How we measure whales: the full method

Ranking whale wallets by how much money they make does not identify wallets worth following. We measured it two ways, and both say so. This note is the full method — the sample sizes, the definitions, and the measurement that indicted our own delivery.

Why this note exists

Every number on darktrade.ai comes from one of two places: the live API, or our receipt log. Neither is useful to you unless you can see how it was computed. This page is the reference the rest of the site links to.

We cite no external studies here. Everything below is our own measurement of Hyperliquid wallets, and where our data is thin we say so rather than rounding it into a headline.

The sample

We evaluate a universe of 1,300+ Hyperliquid wallets on a rolling 90-day window, recomputed daily. A wallet's metrics are built from its closed positions in that window, on whole capital — spot and perpetual balances together, not one position in isolation.

Whole capital is the denominator that matters and the one most copy-trading products avoid. A wallet that risks 4% of a $10M account behaves nothing like a wallet that risks 4% of the $200K it happens to have on one venue, and only the whole-capital view can tell them apart.

The two populations, and why they differ

This is the finding that reshaped our product, so it is worth stating precisely — including which ranking the wallet below actually topped. There are two different sets of trades in any copy-trading system, and they are not the same set:

• Graded — the positions our selection metrics score. To be graded, a position must be reconciled, untruncated, and held at least four hours. All three are knowable only when the position closes.
• Emittable — the positions a follower could actually have received an alert for. An alert fires when the position opens.

Selection therefore scores one population and delivery serves another. Over a 90-day backtest of five live wallets:

Twelve points of difference, from grading a non-representative slice. We grade on 52.8% of emittable positions; an emitted signal lands in the graded set only 48.1% of the time — a coin flip.

For one wallet the gap was far worse. On graded win rate — the number our own scoring ranked by — it was the best in the set:

A 34-point gap on the same wallet over the same 90 days. Worse than a coin toss on the trades a follower could take, while looking like our strongest performer on the trades our metrics happened to grade.

One caveat we owe you, because it cuts against the obvious reading: this wallet was not the most profitable one in the set. By expected return to a follower it ranked 103rd of 130, with the lowest monthly ROI of the five. It topped the set on measured win rate, not on money. The point stands either way — the ranking we trusted did not survive contact with what a follower could actually take — but "the most profitable whale is the worst to follow" is not what this particular measurement shows, and we are not going to imply it.

What we actually delivered

The honest chain does not stop at "emittable". Our public receipt log records what we broadcast and how it closed. Over 264 closed signals, the delivered win rate was 36.7%.

That is the worst number in this note and it is ours, not a whale's. It breaks into two distinct losses:

1. Selection illusion — 12 points. 68.0% → 56.0%, from grading a slice nobody receives. Described above.
2. Execution — 19 points. 56.0% → 36.7%, from a legacy fixed 1% stop at 20x leverage. Across 260 legacy signals, 64% were stopped out, winning 36.2%.

The stop rule was the larger problem and it was entirely self-inflicted. We retired it. The replacement risk model has closed too few signals to quote a win rate for, so this note does not quote one — the receipts accumulate in public and you can watch the number form.

Why profitability does not predict a good whale to follow

The intuitive metric is whale profit. We tested it and it does not survive.

We built an expected-user-return model — how much a follower earns per position, given the whale's returns and how tightly the position can be sized against its worst drawdown. Correlated against what a copier actually captures per position, across all 130 measured wallets:

correlation = −0.06

Statistically nothing. Five of the top twenty wallets by expected return had negative realized per-position expectancy.

The mechanism is sizing, not skill. Two wallets from the same sample:

Wallet A wins half as often and makes a third as much, and is worth 2.8× more to someone following it — because the worst case a follower has to size against is 8.7× tighter, so identical risk buys 8.7× the position. Win rate does not enter that calculation at all.

("Worst case we size against" is the larger of the wallet's worst per-position drawdown and its worst stop distance, measured on whole capital. It is not drawdown alone, and the distinction matters: for some wallets the stop is the binding number.)

This is why the site ranks by risk colour rather than by profit, and why the leaderboard's headline number is return on capital rather than dollars earned.

The five filters

A wallet must clear all five to be listed. These are the thresholds the engine actually runs — no discretionary admissions:

1. At least 50 measured positions. Under 50, a lucky streak and a real edge are indistinguishable.
2. Win rate 60% or better, after fees and funding, on positions a follower could have taken. Win rate is the one trait we measured that genuinely persists between independent halves of the sample (rank correlation +0.66, versus +0.34 for expectancy) — the weight follows the measurement, not the intuition.
3. Max concurrent drawdown under 50% of whole capital. The worst simultaneous open loss the account ever carried, against the entire account.
4. Median hold of at least 2 hours. A copyability rule, not a quality one. One wallet we tested ranked #2 of 20 on our approved ranking metric (expectancy × multiplier) and closed 71.3% of its positions before a follower could act. Excellent, and un-followable — and a ranking metric cannot see this, which is exactly why it is a gate.
5. All-time profit positive. A sign test, never a threshold.

Passing once is not a permanent seat: the set is re-scored daily and a better wallet takes a worse one's place. Nobody is deleted — a demoted wallet keeps its page and its receipts and can climb back.

The colour system

Colour comes from one number: the wallet's maximum drawdown per position on whole capital. The bands are published on every page that shows a colour, and the thresholds are served from our own config rather than written into the page, so the words cannot drift from the numbers.

A colour is a statement about risk taken, not about returns coming. A white whale risks less of its account per position than a red one. That is the entire claim. It does not make any given week profitable, which is why cohort performance is published separately with the number of closed positions behind every percentage.

What we do not know

• Your result. Every profit figure on this site is what the whale realized on its own capital. We do not know your position size or when you acted, so we never state what you would have made.
• Peak versus realized. These are separate fields everywhere, always labelled. A peak is what a position reached at its best moment; realized is what closed. Substituting one for the other is the most common way a track record gets inflated, and we treat it as a correctness bug.
• Regimes. The 90-day window covers one market regime. No walk-forward test across regimes exists yet. Every threshold above was derived on that window and may not survive a different one.
• Sample size, honestly. Our public receipt log began 2026-08-12, and only closes from 2026-08-10 onward are ever shown. Some cohorts have very few closed positions. Where that is true, the site says so on the number rather than in a footnote.

Checking us

Every wallet behind these numbers is on the leaderboard with its own profile, its own receipts, and its address. The addresses are real and public — Hyperliquid's data is on-chain, and nothing stops you recomputing any of this yourself.