Independent research · public results

We test what people believe about Hyperliquid.
Most of it fails.

Measured on Hyperliquid's own data. We publish the failures first, with the number that killed them.

01 — What we measured

Eight results. Six negative.

Market making loses on arithmetic, not skill.

0.12–2.6 bp of spread against a 3.0 bp maker round trip. 0 of 85 configs profitable.
A profit leaderboard predicts nothing.

Top 100 of one quarter: no predictive power over the next. Three methods agree.
Picking the right trader loses to picking none.

Our own score loses out of sample to a control basket. Bound risk, don't rank people.
A funding rate on a screen is a spike, not a level.

XMR median 11%/yr, p99 213%, spikes last two hours. 23 bp to enter and leave.
On a thin market, your own order is the market.

She made +30.8%; a faithful copy made −0.5%. 23% of lines, 1.8% of notional.
Order-flow imbalance is real, and dies to fees.

The signal is in the book. 0 of 42 configs survive taker fees. Right ≠ right after costs.
Crowd positioning describes. It does not forecast.

A signal on average — flips sign 2 quarters in 5. Shown as a description.
Language models that trade live mostly lose.

Four of six frontier models finished down on real capital. Ours run in paper, against a control.

02 — What survived

Correlation between traders is +0.035. Spreading across accounts multiplied the Sharpe by 4.6; picking the best one never worked.

Hold the spot, short the perp, collect a floor written into the protocol: 0.01% per 8 h, or 11.0%/yr. A cash flow, not an opinion.

Caveat Not risk-free: the spot leg is a bridged token that can decouple, and funding is negative ~22% of hours. It runs in paper on the real book.

03 — What we still can't answer

Your keys stay in your wallet. No custody, no subscription — a builder fee of 4.5 bp per trade, the rate Hyperliquid charges a taker.