Most trading bots are slop. We have the receipts.
Every “build a bot, get rich overnight” video sells the same dream. We wrote 171,685 lines of analysis code, ran 3,036 automated tests against it, and killed 41 strategies that didn't survive them. This isn't “don't build a bot.” It's how to not build a bad one, and go in knowing more than you would have. No hype, no signals, nothing to buy.
Single investigations, not the total. The biggest sample we publish is 7.8 million ticks; the smallest kill is n=8, and we killed that one because eight trades is too few to trust.
Watch a “winning” bot drain your money.
Our real HYPE-long bot won plenty of its individual trades and still lost 31.9% in a week, while holding the coin lost 11.5%. Put your number in; the return is locked to what it actually did.
Every kill on the ledger has this calculatorLocked to this strategy's real measured result of -31.9% per week, compounded. You choose the amount and the time. We don't choose the return, the strategy already did.
Live data in. Dead strategies out.
A live slice of the market feed, and the research gauntlet it runs through: features become hypotheses, every hypothesis gets a pass/fail line locked in writing, and then the tests run. Almost nothing survives.
The feed is live from Hyperliquid's public API. The flow is an illustration of the pipeline, not a measured throughput; the counts are real and re-derived from the repository.
See the full engineSo does anything actually work?
Yes, but rarely, and never the way the videos say. It comes down to four things, and none of them is a bot you can buy.
- 01Scale
They have a lot of capital, and bet big.
Size lowers the trading costs that quietly kill small accounts, and unlocks edges you can't run with $500. Not a clever bot, a different cost structure.
- 02Skill
They trade better than today's retail AI.
A tiny minority really read context and risk better than an off-the-shelf bot. The AI that beats them belongs to firms like Citadel, and it isn't for sale.
- 03Information
They have information you don't.
Listings before they're announced, flow, team and market-maker knowledge. Real, and structurally unavailable to a downloaded strategy.
- 04Luck
They got lucky on one coin that covered the rest.
Crypto returns are lottery-shaped: a few coins carry it, most go to zero. One winner covers a book of losers, and that's who posts the screenshot.
We kill almost everything. Two things survived.
Out of everything we have tested, two edges survived a real holdout on data they had never seen. That is why the counter above says zero: zero you could buy, not zero that exist. We publish the proof they hold and never the mechanism, which is the only part worth protecting. We are not selling these, which is precisely why we can afford to keep them, and we show exactly why neither is a bot anyway.
Phantom Edge
Real edge, institutional-cost onlyacross which the link between order-book thinness and price reversion held in every one of the ten bands we measured, and still sits below the cost of trading it
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 bankout-of-sample expectancy, with almost no decay from training, the lone survivor of a holdout that killed the same idea on every other asset
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.
Heard a strategy that “works”?
Send us the bot, the video or the screenshot and we will run it against the same data and publish what we find, whichever way it goes. No sign-up, no email list, nothing to unsubscribe from later: we do not collect any of that.