Research
Why the bot does what it does
Every rule was tested on Binance futures data from 2021 to 2026 before it went into the bot. These notes show the tests, the numbers, and what we decided, including the changes we rejected. Backtests are hypothetical and chosen with hindsight; we say so on every page.
- Does shorting help a crypto trend bot? Long-only vs both sidesLong-only made 35.5% a year with a −14.7% worst drop. Long and short made 45.7% with −18.5%. Almost all of the difference came from one year: 2022, when long-only lost 1.8% and the shorts made 36.7%.
- How much do fees cost a trend bot? The same rules at six cost levelsAbout 688 orders a year, trading 45 times the account's value. Every 0.01% of cost per trade takes roughly half a point a year off the return, and at 0.25% per trade the weak year 2025 goes negative.
- We could cut 39% of our orders. We decided not to.Resizing less often cut orders from 687 to 420 a year with the same long-run return, but recent years were weaker. Wider still looked great, but only because of 2021.
- Funding: the cost our headline number leaves outEvery backtest figure on this site says "funding not included". So we added it: Binance's real funding history for all ten coins since 2021. It costs about 2.7% of the account a year, most of it in 2021. The return drops from 45.7% to 41.8% a year.
- How bad does it get? Every drop over 8% in six yearsTwelve drops deeper than 8% in under six years. The worst was −18.5%; the longest wait for a new high was 241 days. Four in ten months lost money. This is what copying a trend bot feels like between the good months.
- Leverage set to 5×, used at 0.3×: what the bot really holdsThe exchange setting says 5×. The positions averaged 0.30× the account, were below 0.63× nine-tenths of the time, and peaked at 1.02×. The leverage setting and the risk taken are different things.
- Our cheap orders were costing us money on every exitWaiting 45 seconds for a cheaper fee made sense for entries. On exits it never filled, the price moved away, and it cost 9–22 bps to save 2.
- How 16 trend votes become a trade: one real 4-hour cycleAt 08:00 UTC on 29 September, AVAX scored 0.89 and was not bought, while Bitcoin scored 0.62 and was kept. Here is every vote and why that is the rule working, not a bug.
- How small can a copy account be? Binance minimum orders vs our tradesWith a $100 copy, 68% of the bot's orders would be below Binance's minimum order size. At $500 it's 17%, at $1,000 7%, and from $2,500 almost none. Below about $1,000, your copy isn't really a copy.
- If altcoins just follow Bitcoin, why hold four of them?They mostly do: 77% of 4-hour candles move with BTC. But holding fewer coins cut returns to 19–37% a year, and one stop-loss would cost 2–3 times more.
- Does the profit lock need to react in minutes? We replayed it on 5-minute candles.In six years, positions reached +100% 14 times. Checking every 5 minutes instead of every 4 hours changed nothing: same 3 exits, same returns to 0.01%.
- We tested take-profits and trailing stops. Mostly, they made things worse.Take-profits never helped. Tight trailing stops cost up to 9 points a year. Wide ones looked better, but only because of 2021.
- Why the bot holds four coins, not tenHolding only the strongest 4 coins, one new one per cycle, matched all-10 on returns with far fewer orders. Without a per-coin cap, one coin once reached 101.6% of the account.