How the liquidation map is built

Hyperliquid positions are public, which is why maps of it get called exact. The catch is how they are public: you can read any account’s positions, but there is no list of all accounts. A map is only as complete as its builder’s address book, and ours is not complete. This page says how incomplete.

What we measured before building it

• We read all 45,138 accounts on the venue’s public leaderboard on 2026-09-17. It took 80 minutes at the public rate limit, with no errors.
• Those accounts held about 44% of Bitcoin long open interest and 41% of short. The figure was nearly identical on every large market.
• Account size does not predict positions: the 400 largest accounts held 10% of Bitcoin longs.
• In a 90-second sample of the Bitcoin trade tape, 68% of the accounts trading were not on the leaderboard, one of them holding $10.9 million.
• The best-known “exact” map of this venue plotted about 36% of Bitcoin open interest on the day we checked, and stated no coverage figure.

The rules

• The exchange’s own liquidation price, never ours. Where the venue reports none, the position is counted in the coverage table and not drawn. We do not estimate one.
• Coverage is printed on every map, reconciled to the venue’s open interest: plotted, beyond the ±60% axis, no liquidation price, and in accounts we cannot see.
• A coin is withheld when readable accounts hold under 20% of open interest on either side, or when under 10% of open interest lands inside the chart. Withheld coins are listed, not hidden.
• It is a rolling snapshot. Accounts known to hold positions are re-read about every hour; everyone else more slowly, which is how new positions are found. Reads older than 6 hours are dropped, and each map prints the age of its reads.
• The address book grows. Every refresh reads the public trade tape of the largest markets and enrols accounts it has not seen, so coverage should rise over time. The counts above show where accounts came from.
• Bars are coloured by the positions’ actual leverage, as reported by the venue.
• Every open position counts, whatever its age. Maps that offer a “lookback” window count only recently-opened positions, which is a smaller number under the same name. The only age rule here is on our own reads.
• No forecast. We say where the largest cluster is. We do not say price is drawn to it, because we have not seen that measured — so we are measuring it, below.

Are clusters actually reached? (FI-564)

All three clusters a map names — the largest, and the nearest on each side — are recorded the day we draw them, each with a mirror band the same distance from the mark in the opposite direction, where no cluster sits. Both are graded 14 days later on the exchange’s own daily candles. A cluster reached no more often than its mirror is a cluster that attracts nothing, and that result gets published here the same as any other.

The nearest clusters were added after the register’s first days showed why the largest one alone cannot answer this: it sat a median 37.5% from the mark, and Bitcoin travels that far inside a fortnight about 2% of the time. Both bands have to be somewhere price can reach before the comparison says anything, so the cards the page actually leads with are now registered too.

Graded 2,034 claims on 70 coins since 2026-09-18, in 133 coin-fortnight blocks: clusters were reached 39% of the time and not reached 61%; mirrors 35% and 65%.

The gap, clusters minus mirrors: +3.9 points, with a 95% interval of +0.5 to +7.5 points; clusters came out ahead in 99% of resamples. The interval sits above zero, though the true interval is wider than this one. A coin's claims on consecutive days share 13 of their 14 days and often name the same cluster, so claims are grouped by coin and fixed 14-day block, and the blocks are resampled with replacement 2,000 times from a fixed seed, so every build prints the same numbers, keeping the middle 95% of the results. Coins that move together inside one fortnight are not modelled, and the register so far spans 3 fortnights, so the true interval is wider than this one.

Bands and the 20-claim floor below which a row withholds its rate were both fixed before a single claim was graded, so neither can have been chosen against a result. A claim whose window the exchange cannot cover with candles is abstained, never counted as not reached (1 so far). This register is also listed on the settlement referee.

Known bias

Accounts we know skew toward larger and more active traders, who tend to use less leverage than small accounts. The unseen remainder therefore probably liquidates closer to the current price than this chart suggests. We cannot measure that, which is why it is stated here rather than corrected for.