Counting Fills Misrepresents Liquidations: Evidence from a Complete On-Chain Derivatives Ledger

Abstract

Forced liquidations in cryptocurrency derivatives are almost always counted from exchange-published event feeds, in which one economic liquidation appears as several records. We quantify the resulting distortion on Hyperliquid, an on-chain perpetual-futures venue whose node-fills archive is complete and attributed to the liquidated account — a combination no centralised venue publishes. On a fixed-hour stratified sample of 351,540 liquidation episodes reconstructed from 2,010,042 fills over one year, counting fills rather than episodes inflates the event count by a factor of 5.72 (day-resampled 95% CI [5.47, 5.99]). The inflation is not a constant: it is size-dependent, rising from a median of 2 fills per episode below the median size to a median of 72 in the top percentile, so the bias is concentrated on precisely the events that liquidation studies are about. The top 1% of episodes generate 23.1% of all fills. We also characterise the episode-size distribution and report a negative result: the tail is unambiguously heavy — an exponential is rejected decisively — and it is not a power law, but it cannot be named beyond that. Lognormal and Weibull specifications both dominate a Pareto fit at every estimable threshold across nearly three orders of magnitude of cut-off, yet which of the two wins reverses with the threshold, so any named family would report the analyst's cut-off. Finally, the bias already documented in the literature runs the other way: rate-limited centralised-exchange feeds undercount, publishing per-second maxima rather than samples, while fill-level on-chain records overcount. Pooling the two sources combines a downward-biased count with an upward-biased one, and they do not cancel. All data, code and pre-registrations are public.