The machine that does the killing.
Market microstructure goes in. Features become hypotheses, hypotheses get a criterion locked in writing, and then the tests run. Almost nothing survives. That funnel, run at this volume, is the work: it is far cheaper to kill a bad idea than to trade it.
Counts are real and re-derived from the repository. The flow is an illustration of the pipeline, not a measured throughput. Nothing here uses private data.
Market data
Sub-second order-book and trade data from Hyperliquid, collected continuously. The raw material.
Features
The signals we derive from it: funding, basis, volatility, flow. Candidates, not conclusions.
Hypotheses
Each becomes a written prediction with a pass/fail line, locked before we look at the result.
Tests
The criterion runs against fresh data. Over 2,000 automated tests guard the machinery itself.
Verdicts
Almost everything is killed. What survives is rare, small, and only works at institutional scale.
Why it matters
A bot you can buy never shows this. The funnel is the proof the work was actually done.
Zero you could buy. Not zero that work.
Out of everything the funnel tested, two things survived a proper holdout. Neither is a product you could buy, and we will not hand over the signal. What we publish is the proof they exist: the validation, the sample, and exactly why each is an institutional play, not a retail bot.
Phantom Edge
Real edge, institutional-cost onlyWhen Hyperliquid's vault absorbs forced flow through a thin order book, the thinner the book, the harder price reverts afterwards, and it ranks in the same order across all ten thinness bands (a Spearman correlation of negative one, across 24,671 events). The directional trade on the most extreme tail did NOT survive: frozen and tested on fresh data it collapsed to +1.12 bps at p=0.449, a textbook case of an edge that lives in-sample and dies on contact. What survives is the relationship itself, a real piece of market structure that did not exist in the literature before.
The catch. This is a quoting input, not a bot you could buy. Every measured effect (2.63 bps reversion at 60 minutes) sits below the retail cost to trade it, and it only crosses into profit at institutional fee tiers. It is capacity-bounded to thin-book assets. We show the exact economics below rather than imply a return.
Asia Range
On trial, sample too thin to bankA session-liquidity breakout on BTC was the only thing left standing after a proper holdout test killed the same idea on ETH, SOL and HYPE. Out-of-sample expectancy was +0.188R per trade, against +0.189R in training, near-zero decay, on the single asset and filter that survived.
The catch. On just 19 out-of-sample trades, which is thin enough that it could still be noise, and we say so in our own notes. It has only ever been paper-traded, never run with real capital. We publish it as a survivor on trial, not a proven winner.
The edge is real. Your fee tier decides if you keep it.
Read this first: not worth running. The figures below are what this edge would pay before you park the capital, run the infrastructure and carry the risk of every position. It is a rubbish return, and we say so in full underneath the table.
Research results, not a recommendation and not a return you can achieve. Trading crypto perpetuals is high risk and you can lose all of your money. There is nothing on this site to buy, and we are not selling a bot, a signal group or a course.
The move is fixed at 2.63 bps. To capture it you enter and exit as a maker, so your cost is two maker fees. At retail rates that is more than the edge, so you lose. As your 14-day volume climbs Hyperliquid's real fee tiers, the maker fee falls, and somewhere on this ladder the edge crosses into profit.
Every bps figure below is per round trip, not per day and not per year. One event is one enter-and-exit. The last column turns that into money, so the return has a time period attached to it.
Put $250k in at 20x and you reach Tier 1, where the edge first clears its cost. That returns +$1,499 a year, or +0.60% on that capital.
Drag the slider and the dollar figure does not move. Turnover changes how much capital you have to park to reach a tier, never what the edge pays: that is set by the volume you must trade, not the money you hold. And every figure in this column is a ceiling, because it assumes every dollar you trade is one of these events. It will not be.
The verdict: not worth running. That is the honest read of the number above, and it is a rubbish return. To earn it you park the capital, run the infrastructure, carry the risk of every position, and watch the book. The figure is gross of all four. The edge is real and it survived a holdout, which is why it sits on this page at all. It is still not a business, and we are not going to pretend otherwise just because we are the ones who found it.
The honest catch. 2.63 bps is the density-gradient reversion, the robust structure. The directional trade on the most extreme events died on fresh data. It is capacity-bounded to thin-book assets, and net stays a thin sliver even past the line. This is what “killed at retail cost” means: not no edge, an edge only an institution's fee tier can reach. The percentages punish careless reading of the slider. A high turnover makes the percentage look investable, but the effect only exists in thin books: the deep, liquid assets produced no measurable effect at all. Turning your capital over dozens of times inside a thin book is exactly the thing that book will not let you do without moving the price you are trying to capture. Fee tiers are Hyperliquid's published schedule; the effect size is our own measurement.