Seven ways copying a wallet goes wrong

Following another address is the most intuitive thing you can do with on-chain data, and it is intuitive precisely because it looks like it removes the hard part. Someone else did the analysis; you copy the position; their edge becomes your edge.

It does not, and the reasons are specific rather than philosophical. Each one below is a concrete mechanism that puts distance between the position you copied and the position you now hold. Most of them are fixable. None of them fixes itself.

1. The track record was selected after the fact

The wallet you are watching came to your attention because it was right about something. That is the selection, and it happened before you started measuring.

On our own store of 145,141 wallets, 36,354 carry a win rate of exactly 100% — one address in four. Of those, 28,710 have traded three markets or fewer, and 18,914 have traded exactly one. The column as a whole barely has a middle: 126,582 of the 145,141 sit at exactly 0% or exactly 100%. (Measured 17 August 2026 at 10:30 UTC.) Pick a wallet off that list after the fact and you have picked a coin that came up heads twice.

The fix is unglamorous: decide the filter before you look at the outcomes, then apply it to everyone. Distinct markets, not trades. A statistic, not a percentage. A window with dates on it. Require twenty-five distinct markets before a wallet is readable at all and about 2,700 addresses qualify — 1.87% of the more than 140,000 that appear in our trade record. Their median win rate is 42.41%, and 1.13% of them are still at 100%, which tells you the filter is a floor and not a proof. What "smart money" means, and what it does not works through the version we apply.

2. An address is not an actor

The unit you can observe is an address. The unit that makes decisions is a person or a desk, and the mapping between them is not one-to-one in either direction.

One actor can run many addresses — routinely does, for exactly the reason that being watched is expensive. Many actors can share one address: a custodial wallet, a fund, a bot operated on behalf of several people. Nothing on-chain distinguishes these cases, and no clustering exists in our data to fix it — not a table, not a column, not a tag. The tags array is empty on all 145,141 wallets, and the only label we carry is a size bucket: 132,544 fish, 9,335 sharks, 3,262 whales, measured 17 August 2026 at 10:30 UTC. That is a fact about balances, not about identity.

So when you follow an address, you are following a channel, and the channel may be one leg of a strategy whose other legs run somewhere you cannot see.

3. You are seeing one leg

This is the same problem in its most expensive form.

A large buy of YES on one market is a directional position, or the hedge of an opposite exposure held elsewhere, or the offsetting side of a position in a correlated market, or inventory taken on by someone making prices and about to lay it off. From the tape, these are identical. From the trader's perspective they are opposite in meaning.

Copying the visible leg of a hedged position leaves you holding the risk they were removing. There is no observation that resolves this, which is why the honest use of wallet data is as a prompt to look at a market, not as a position.

4. You pay a different price than they did

Two costs separate their fill from yours.

The first is the spread, and you cross it twice. Buying at the ask and later selling at the bid costs the whole distance between them. Across more than 2,700 order-book snapshots taken between 15 and 17 August 2026 — the ones with a two-sided book and a mid between 0.05 and 0.95, covering 14 markets — the median spread is one cent, and the median spread measured against its own mid price is 2.50%, with quartiles at 1.34% and 4.88% (measured 17 August 2026 at 10:30 UTC). Against an edge of a few points, that is the same order of magnitude as the entire thing you are trying to capture, and you start paying it on the way in whether or not the copied trade was any good.

The second cost is caused by the trade you are copying. Their order consumed the resting liquidity; yours arrives into a thinner book at a worse price, and every other follower is arriving at the same time. The larger and more visible the print, the worse this gets — which means the trades most likely to catch your attention are the ones with the worst follow-on fills.

5. Their size is not your size

A position that is 2% of their balance may be 40% of yours. The copied trade has the same probability and the same payoff; it does not have the same consequence.

Binary payoffs are high-variance by construction. A position that resolves in your favour 60% of the time loses four in a row 2.56% of the time — once in every thirty-nine runs of four — and the only defence is a stake small enough that it does not matter.

The Kelly fraction for a binary contract, (p - c) / (1 - c), does not behave the way intuition suggests. The same three-point edge asks for 6% of the bankroll at a price of 0.50 and 30% of it at 0.90, because the higher price also carries the higher chance of being paid. That is the number the formula gives, not the number to stake: at 0.90 the estimate is the fragile part, since three points of error there is the difference between losing one time in ten and losing one time in fourteen. A position sizer gives you the number; take a fraction of it.

And the input p is your estimate, not theirs. Copying a position does not copy the conviction behind it.

6. Their horizon is not your horizon

Capital in a prediction-market position is locked until resolution. On our store, about 82,000 markets are active, unresolved and carry an end date still in the future; the median one ends in under a week, but 26.74% run past a month and 4.14% past six months (measured 17 August 2026 at 10:34 UTC). A market at the long end of that distribution holds your money for half a year earning nothing, and against a risk-free alternative a small positive edge over that horizon can be a negative use of capital even when it resolves your way.

The wallet you are copying may be indifferent to this — running a book where that capital is collateral for something else, or holding a position whose horizon was chosen deliberately. You inherit the horizon without inheriting the reason for it.

Relatedly: you do not know how long they have already held. A print you see today may be the tail of an accumulation that started weeks ago at much better prices, in which case you are entering the same trade with none of the cushion.

7. You copy the entries and never the exits

Entries are conspicuous. Exits are quiet, gradual, and often executed as passive liquidity rather than as prints that trip a threshold.

The asymmetry is structural: the shape of a build-up is legible on the tape, and the shape of an unwind frequently is not. Five ways a large position gets built, and what each leaves on the tape covers both sides and is explicit about which are and are not distinguishable.

The practical result is that you hold past the point where they left, because nothing on the tape told you they were leaving. You will find out that they left when the price tells you.

If you are going to do it anyway

There is a version of this that is defensible, and it looks like research rather than mirroring.

• —Use the activity as a filter, not as a signal. A tracked address entering a market is a reason to read that market's resolution rules. It is not a reason to hold it.
• —Form your own probability before you look at the price. If your number and the price disagree by less than the round-trip cost, there is no trade regardless of who else is in it.
• —Decide the exit at entry. You will not be told when they leave.
• —Size from your own balance and your own estimate, and stay under what the formula says.
• —Log it. Tag the trades you took because of someone else's activity and compare that bucket against the rest of your record after fifty of them. That is the only test that settles the question for you, and it is the same construction described in why most prediction-market traders lose money.

The mechanics of the instrument underneath all of this — bounded price, known payoff, scheduled end — are in what a prediction-market order book does that a crypto book does not, and the general question of telling a good forecast from a lucky one is in forecast calibration.