Every strategy we've killed.

Everyone believes they could just buy the dips and take the small bounces, and after seven months of watching charts it feels obviously true. So we counted every dip in seven months of minute-by-minute prices, thirteen thousand of them, and made the computer trade them all under fixed rules. About half bounced. The eye remembers those; the account also trades the other half, which kept falling. And on the one coin where dip-buying did make money, buying at completely random moments made money too, because the coin itself was rising: the dip-spotting added almost nothing, and the whole clever exercise earned less than half of what doing nothing would have.

Our bot was bad at deciding when to leave a trade, so we asked whether a machine could learn better exits from seven months of its own history, and we ran the question the way a lab would: rules written down first, a final exam sealed in advance, and independent checkers ordered to rerun everything. Three learned strategies looked genuinely skilled. The checkers then showed all three were accidentally being told when each trade had ended, information nothing can have in real time, like acing a quiz with the answer sheet stapled to the back. Strip that away and the skill vanishes below a coin flip; the honest version was worse than the bad exits it was meant to fix. The exam was never opened. The only exit improvement that survived all of it: take your profit at a fixed nine percent, no learning required.

Traders watch the funding rate, the small fee that keeps a perpetual contract tied to the real price, and bet that extreme readings tell you where the market goes next. Ours seemed to. We left it for four months, came back with nearly five times the data, and the pattern looked stronger than before. Then we removed the double-counting, where overlapping time windows let the same few hours vote over and over, and it vanished completely. On the newest stretch of data it pointed the opposite way.

A signal on a prediction-market exchange looked like it might pay, and a month earlier we had said we simply needed more data. So we collected five more weeks, seventeen times as much, and it was still indistinguishable from luck. This time we worked out how much data would actually settle it: about ninety-two thousand trades, arriving at twenty-eight a day, which is about nine years. The trouble is not the amount of data, it is that a few enormous trades swamp everything else. Some questions cannot be answered, and saying so is far cheaper than collecting for a decade to find out.

If the crowd is usually wrong, betting against it should pay. A month earlier this looked very slightly positive and we called it noise rather than an edge, which turned out to be the right call. With seventeen times the data it is a real, measurable loss of about a dollar a trade, and it crosses the same evidence bar we set in advance, just in the losing direction. The bar we wrote down before looking stopped us trading a flattering number, then killed the idea once the number became real.

We put real money into quoting both sides of a prediction market, aiming to earn the gap between the buy and sell price. Our orders did fill at good prices. The problem is what happens next: there is almost nobody to sell the position back to, so every fill has to be held until the market settles, where the outcome is close to a coin flip worth fifty cents a share against a gap worth under two. And the fills land on the wrong side, because on a falling market everyone sells into your buy order and nobody touches your sell order. We lost about a hundred and forty-five dollars in thirty-five hours and stopped. The deeper mistake was arming it with no maximum loss, no position cap and no exit plan, because the criteria we had written down were about statistics, not safety.

The plan was to leave a buy and a sell order resting in a market for an hour and collect the difference. The profitable version of that sum was actually measured over two seconds, not an hour. Over an hour the price walks away from your orders every single time, so instead of collecting the gap you get run over. And if you quote fast enough not to get run over, almost nobody trades with you: this market fills about one order an hour, the typical trade is a tenth the size of the smallest order we are allowed to place, and it turns over eighteen dollars a day.

The pitch was that prediction markets have huge gaps between the buy and sell price that a patient trader can sit inside, and that the exchange pays a bonus on top for providing that service. We checked both halves. The gaps are about a hundred times narrower than the pitch claimed, and the bonus programme was simply switched off on every market we would have used. There was nothing to collect, and nothing to cushion the losses that come with sitting in a market all day.

“Just copy the winning wallets” is one of the most-pitched strategies in crypto. We mirrored 205 real trades, then re-ran the test removing one wallet at a time. The entire edge came from a single lucky wallet: drop it and a winning record turns into a loss. A widely-believed edge, falsified on real data.

Hyperliquid's stock perps don't use a live market price directly, they rebuild it from an oracle and a lagging formula. The idea was to exploit that lag three ways: the weekend catch-up jump, the split-second the perp trails the index during the day, and the gap between two versions of the same stock. We simulated all three from the venue's own published formulas. Even in the best case the edge was smaller than the fee to trade it, so every version lost money after costs. A one-hour simulation settled it, before any weeks of data collection.

We bet prices would bounce off the day’s high and low; they didn’t. The setup is rare, the move was tiny, and it went the opposite way to the bet.

“Build a bot that trends HYPE” is the exact pitch in a thousand videos. We ran ours live, on real capital, against a bar set in advance: beat buy-and-hold. Over one week it returned −31.9% while holding lost 11.5%. No config rescued it; there was no skill in it, just leverage.

We assumed a price gap would close within half an hour; across 5,800 cases it usually didn’t, the gap tends to stick around, not snap back.

We trained machine-learning models to guess HYPE's next move from the live order book. Up close the moves were too small to beat trading fees; further out there were too few real examples to trust. No version came close to paying for itself.

An upgraded version of the killed trend bot made only two trades in the test window, far too few to prove anything, so we can’t call it a win.

We turned a profitable friend’s trend-following into a live bot on real capital. It made money for 26 days, but 92% of the profit was a single trade, and the three-month backtest of the same setup lost to just holding the coin.

We built a filter to improve the trend bot, but the bot was shut down before the filter ever saw enough trades to be judged.

The “free” price gap between exchanges was never once wide enough to cover trading costs, not in a single hour out of 4,216.

A Bitcoin-linked stock doesn’t predict Bitcoin; it just moves alongside it after the fact, so there’s nothing to trade ahead of.

We wanted an AI to read price charts like a human, but the only trader it could have learned from traded coins we don’t touch, so we stopped before spending on it.

We tried to profit from the price gap between tokenized US stocks and Hyperliquid's version of the same stocks. The gap is tiny and professional traders already keep it closed, so there was nothing left to capture.

A pitch said you could farm the GRVT exchange for a fat trading rebate and a 10% yield. Against GRVT's own fee page, the rebate was a hundred times smaller than claimed and the yield needs millions in trading volume. It only works on paper.

Reading the order book to call the next tick is the dream of every screen-watching trader. We ran five classic patterns over 7.8 million ticks. The signals are real and rock-solid, and still far too small to cover the cost of the trade. Real, measured, and not bankable.

The plan was to farm the Lighter exchange for its token giveaway, like an earlier farm that worked. But the token had already launched and the giveaway was paid out months before the idea reached us. Nothing left to farm.

Yesterday’s funding rate tells you nothing useful about tomorrow’s price direction, and even if it did, the edge is too small to cover costs.

The volatility around the 8-hour funding reset is real, but it’s spread evenly through the cycle, there’s no special moment to time a trade to.

Trading the speed-up in volatility, rather than its level, added nothing, once you account for the level, the acceleration actually hurts.

Bitcoin and Solana volatility move together at the exact same moment, so there’s no head-start to trade on.

The bot’s own “good day” filter actually picked worse days than the ones it skipped, the filter points the wrong way.

Betting that forced sell-offs keep falling worked in old data but fell apart on fresh data, the effect shrank tenfold and flipped direction.

The trade needed HYPE funding data from Binance and Bybit, but HYPE doesn’t trade there, so there was literally nothing to measure.

Liquidation cascades bounce back rather than keep running, so riding them as they speed up loses money.

“Ride the liquidation cascade” is the classic momentum play. We measured 15,992 cascades and sorted them by intensity; the most violent ones snapped back the hardest, the exact opposite of continuing. The detector works, the trade is backwards. Killed.

HYPE’s buy/sell gap is too thin to earn the rebate for posting both orders, the spread disappears before both can fill.

When several exchanges agreed on funding direction, that agreement still didn’t predict anything you could trade.

Betting against “overstretched” moves looked great on a chart but lost money on every single coin we tested it on.

This pattern made a tiny profit, but over only eight trades, far too few to tell skill from luck.

There’s no reliable reversal at the Asian market open; the result was indistinguishable from random noise.

We tested about ninety chart patterns at once; after correcting for the fact that a few always look good by chance, not one held up.

A score meant to flag a fragile order book worked on old data but couldn’t tell fragile from stable when shown fresh data.

We checked other exchanges for an edge Hyperliquid didn’t already have; none offered one, so we stayed put.