← Research

NOARA RESEARCH · 2026-07-27 · 20 DAYS OF DATA · ALL SOURCES PUBLIC

Big Money Isn't a Signal: 109,000 Whale Prints on Polymarket, Audited

What we did: we run a public-API collector that has been recording every Polymarket trade above $2,000 since late June 2026, plus a rolling snapshot of holder concentration on the markets those trades touch. This report takes 20 days of that feed — 109,460 whale prints, $880.8M in notional, 11,636 markets, 10,535 distinct wallets — profiles the 300 largest wallets by reconstructing their full trade ledgers from public endpoints, and then asks the only question that matters to anyone watching a whale-alert bot:

If you had copied those trades, what would have happened to your money?

Every number below is reproducible from the artifacts listed in §8. Where the data refuses to support a clean answer, we say so instead of rounding it into one.

1. Executive summary

1. We found the copiers on-chain, and all of them lost money. Pulling every trade on the 500 whale-densest markets (1.94M trades) and matching them against 48,338 whale buy prints: 182,355 buys landed within 60 seconds of a whale, at a median ticket of $10 and a median lag of 35 seconds. Every follower cohort finished negative after fees — −6.6% at one minute, −14.1% at one hour. We have not seen this measured anywhere public, though we make no claim to have searched exhaustively; the raw material for it is entirely public.

1. The cost of copying is delay, not slippage. Followers paid the whale's price plus half a cent — the fill is not the problem. But against a price-matched control group of buyers who weren't in any whale window, one-minute followers do +1.6 points better, five-minute followers −5.7, and one-hour followers −12.2. Reacting instantly is survivable. Seeing an alert, thinking about it, and buying ten minutes later is what costs money.

1. Copying is negative even with a time machine. Give a copier the whale's own fill price with zero latency and infinite depth, across all 76,069 resolved whale buy prints: they win 69.5% of their bets and still finish at −3.11% ± 0.63% per dollar after fees. Market- clustered: −1.22% ± 0.99%. High hit rate, negative expectancy — the odds do the killing.

1. The average "whale" wins about half its events. Public leaderboards show ~86%. The gap is one mechanism: losers are never closed. Counting a position whose price has collapsed below $0.02 as the loss it already is, the 280 profiled wallets show a median true win rate of 47.3% against a median displayed win rate of 90.9%. 84% of wallets are flattered by this, and 102 of 278 are flattered by 50 percentage points or more. On the subsample where we can see complete trade history (n=56), the numbers are 55.6% true vs 74.1% displayed — an independent replication of PANews's January 2026 audit (53.8% vs 73.6%) on a different, larger, more recent sample.

1. Only 2 of the 300 biggest wallets look like actual forecasters. By behaviour, the money is 35% structural arbitrage, 23% near-certainty yield farming, 18% position management, 17% directional gamblers — and 0.7% pure predictors, who account for 0.5% of the notional. When you get a whale alert, the prior is overwhelmingly that you are watching an arb leg or a yield trade with no directional content at all.

1. Where the losses concentrate is exactly where retail looks. Following buys priced 0.20–0.50 returned −9.05% ± 2.05%; below 0.20, −15.07% (wide error bars). Following buys above 0.90 — the boring yield trades nobody screenshots — returned +0.22% ± 0.21%. The trades that look like an opportunity are the ones that cost money.

1. The mirror trade is not free money either. Taking the other side of every whale buy returns −8.61% ± 4.26%. Both sides of the fee-and-spread structure lose. "Fade the whales" is not the fix for "follow the whales."

1. We tested this with our own money-shaped experiment and lost. Our own smart-money paper tracker — real prices, real signals, real settlement, simulated capital — went −20.3% over 55 settled positions and we shut down the live bridge feeding it. Details and the full ledger in §6, because a report about copy-trading traps that hides the author's own is worthless.

2. Method

2.1 Where the data comes from

Everything is public and read-only. No account data, no order flow of ours, no paid feeds.

2.2 True win rate vs displayed win rate

This is the single most important definition in the report.

A Polymarket position that goes to zero does not have to be closed. It can simply be left in the wallet forever. If you compute a wallet's record only from realised P&L — which is what a leaderboard does — those positions never enter the denominator, and the record shows only the trades that were closed, which are overwhelmingly the winners.

• Displayed win rate — events with realised P&L only. This is the flattering number.
• True win rate — same events, plus every position currently priced ≤ $0.02 counted as the loss it already is, and every position priced ≥ $0.98 counted as the win it already is.
• Zombie gap = displayed − true. This is not noise; it is a behavioural fingerprint. A wallet that never leaves dead positions lying around is a wallet that closes its books.

Event-level adjudication uses on-chain resolution where available (Gamma closed=true), redemption records, and current price for positions that are decided but unredeemed. Anything still genuinely open counts in neither numerator nor denominator.

2.3 Behavioural classification

Wallets are typed by rules anchored to the profiles PANews documented, not fitted to P&L — the thresholds were fixed before this sample was collected and were not tuned for these results:

Hedging detection covers both same-market two-sided buying and the "buy every line in one event" pattern across the separate condition IDs of a multi-outcome event.

2.4 The copy-trade simulation

For every whale buy print whose outcome has since resolved, we compute the return of buying that outcome at the whale's own fill price:

gross per $1 = (settlement − price) / price
fee  per $1  = 0.07 · (1 − price)        # taker-only, rate·p·(1−p) per share
net  per $1  = gross − fee

Three deliberate choices, all of which flatter copying:

• Zero latency, zero slippage, unlimited depth. A real copier sees the print after the fact and pays worse. This is an upper bound on copy-trade performance, not an estimate of it.
• Fees only, no spread. Crossing the book costs more than the fee alone.
• Unresolved markets are dropped, never assumed. 86.1% of buy prints are adjudicable; the rest are still open and simply absent.

We report equal-weighted (each decision counts once — the investor's view), notional-weighted (the money's view), and market-clustered (one market counts once — the conservative statistical view, since prints in the same market share a single outcome and are not independent draws).

3. What the big wallets actually are

300 wallets, ranked by cumulative notional in the feed, covering 60.9% of all whale-print notional. Medians within each type:

Three things worth sitting with:

The largest single block of whale money has a 25% true win rate — and that's fine. Structural arbitrageurs are supposed to lose most of their individual legs; they are buying both sides and harvesting the spread. Their "buy" prints are not opinions. Nearly half of all whale notional in this sample is this. An alert bot cannot tell the difference, so half of what it shows you is noise by construction.

Farmers have the best true win rate in the dataset (83%) and the least useful signal. They buy at $0.90+ and collect the last few cents. Copying them at their price earns roughly the same few cents minus fees — see §4 — and no amount of hit rate turns that into an edge.

Pure predictors exist but are rare and quiet. Two wallets out of 300, trading 1.6 times a day at an average ticket of $7,579, with a win/loss size ratio of 9.5. Their combined footprint is 0.5% of whale notional. If a wallet is printing 200 times a day, that alone rules it out of this category.

3.1 Leaving dead positions lying around predicts being bad at this

Split the profiled wallets by whether they carry any zombie gap at all:

A 26-point spread in true win rate, from a variable you can compute for free from the positions endpoint without knowing anything about the trader. The caveat is real and we state it plainly: the clean-book group has far fewer adjudicated events per wallet (median 32 vs 127), so part of that spread is small-sample luck rather than skill. It replicates a pattern we saw in an earlier scan on a different sample, which is why we report it — but it is an observation, not a validated selection rule.

4. The people who actually copied these trades

Everything above this point is inference about whales. This section is about the people watching them — reconstructed from the chain, not modelled.

4.1 How you can see a copier

Whale-alert users are invisible in aggregate statistics, but they are not invisible on-chain. If a $40,000 buy prints on an outcome at 14:02:10 and 180 addresses buy the same outcome in the next sixty seconds with a median ticket of $10, those addresses are, at minimum, acting immediately after the whale. So we pulled every trade — not just large ones — on the 500 markets where whale activity is densest: 1,944,449 trades across 1,000 outcome tokens, covering 68.7% of all whale notional in the feed. Then, for each of the 48,338 whale buy prints landing inside that tape, we collected every other address that bought the same outcome within 1, 5, 15 and 60 minutes.

Two things this method is not:

• It is not proof of causation. Someone buying 40 seconds after a whale may be reacting to the whale, or to the same goal / injury / headline the whale reacted to. We can measure what happened to them; we cannot prove why they clicked.
• It is not usable without a control group. "Copiers lose money" means nothing if everyone buying in these markets loses money. So every number below is compared against a control: buyers of the same outcomes who were not inside any whale window, with whale addresses themselves excluded from both groups.

4.2 Who they are

The median follower ticket is $10. These are not funds mirroring each other; this is retail, clicking within about half a minute of a print, in size that would not cover the gas on most chains.

The most surprising number here is the smallest: followers pay only half a cent more than the whale did, and only about half of them pay worse at all. The intuition that copiers get destroyed by slippage is wrong — at this size the book absorbs them. Whatever it costs to copy, it is not being spent on the fill.

4.3 What happened to their money

Comparing raw averages between followers and controls would be invalid: whale windows sit on expensive outcomes (mean entry 0.653 for one-minute followers vs 0.426 for controls) and returns depend heavily on entry price. So both groups are bucketed into 29 price bands (0.02 wide below 0.30, 0.05 above) and the control is re-weighted to the followers' price mix. After that alignment, the largest within-bucket price difference between the two groups is 0.005 — the comparison is clean.

Read the equal-weighted column first — with a median ticket of $10, an equally-weighted decision is what a retail follower actually experiences.

Every follower cohort loses money in absolute terms: −6.6% to −14.1% after fees. That is the headline finding of this report and it is not close.

But the comparison is more interesting than the level, and it does not say what a copy-trading sceptic would expect:

• Followers who act within 60 seconds do no worse than comparable buyers (+1.61%, i.e. slightly better). The immediate reaction is not the mistake.
• The penalty grows monotonically with delay. Five minutes: −5.7 points. Fifteen: −10.2. An hour: −12.2. The people who see an alert, think about it, and buy ten minutes later are the ones paying.
• Notional-weighted, the gap mostly disappears (+7.9% to −0.3%). Large tickets inside whale windows do fine; small ones do badly. The damage is concentrated in exactly the retail-sized clicks that make up the median.

The honest summary: copying is not a fast way to lose money because whales are wrong or because fills are bad — it is a slow one, driven by acting late and small in markets where all retail buyers are already losing (the control group loses 2–8% too, depending on what it buys).

4.4 The same test, idealised

The follower analysis above covers 500 markets. Widening to every whale buy print in the feed and asking a different question — what if you could fill instantly at the whale's own price, with no delay and no slippage at all? — gives the theoretical ceiling of copy-trading:

76,069 resolved whale buy prints, $634M of notional, filled at the whale's own price:

The gross number is the interesting one: −0.98%. Whale prints are, on average, priced almost fairly. There is no large systematic mispricing to harvest by following them — and once you pay to participate, the small negative becomes a reliable one. Copying does not fail because whales are wrong. It fails because they are roughly right, and being roughly right is not worth the toll.

4.5 Where the idealised version fails hardest

The monotonicity is the finding. The cheaper and more exciting the ticket, the worse copying does. The one slice that is positive after fees is the one that looks like nothing: buying at $0.90+ for a handful of basis points, which is a yield trade requiring size and patience, not a signal.

Bigger prints are not worse to follow — if anything the reverse — but the two largest buckets have error bars that swallow their point estimates. We cannot tell you that following $100k+ prints works. n=457 and ±7.53% means we cannot tell you anything about that bucket.

4.6 Sorting by wallet type does not rescue it

This is the part of the report we most wanted to come out differently. Our own thesis going in was that size carries no information but behaviour type does. On this test, it doesn't:

Every type is negative equal-weighted. Two types flip positive when weighted by notional, driven by a handful of large tickets — that is variance, not an edge.

On pure predictors specifically, we decline to answer. Two qualifying wallets produced 33 adjudicable prints in this window with a standard error of ±29.5%. That is not a result in either direction. The honest statement is: the wallets whose behaviour suggests genuine forecasting are so rare and so inactive that 20 days of data cannot evaluate copying them. Anyone claiming otherwise from a sample this size is guessing.

4.7 Fading them doesn't work either

Taking the opposite side of every whale buy — buying the complementary outcome at (1 − p) — returns −8.61% ± 4.26% after fees. The structure is not zero-sum for participants: both directions pay the toll, and the mirror of a fairly-priced trade is a fairly-priced trade minus another fee.

5. Concentration: who is actually holding these positions

12,988 unique (market, outcome) tickets snapshotted across the window, 11,500 of them since resolved.

5.1 A data trap worth knowing about

4.8% of the tickets we sampled show a "largest holder" that is not a trader at all. One address appears as the top holder on 620 tickets across 1,086 markets, holding identical enormous balances across every outcome of the same event — including every 2024 presidential candidate simultaneously, at an initial value of $0. Its activity log contains no trades of any kind, only protocol yield entries, and it has never once appeared in 111,000 whale prints.

This is protocol plumbing: depositing collateral into the conditional-token contract mints one share of every outcome, and the resulting balances sit in a vault address that the public /holders endpoint reports alongside real traders. Any tool that reads that endpoint and announces "one wallet holds 99% of this market" without filtering it is reporting on a smart contract.

We detect these two ways — identical balances across the outcomes of one market, and "appears constantly as top holder but has literally zero trades" — and exclude them below. Removing them moves the headline concentration numbers by 1–2 points, which is small, but the extreme tail is where it matters: tickets showing a ≥99% top holder drop from 2.1% to 1.4% of the sample.

5.2 Concentration is real, and it is not a signal

After excluding protocol addresses (12,368 tickets):

• Median top-holder share: 35.1% of the top-20 holdings
• 29.8% of tickets have a single holder above 50%
• 6.0% above 90%, 1.4% above 99%
• Median top-5 share: 77.7%

Note the denominator carefully: /holders returns the top 20 addresses, so these are shares within the top 20, not of the entire market. They describe how lopsided the visible top of the book is, not what fraction of all outstanding shares one person owns.

Does buying the outcome that a dominant holder has cornered work? We bucket every resolved ticket by its first concentration snapshot — deliberately not the last, because near resolution the losing side clears out and the winning side stays, which manufactures a correlation that has nothing to do with foresight — and compare the realised settlement rate to the volume-weighted price whales paid:

Every bucket is negative: across all concentration levels, the outcome whales were buying settled less often than the price they paid implied. Concentration does not flip the sign. There is a mild gradient — the most concentrated bucket is least bad — but at ±4.2% it is not something to trade on, and it is equally consistent with concentrated markets simply being closer to decided.

For the extreme cases specifically (top holder ≥95%, 441 resolved tickets): they settle "yes" 58.3% of the time at an average whale price of 0.706. Cornering a market does not make the corner right.

6. What happened when we tried this ourselves

We are not observers here. We built a copy-trading tracker, ran it forward on live signals, and it lost money.

Design: simulated capital, everything else real. Entry price is the whale's actual fill VWAP plus one-sided cost — we never assume a perfect fill. Signals require ≥$5,000 net flow from ≥3 independent wallets, must be under 30 minutes old, and near-certainty farming prices are excluded. Settlement is on-chain; unresolvable positions stay pending rather than being guessed.

Result, 2026-07-13 → 07-23, 55 settled positions:

The same shape as the market-wide result: won more often than not, lost money anyway.

The part that hurt most: the stronger the consensus, the worse the outcome.

Every filter we would intuitively have tightened — more wallets, more one-sided, higher conviction — selected harder for losses. The worst single position was a World Cup semi-final line with 69 independent whale wallets behind it, one-sided, maximum conviction: −$118.92 on a $118.92 stake.

We had pre-registered a kill rule before this experiment started (kill at n≥30 if returns after costs were ≤0). We executed it: the bridge that mirrored these signals toward live orders was switched off on 2026-07-24 and has not been switched back on. The paper track keeps running because we would rather keep collecting evidence than stop looking.

Sample caveats, stated rather than buried: 55 positions is small, 50 of them are sports, and the window is a World Cup fortnight. This is a consistent story alongside the 76,069-print market-wide result, not independent confirmation of it.

7. What this means if you trade

Stated as plainly as we can, with the confidence each one actually has:

1. A whale-alert notification, on its own, is close to worthless — and slightly worse than worthless after fees. (Strong: 76,069 prints, negative on every weighting.)
2. Be most suspicious of the alerts that look most attractive. Big buy, mid-range price, lots of room to run — that's the −9% bucket. (Strong: n=15,598, ±2.05%.)
3. Ignore win-rate leaderboards entirely unless they tell you how they treat open losing positions. A wallet showing 90%+ is usually showing you its closed trades only. Check the positions endpoint for how much dead weight it is carrying. (Strong: 84% of wallets flattered, median 90.9% displayed vs 47.3% true.)
4. Frequency is a fast filter. Wallets whose behaviour is consistent with genuine forecasting trade about 1.6 times a day. Hundreds of prints a day means arbitrage, market making, or farming — whatever it is, it isn't a forecast. (Strong as a description of who's who; unproven as a trading rule.)
5. "Fade the whales" is not the answer to "follow the whales." Both sides are negative after costs. (Strong: −8.61%, though with wide error bars.)
6. Concentration tells you a market is lopsided, not which way it will go. And check whether the "whale" holding 99% is a smart contract. (Strong.)
7. We do not know whether following genuine forecasters works. Two candidate wallets, 33 prints, ±29.5%. Anyone who tells you they've proven this either way on public data from a window this short is overreaching. (Explicitly unresolved.)

8. Limitations

We would rather you discount this correctly than trust it wrongly.

• 20 days, one season. 2026-07-06 to 07-27. Sports and esports are 92% of the $634M we analysed in §4 ($515.5M + $68.4M) because the window sits on a World Cup. Politics is 91 prints and $0.6M. Nothing here should be extrapolated to an election cycle, which is exactly when whale-following gets most attention.
• Resolved-only. 86.1% of buy prints are adjudicable; long-dated markets are structurally underrepresented because they haven't settled yet. If whale skill lives in year-out contracts, this design cannot see it.
• Trade history is windowed. Per-wallet reconstruction pulls the most recent ~1,500 activity records. For 224 of 300 wallets that window is full, meaning older redemptions are cut off while their still-open dead positions remain fully visible. This biases true win rates downward. It is why we report the complete-history subsample (n=56: 55.6% true / 74.1% displayed) separately and treat it as the better estimate; the full-sample figure (47.3% median true) is a floor, not a point estimate.
• "Follower" means "bought right after", not "copied". §4.1–4.3 identify people by timing, not intent. Someone buying 40 seconds after a whale may have been reacting to the same headline. The control-group design removes the "everyone in these markets loses" explanation; it does not establish that the whale print caused the follower's click.
• Equal-weighted and notional-weighted follower results point different ways. Equal-weighted, the one-hour cohort trails its control by 12.2 points; notional-weighted, by 0.3. Small tickets drive the damage. We lead with equal weighting because the median follower ticket is $10, but a reader who cares about aggregate dollars should read the other column.
• The tape covers 500 markets, not the whole exchange (68.7% of whale notional, 1,000 outcome tokens). 159 of those markets hit our 6,000-trade fetch cap, so on the very busiest markets we hold the most recent trades rather than all of them.
• Idealised copy returns are an upper bound. Zero latency, zero slippage, infinite depth, fees but no spread. Real copying is worse than every number in §4.4–4.7 — as §4.3 shows directly.
• Fee model is a ceiling. We use rate·p·(1−p) at rate = 0.07, taker-only. Actual costs vary.
• Correlated observations. Prints in one market share one outcome. We report market-clustered standard errors for the headline; the per-print error bars are optimistic by construction.
• Concentration is top-20-relative, sampled only on markets with whale activity, and only for markets our collector saw. It is not a census of Polymarket.
• We cannot see intent. A wallet buying both sides is classified as an arb leg. It might be a forecaster changing their mind. Behaviour is observable; motive isn't.
• Survivorship in the wallet sample. We profiled the 300 largest wallets by activity in this window. Whales who blew up before 2026-07-06 or who trade below $2,000 per clip are invisible here.

9. Reproducing this

All figures come from five artifacts, each stamped with its own generation time (the underlying feed is live and grows, so row counts differ slightly between artifacts):

The collector and the analysis tooling are ours; the inputs are entirely public endpoints, so the whole thing can be rebuilt from scratch by anyone with the same three API calls and enough patience to wait 20 days.