Hyperliquid leaderboard explained: why most rankings are misleading

How the public leaderboard is computed

Hyperliquid's data is public, so a leaderboard is a straightforward aggregation: take accounts, compute profit over a window, sort descending. Most public boards, including third-party ones, offer a handful of windows — a day, a week, a month, all time — and let you sort by absolute PnL or by return.

Two details do most of the work in shaping what you see. The window determines how much of an account's history is being summarised, and the metric determines what 'best' means. Absolute PnL favours large accounts almost mechanically, because a 2% move on $50m outranks a 60% move on $200k. Return favours small accounts for the mirror-image reason.

There is usually also a scope question: whether the figure includes unrealised PnL on open positions. When it does, an account can sit at the top of the board holding a large open position that has not been closed, and the ranking is reporting a mark-to-market number that may never be realised.

None of this is a criticism of the exchange. The board is an accurate report of what it says it reports. The problem is what people do with it.

Leverage distortion: same return, different risk

The single largest distortion is that returns are reported without the risk that produced them. Two accounts can both return 40% over a month, one by running 2x on liquid majors and one by running 20x on alt perps. The board shows one number for both.

In leveraged instruments, this is not a nuance. The second account's distribution of outcomes includes liquidation with meaningful probability; the first account's does not. Ranking them adjacently implies a comparability that does not exist.

The asymmetry matters because the board only ever shows you the good draw. A strategy with a 90% chance of returning 40% and a 10% chance of returning -100% will, in a large enough population, put plenty of accounts near the top every month — different accounts each time, and each of them looks excellent in the window where they appear.

Short windows and variance

A weekly ranking of leveraged perpetual futures accounts is, statistically, mostly noise. There are not enough closed trades in a week for the ordering to reflect process rather than which direction happened to work.

This produces a specific, expensive pattern. A trader is at the top this week because their bias matched the week; capital follows them in; the following week's conditions differ and the same bias underperforms. The copier's entry is systematically at the point of maximum recent performance, which is the worst available entry if any mean reversion in performance exists.

Longer windows help, but not as much as people expect, because they trade one problem for another: a strong twelve-month record can be entirely composed of one exceptional quarter followed by nine months of decay, and the sorted column cannot tell you which.

Deposits, withdrawals and the ROI denominator

Return figures need a denominator, and on a live trading account the denominator moves for reasons that have nothing to do with trading.

A trader who withdraws half their balance halves the base against which subsequent returns are measured, so identical trading produces roughly double the percentage. A trader who deposits substantially dilutes their own reported return. Neither has changed anything about their process.

This is not usually deliberate manipulation, though it can be. Mostly it is just noise that the ranking presents as signal. It is one reason to prefer measures computed from closed trades — profit factor, payoff ratio, distribution of realised PnL — over anything with account equity in the denominator.

One-market outcomes read as skill

A large share of top-of-board results come from a single market making a large move while an account happened to be positioned in it with size. The board cannot distinguish that from a diversified process that performed consistently across many positions.

The test is trivial and almost never applied: remove the largest single trade and recompute. Records that collapse under that test are describing an event. Records that survive it are describing a method. Both look identical in the sorted column.

Survivorship bias: the accounts you never see

Every leaderboard is a survivorship-filtered sample, and this is the distortion that most changes the interpretation of everything else.

Accounts that were liquidated do not appear. Accounts that stopped trading after a large loss do not appear. What remains is the upper tail of a distribution whose lower tail has been deleted, and the visible tail looks like evidence of widespread skill because the counter-evidence has been removed from view.

This has a practical consequence for anyone estimating what copy trading can produce. Inferring an expected outcome from the accounts on the board is inferring from a sample selected on the outcome you are trying to estimate. The estimate will be too high, and the error is not small.

The correction is not available from the board itself, because the missing accounts are missing. The best available substitute is to weight survivability heavily as a selection input — drawdown depth, recovery, liquidation history, account age — so that the accounts you choose are the ones whose survival is explained by their behaviour rather than assumed from their presence.

What a useful ranking would optimise for

A ranking intended for allocation, rather than for display, should optimise for repeatability. That means measuring the properties that persist rather than the outcome that happened.

It should use closed trades rather than account equity, so the denominator cannot be moved by transfers. It should require a minimum sample before ranking an account at all. It should read win rate and payoff ratio together, since either alone is trivially gameable. It should penalise sizing escalation and leverage spikes rather than ignoring them because they did not blow up this time. And it should treat survivability as a first-class input, not as a footnote.

The output of such a ranking will look boring next to a PnL board. Accounts with spectacular recent returns will often be absent, because their sample is too small or their sizing is unstable. That absence is the ranking working.

Composite scoring versus raw performance

A composite score combines the properties above into one comparable figure: realized pnl consistency, win rate, profit factor, position discipline, account survivability. Each has a floor a candidate must clear independently, because averaging lets a strong factor hide the weak one that ends accounts.

The score is then used for two things, not one. It decides who is in the basket, and it decides how much capital each qualified leader receives — allocation proportional to score rather than split evenly, because equal weighting asserts that a leader who barely cleared the floors is as convincing as one who cleared them comfortably.

The honest limitation is that a composite score is still computed from history. It is more robust than a PnL rank because its inputs are more persistent, but it cannot see a regime turn before the closed trades reflect one. Scoring reduces the number of ways you can be wrong; it does not remove them.

Using the leaderboard intelligently

The board is a reasonable candidate generator and a bad final answer. Used deliberately, it earns its place in a process.

• Treat the board as a shortlist to investigate, never as a ranking to act on.
• Pull the longest window available and compare it against the shortest. Divergence is informative.
• For each candidate, count closed trades before reading any derived statistic.
• Remove the largest winning trade and recompute. Discard anything that collapses.
• Chart position size against equity over time and look for escalation after losses.
• Check explicitly for liquidations and for account age.
• Ask which regime produced the record and whether it currently holds.
• Assume the accounts you cannot see failed, and let that discipline your position sizing.

What scoring still cannot tell you

A better ranking is still a ranking of the past. It cannot see intent, hedges held elsewhere, or whether the person behind an address is about to change what they do. It cannot tell you that a strategy is about to stop working, only that it has stopped.

Diversifying across several scored leaders reduces dependence on any one of those unknowns being benign. It does not remove market risk: in a correlated deleveraging event, well-scored leaders fall together, because they are trading the same market with the same liquidity and the same funding.

Perpetual futures are leveraged instruments and carry a substantial risk of loss, including the loss of your entire position. Past performance is not indicative of future results.

Conclusion

A leaderboard ranks outcomes. Allocation requires a ranking of process, and the two lists overlap far less than the interface suggests.

The basket we run publishes each leader's composite score and weight alongside the raw figures, so the two orderings can be compared directly rather than taken on trust.