Two lists, ten items. Each one is backed by a number from the logs, and each one cost or saved real money. Start here before the API guides.
Do
- Log the order book before anything elsePublic win rates are marketing. A week of your own book snapshots settles more arguments than any thread, including this site.
- Backfill official resolutionsWinners derived from the last price are wrong 2 to 7 percent of the time. Fetch the settled result by market id, every market, every time.
- Host in the exchange's regionMoving from a far region into the exchange's cut one order path from about 700 ms to about 190 ms and turned a flat model positive for a week.
- Set the kill bar before launchWrite down the loss that stops the bot before it trades. Every edge in the case studies decayed within weeks.
- Hold winners to settlementNo exit rule of any shape beat holding: price stops, time stops, trailing, learned. Redemption is free; selling early pays the spread twice.
Do not
- Copy leaderboard walletsThe skill is real and does not transfer: two thirds of their trades have no book left to fill by the time they are visible, and the third that does is the losing third.
- Buy cheap underdogsBelow 65 cents loses on every dataset, exchange and season tested. The book is calibrated; the fee is not.
- Trust a 5-second-grid backtestSampling the book every 5 seconds overstates late entries about twofold. Tick data or nothing for anything that fires in the last minute.
- Rent a burstable VPS tierCPU credits run out in days on a bot that polls a book, then the box throttles and every latency number you measured is wrong. Dedicated vCPU only.
- Chase 15-minute momentumThe 15-minute markets went negative on every model trained on them, three times over, on the same days the 5-minute markets paid.
The one rule behind all ten
Measure at the resolution you trade at, with the labels the exchange actually settles on, and decide the stop before the start. Everything else on this site is a special case of that sentence.