Does it actually step aside?
We say the bot "steps aside when the market turns dangerous". That is a claim about drawdown, and until 2026-09-09 it had no measurement behind it. So we measured it — and we are publishing the half where it does not hold, because a claim you only test where it passes is not a tested claim.
The short version
At 4-hour cadence the claim holds comfortably: 39% drawdown against buy-and-hold's 72%. At 15-minute cadence it fails: 87–100% against 81–97%. The mechanism is the same in both cases — a long-only trend follower is out of the market roughly 57% of the time — but at 15 minutes the trading costs compound the account toward zero, and an account going to zero has a ~100% drawdown however well it steps aside.
Where it holds — 4-hour bars, BNB/USDT
| Strategy | Exits | Strategy drawdown | Buy & hold | Verdict |
|---|---|---|---|---|
| BREAKOUT_V1 | ATR stop | 39.24% | 72.49% | holds |
| EMA 12/26 | ATR stop | 48.10% | 72.49% | holds |
| BREAKOUT_V1 | none | 58.37% | 72.49% | holds |
| EMA 12/26 | none | 55.49% | 72.49% | holds |
| TSMOM_V1 | none | 66.95% | 72.49% | holds |
| TSMOM_V1 | ATR stop | 84.37% | 72.49% | fails |
11 of the 12 configurations tested at this cadence beat buy-and-hold, several by more than 30 points. One does not, and it is in the table.
Where it fails — 15-minute bars
| Series | Strategy drawdown | Buy & hold | Verdict |
|---|---|---|---|
| BTC / ETH / SOL, all three strategies | 87.3 – 100.0% | 81.5 – 96.9% | fails, 34 of 36 |
| ADA, LINK at 1 hour | 90.4 – 99.1% | 90.5 – 95.5% | mostly fails |
This is not a different mechanism, it is the same one losing a race against cost. A signal worth about 2 basis points per round trip, traded against a 10 basis point round trip several thousand times, drives the account down regardless of when it chooses to be out.
What this does not say
It does not say the bot makes money — we measured that separately and it does not. On 504 out-of-sample configurations across 56 symbols the same strategies land at the 52nd percentile of random timing: entering at random, with the same exposure and the same costs, earns the same. Avoiding drawdown is a real property. Picking the moment is not a skill we can demonstrate, and we do not claim it.
Method: drawdown is peak-to-trough of the per-bar, net-of-cost account value, so it sees
losses inside an open position rather than only at the exit. Buy-and-hold is peak-to-trough
of price on the same bars and pays one round trip, so it is cost-advantaged. USDⓈ-M
perpetual bars, 2021-01 to 2026-08, Binance public archives, every file checksum-verified.
Full write-up and the code that produced it are in our repository:
docs/research/exit-trigger-gap-and-drawdown-2026-09-09.md.
If you want the rest of the week that produced this — including the one result we published as a success and then killed ourselves six hours later — the beta record is here.