Hyperliquid routed $576M off-book, averting on-book pressure

A research preprint reports that at 21:19 UTC on Oct. 10, 2025, Hyperliquid recorded about $641 million of forced sales. The paper reports roughly $576 million of that value was absorbed by the platform’s backstop while about $64 million reached the public order book. The study treats backstop-filled trades as a separate off-book series.

The preprint measures the backstop share at 62.6% of forced-sale value after the event’s onset. Forced selling was concentrated in time: 87.8% of post-onset forced sales occurred within 30 minutes and 96.5% within one hour.

Across a 15.7-hour post-onset window, the analysis traces $733 million of book-directed forced-sale value, with $644 million occurring during the initial nucleation phase.

The paper describes how Hyperliquid handles liquidations. The venue first attempts to close distressed positions with market orders on its public order book. Under specified conditions, a liquidator vault, a strategy inside the Hyperliquidity Provider protocol vault, can take over and absorb the position off-book.

To study cascade dynamics inside the venue, the authors model a branching ratio: the average number of additional liquidations triggered by each forced sale. Structural branching estimates at the venue level remained below 0.2 across measured regimes, with a value of 0.195 during nucleation and 0.140 at the event peak. A separate amplification calculation produced a ratio of 0.122.

The preprint notes limits to its measurements. The data come from Hyperliquid’s public fill-log archive, which begins on May 25, 2025, making the October event the paper’s only live-case example. The paper has not completed peer review and does not measure broader cross-exchange effects.

The authors present this analysis as Part II of a project that compares seven major Bitcoin perpetual futures cascades from 2022 through 2025. The Part I companion examined early-warning variables across those episodes and found no single indicator that was invariant across events. The current paper frames higher realized branching on venues without comparable backstops as a hypothesis for future cross-venue testing.