Two weeks of a rule-perfect trading server
A dedicated server in Seoul ran our best trend rule for two weeks with ₩10,000,000 of paper money. It followed every rule, made no errors, and lost −₩392,842 (−3.93%).
How this was made: the code, the research runs and the first drafts of this text were produced by an AI assistant (Claude) working under the operator's direction; the operator made every decision and checked the numbers. All trading was on paper: real prices, no real money. Nothing here is investment advice.
The setup
- Rule: buy when the four-hour close breaks the 55-bar high, sell on a trailing exit two average ranges below the peak, stay out while Bitcoin is under its 100-bar average, 5% stop. Up to four coins from the eight most traded on Upbit. From 7 October, half the money ran a second trend rule (AND-3) and an intrabar −15% emergency stop was added.
- Paper fills used Upbit's live order book at the moment of the order, with Upbit's 0.05% fee per trade.
- A small cloud server in Seoul, a standalone Python program, Telegram alerts, and an external heartbeat monitor.
Every trade
| Coin | Bought (KST) | Price, KRW | Sold (KST) | Price, KRW | Held | Result, KRW | Exit |
|---|---|---|---|---|---|---|---|
| XLM | 10-01 09:01 | 308 | 10-02 01:01 | 298 | 16 h | −83,586 (−3.34%) | Trailing exit |
| ETH | 10-02 21:00 | 3,713,277 | 10-03 09:01 | 3,629,637 | 12 h | −58,263 (−2.35%) | Trailing exit |
| BTC | 10-02 21:00 | 116,741,000 | 10-04 09:01 | 115,182,010 | 1 d 12 h | −35,551 (−1.43%) | Trailing exit |
| WLD | 10-03 21:00 | 822 | 10-05 13:01 | 769 | 1 d 16 h | −160,669 (−6.54%) | 5% stop (on the bar close) |
| BTC | 10-05 09:01 | 116,789,986 | 10-07 13:01 | 114,270,892 | 2 d 4 h | −54,772 (−2.25%) | Trailing exit, filter off |
Where the loss came from: price moves −₩368,479, fees −₩12,147, fills versus the signal price −₩12,215.
The audit
- All 87 four-hour bars processed, 45–64 seconds after each close. Zero errors, no unplanned restarts.
- An independent replay, written separately and fed with public exchange data, made the same 10 decisions on the same bars for the same reasons.
- The coin universe matched an independent calculation on all 15 days, in the same order.
- Fills averaged 0.05% worse than the signal bar's close, inside the 0.10% the research assumed.
- From 7 October 13:00 Bitcoin stayed below its 100-bar average, so the server held cash. A copy without that filter bought NEAR and ONDO on 12 October and lost −1.83%.
What it means
In nine years of replay, 12.9% of two-week windows were this bad, so the result is unlucky but not unusual for this rule (why two weeks cannot prove anything). The operator's judgment is still clear: the system worked, the method did not make money, and holding Bitcoin would have done better. Trading research is paused until better AI models are available.