Three things happened in the weeks after each edge in the case studies was found. None of them were bugs.
The late books emptied
The late-window favourite depended on someone still quoting an ask at 285 seconds. Within a week the share of windows with an empty late book went from 23 percent to 84 percent. The makers had not been beaten; they had left. Whoever else found the same edge found the same empty room.
The leaders rotated
Of the six most profitable wallets on one exchange on one day, three had not existed three days earlier and one had flipped sign. Wallets that were positive for a week were, on average, negative the next. Copying any of them, at any delay, lost.
The rules changed
One exchange switched its short crypto markets from settling on the closing price to settling on a 60-second average, with a docs note and no other announcement. Every bot modelling "spot at close" was modelling the wrong thing from that morning. A model that was right for the old rule was wrong for the new one on about one window in fifteen.
What this means for a build
- Set the kill bar before launch: the loss, in dollars, that stops the bot. Decide it while you are calm.
- Check the settlement rule in the exchange's docs weekly, and log the exchange's own outcome, not yours.
- Assume any edge has a half-life of weeks. Budget the engineering accordingly.
- Do not switch strategies on recent performance. A test of "run whichever model did best in the last 12 hours" lost to leaving the best model alone, every time.