What 284 builder-deployed perpetual markets look like from the inside

Census 2026-09-12 · 57 fillable markets · 209,460 market-hours · Hyperliquid info API · github.com/perpcensus/hip3-perp-research

284Markets ever deployed

51%Delisted and settled

0Listed markets sitting idle

131 of 140Live markets at 0.90 bps, not 9.00

27.84 bpsCheapest hedge that exists

No configuration shows a significant positive edge

Twenty-five threshold and holding-period combinations, tested out-of-sample on the held-out second half of each market's history, hedged, and priced against order-book depth measured rather than assumed. Only 2 of 25 remain profitable at all. The best reaches a clustered t of 0.35 — below even the naive 1.96, against a bar of 3.08 that a 25-cell search demands.

0.35Best profitable cell, clustered t

3.08Bar the search demands

−31.0Strongest losing cell, clustered t

17 of 25Cells that reliably lose

The bootstrap critical value of 3.08 was reached from the data — resampling whole calendar days and taking the maximum |t| across all cells per replicate. The Bonferroni bound for 25 tests is 3.09; the two bars agree.

Why this is a result rather than an underpowered test Twenty-three cells lose money and seventeen of them clear the same 3.08 bar on the losing side, one at t = −31.0. The test detects effects of this size without difficulty. It finds that trading the signal frequently and paying the spread is a dependable way to lose, and finds no significant winner anywhere on the surface.

Hedged returns, out-of-sample, net of per-market fees and measured L2 slippage at $10,000 per side. 57 of 60 cached markets can fill that size; xyz:CXMT, xyz:DELL and xyz:JP225 are dropped, not costed at zero.

Profit surface: every cell searched clustered t, out-of-sample

Entry	1h	4h	8h	24h	72h
20%	−29.46	−31.00	−29.59	−15.52	−0.91
50%	−23.04	−23.72	−22.11	−11.83	−1.91
100%	−17.74	−18.09	−15.42	−8.04	−1.42
200%	−11.20	−10.32	−8.66	−4.40	−1.16
400%	−5.68	−3.06	−1.47	+0.32	+0.35

Entry is the funding dislocation required; columns are hours held. The only two positive cells sit at 400% / 24h (+1.09 bps, t = 0.32) and 400% / 72h (+1.77 bps, t = 0.35). After measured costs the leftover is those two weak cells, not a monotonic surface.

The hedge leg is priced out by protocol design

Of 60 cached markets, exactly one can be hedged on Hyperliquid itself: xyz:GOLD against the native PAXG perp. That is the most favourable hedge that could exist — same exchange, same collateral, no settlement risk, both legs trading around the clock.

Leg	Fee / side	Half-spread	Cost / side	Round trip
xyz:GOLD	9.00 bps	0.11	9.11 bps	18.22 bps
PAXG (hedge)	4.50 bps	0.31	4.81 bps	9.62 bps
Total				27.84 bps

Cheap fees and a cheap hedge are mutually exclusive Growth mode — the 90% fee cut the entire result depends on — is only available for markets disjoint from validator-operated perps. But being disjoint from native perps is exactly what makes a native hedge unavailable. xyz:GOLD pays full freight at 9.00 bps precisely because PAXG exists to hedge it with; the 131 markets on 0.90 bps qualify precisely because nothing on Hyperliquid can hedge them.

27.84 bps against a gross edge of 1.77 bps that is already not statistically significant.

Half of everything ever deployed has been shut down

284 markets across 10 venues; 144 delisted and settled. The obvious reading is wrong: there are no zombies. Every listed market has real flow — the quietest did $2,025 in 24 hours and not one of the 140 falls below $1,000. This is a 51% mortality rate, not 51% idleness.

Deaths by month 88 of 144 dated

Dec 20251

Jun 202657

Jul 202612

Aug 20267

Sep 202611

Count of markets by month of death (UTC). Dates recovered for the 88 dead markets the API still serves funding for; 56 return nothing at all.

Four venues died inside three weeks 57 markets stopped trading in June 2026 — 65% of every dated death. km 06-10 to 06-17, vntl 06-12 to 06-18, flx 06-18 to 06-19, then cash 06-30 to 07-02. Each venue died all at once.

Venue	Operator	Deployed	Still trading	24h volume
`xyz`	XYZ / trade.xyz	120	104	$1,867.6M
`para`	Paragon	35	26	$3.0M
`io`	EntropyIO	9	6	$31.3M
`mkts`	Kinetiq — still trading	23	4	$10.2M
`hyna`	HyENA	25	0	—
`km`	Kinetiq — dead, June 2026	23	0	—
`cash`	dreamcash	17	0	—
`flx`	Felix Exchange	16	0	—
`vntl`	Ventuals	15	0	—
`abcd`	ABCDEx	1	0	—

Kinetiq appears twice: km deployed 23 markets and killed all of them in June 2026; mkts is a second venue from the same operator and is still trading.

Six things this research got wrong before publication

Each was caught by re-checking an input rather than admiring an output.

1Fees 10x too high

Assumed the published 0.090% taker rate applied universally.

131 of 140 live markets run growth mode at 0.90 bps. Inverted the result from 0 of 25 profitable to 15 of 25.

2Mortality read as idleness

Reported 145 markets with working oracles and no order flow.

All were delisted and settled. Not one listed market is idle — a 51% mortality rate, not 51% idleness.

3t-statistics assumed independence

Quoted t = 2.19 on the best of 25 searched cells.

Clustering by day widens errors up to 4.92x and the search demands |t| > 3.08. Best falls to 0.35.

4Eleven live markets dated as dead

The funding bisection walked to the newest row when the tail held no zeros.

A real death leaves zero rows after it. xyz:NVDA has none. Removed 11 spurious September deaths.

5Unfillable books and optimistic fills

Three names that cannot take $10,000 were costed at zero slippage; price PnL filled at the signal close; holds that spanned a missing hour were kept.

Drop the three thin books, enter at the next hour's open, refuse a hold that crosses a hole. Best cell moves from +8.83 bps (t = 1.90) to +1.77 (t = 0.35).

6Claimed novelty that did not exist

Wrote that no joined HIP-3 market table was published anywhere.

Hyperliquid Guide publishes one live and Dune curates a registry. Repositioned around what is actually new.

What is genuinely new here

Lifecycle dating. Birth and death dates recovered by bisecting funding history, which the leading public registry explicitly declines to attempt.

Data perishability. 39% of dead markets return no candles and no funding.

Per-market true fees. The published flat 0.09% taker figure is wrong for 94% of live markets.

Measured book depth and the growth-mode / hedge exclusivity above.

What is not new, and whose it is

A joined HIP-3 market table is already public. Hyperliquid Guide publishes one live, and Dune curates an RWA market registry. The census here reproduces that rather than introducing it.

Stated because an earlier draft claimed the census was unpublished anywhere. It was not true.

Limitations what would change the answer

The hedge is measured for 1 of 60 cached markets. Pricing an equity or NYMEX hedge properly needs a broker statement, not this API.

Execution cost was never sampled during a US cash session. Both samples fall outside 13:30–20:00 UTC.

105 day-clusters is enough to widen errors honestly, not to detect a small effect. The claim is “no demonstrable edge”, not “no edge”.

Growth mode is a deployer setting, not a protocol constant. It carries a 30-day cooldown per asset and can be withdrawn.

Every figure is produced by a committed script against the public, credential-free Hyperliquid info endpoint. 30 offline invariant tests pin the ways the analysis could be silently wrong in a flattering direction. Source: perpcensus/hip3-perp-research.