlated.
The broad rule
| 5+ holders, 60 %+ on the side, price 0.05–0.95, held | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Buy the consensus side (1¢ + taker fee) | 10,051 | −4.8 % | −6.5 to −3.1 % |
| Same, without the taker fee | 10,051 | −3.2 % | −4.9 to −1.4 % |
| Same, at a 2¢ half-spread | 10,051 | −6.4 % | −8.0 to −4.8 % |
| Buy the other side instead | 10,056 | −16.5 % | −19.3 to −13.7 % |
The consensus side won 64.7 % of the time at an average price of 0.638: about one point better than its price, before costs.
By price
| Price of the side | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| 0.05 – 0.30 (longshots) | 1,625 | −29.5 % | not reported |
| 0.30 – 0.50 | 2,793 | −6.5 % | not reported |
| 0.50 – 0.70 | 3,834 | −3.2 % | not reported |
| 0.70 – 0.90 (favourites) | 4,327 | +0.4 % | −0.9 to +1.7 % |
| 0.90 – 0.99 | 4,210 | −0.2 % | not reported |
Slices that looked better (picked after seeing the tables)
| Rule | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| 80 %+ holders, $50k+ staked on the side, price 0.50–0.95 | 172 | +10.3 % | +2.1 to +18.3 % |
| 80 %+ holders, side worth $100k+ | 91 | +13.2 % | +2.0 to +24.4 % |
| Quality net +3 or more when the consensus forms | 358 | +7.2 % | +0.9 to +13.6 % |
| Tennis, ATP/WTA tour, 70 %+ holders (window A) | 164 | +16.6 % | not reported |
| Tennis, ATP/WTA tour, 70 %+ holders (window B) | 120 | +19.0 % | not reported |
These were chosen from more than 1,300 tested rules, so even the intervals that exclude zero are optimistic. The quality-net result depends on a few events: the five best gave 66–75 % of its profit. The tennis rule was selected in one window and confirmed in the other, but still fails the multiple-testing correction.
Not confirmed out of sample. In the full-year test without hindsight (above): $50k+ on the side −2.7 % a trade (641 trades); 80 %+ with $50k+ on a favourite +0.4 % (1,558, CI −2.3 to +3.0 %); 80 %+ with a side worth $100k+ −0.3 % (835); 3+ proven winners with two thirds on the side −1.9 % (49,161); tennis −3.4 % (12,026). These slices are not an edge.
What lost in both windows
| Category (5+ holders, 60 %+, $1k+ on the side) | Window A | Window B |
|---|---|---|
| Basketball | −12.0 % (CI −22 to −2 %) | −9.2 % |
| Culture | −7.1 % (CI −15 to −0.1 %) | −15.4 % (CI −35 to −3 %) |
Costs and exits
- A cent of half-spread costs about 2 points a trade; the 2026 taker fee about 1.4–1.8 % of the stake.
- Holding to resolution was best or tied for every rule. Taking profit at 0.95 cost 0.5–1.8 points a trade, switching sides on a flip 1.5–8.4.
Sources: the October 22 2026 backtest reports (core, skill, categories and realism). Nothing here is significant after correcting for the number of rules tested.
October 25, 2026: one year of today's top traders (hindsight, superseded)
Flattering by design. These are the 1,937 wallets on Polymarket's leaderboard today, replayed over the 12 months before (Nov 2025 – Oct 2026). They were chosen because they did well, so every figure below is an upper bound: on the clean data the same plain rule lost about 4 points a trade more (+0.5 % here against −3.5 % without hindsight, same window), and none of the positive rules below held up out of sample. Same method and costs as above (1¢ half-spread + taker fee, held). Never pooled with the windows above.
| Rule | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Buy the consensus side (5+ holders, 60 %+, 0.05–0.95, held) | 45,101 | +0.4 % | −0.4 to +1.3 % |
| Same, without the taker fee | 45,101 | +2.2 % | +1.3 to +3.1 % |
| Same, at a 2¢ half-spread | 45,101 | −1.4 % | −2.2 to −0.6 % |
Positive in 7 of 12 months.
| Rule | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Price 0.05 – 0.30 | 7,281 | −3.5 % | −8.2 to +1.3 % |
| Price 0.30 – 0.50 (positive in 10 of 12 months) | 13,382 | +3.1 % | +1.1 to +5.0 % |
| Price 0.70 – 0.95 | 25,129 | +0.7 % | +0.2 to +1.2 % |
| Politics and geopolitics, 70 %+ favourite, switch on a flip | 1,535 | +5.8 % | +1.7 to +9.6 % |
| Esports favourites | 5,669 | +3.5 % | +1.7 to +5.3 % |
| Tennis, ATP/WTA tour | 3,424 | +0.9 % | −2.9 to +5.0 % |
| Tennis, lower tiers | 3,703 | −4.9 % | −7.9 to −1.9 % |
| Fading basketball (the other side) | 2,192 | −6.6 % | −11.8 to −0.4 % |
The politics rule is the only one that passed the multiple-testing check on this set, including with entries a day late. The tennis tour rule that looked best in the short windows (+16 % to +19 %) is about break-even over the full year.
| Signal on the consensus side | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| 1+ holder betting at least 3× their usual size | 7,789 | +2.6 % | +0.8 to +4.4 % |
| 2+ such holders | 1,715 | +5.1 % | +1.5 to +8.6 % |
| 3+ such holders | 450 | +8.3 % | +1.8 to +14.5 % |
| The elite disagree with the crowd: buy the elite side | 3,171 | −14.8 % | −20.3 to −9.3 % |
Bigger-than-usual bets show a clear dose effect on the year (it was not readable in the short windows), but no signal passed the multiple-testing check. See the signal explorer.
| In-play rule | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Follow the consensus, sell when it fades (the first in-play bot) | 60,801 | −1.4 % | −2.0 to −0.9 % |
| Follow the consensus, hold to the result | 41,129 | +1.8 % | +0.8 to +2.7 % |
| Hold, match winners only, price ≥ 0.20, strongest consensus | 9,840 | +1.0 % | −0.4 to +2.4 % |
| Hold, tennis only | 7,141 | +1.4 % | −1.0 to +3.7 % |
| Hold, esports | 12,038 | +4.8 % | +3.0 to +6.5 % |
At a 2¢ half-spread every in-play rule falls to about zero or below. Priced with minute-by-minute order-book prices on August–October only, the held rules came out higher (+5 % to +8 %), but those are the three months the hindsight flatters most.
October 25, 2026: the first skill-weighted model (superseded)
| Data | Trades | Return a trade | 95 % CI |
|---|---|---|---|
| Spring 2026, hand-picked wallets (clean) | 1,639 | −5.6 % | −11.4 to +0.4 % |
| October 2026, every tracked wallet (provisional) | 736 | −2.5 % | −8.5 to +4.1 % |
| One year, today's top traders (hindsight) | 8,058 | +2.5 % | +0.3 to +4.5 % |
On the clean data every version predicted worse than the market price; the one-year gain was the hindsight, not the skill (it did better with the trader information removed). The full-year model above, without hindsight, confirms it.
The other figures on this site
The accuracy and ROI figures on Analytics and the homepage describe the smart money's own result at resolution: which side the wallets still holding at the end were on, and what their own entries returned. That is not what a follower earns. It leaves out everyone who sold before the end, holders who redeemed their winning shares as soon as a market settled can drop out of the count, it uses the holders' own entry prices rather than the price when you could have followed, and it includes no spread or fee. This page is the one that measures following.
How the consensus is counted, and how these studies fit together, is on the methodology page.