34 indicator rules on hourly crypto charts: all of them lost
We took 34 textbook indicator rules, ran them unchanged on 1-hour bars of 22 Upbit coins with realistic costs, and none made money. On daily bars most of the same rules did.
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 test
In September 2026 we asked a simple question: which indicator wins most often, and can it run a trading bot? We coded 34 standard rules exactly as textbooks describe them, with default settings and no tuning: moving-average crosses, MACD, RSI, Bollinger Bands, Keltner channels, Donchian channels, stochastic, Williams %R, CCI, Ichimoku, breakouts with volume and more. Signals were taken at a bar's close and filled at the next bar's open, buying only, with round-trip costs of 0.25% and 0.50%.
The result
| Test | Profitable at 0.25% costs | Profitable at 0.50% costs | Holding the same coins |
|---|---|---|---|
| 22 coins, 1-hour bars, Aug 2025 – Sep 2026 | 0 of 34 | 0 of 34 | −48.5% |
| 4 large coins, 1-hour bars | 3 of 34 | 0 of 34 | — |
| 22 coins, daily bars, Jul 2024 – Sep 2026 | 28 of 34 | 27 of 34 | −5.1% |
| 223 coins, daily bars, Oct 2024 – Sep 2026 | — | 4 of 34 | −44.1% |
On hourly bars every one of the 34 rules lost money at both cost levels. The best, a 20/50 moving-average cross with an ADX filter, lost 0.6% at 0.25% costs and 8.6% at 0.50%. The most common MACD rule, buying the signal-line cross, lost 90%. Holding while the close was above its 20-bar average lost 97.5% over 13,928 trades: the costs took everything.
The oscillators again won often and lost money: RSI reversal, the Bollinger lower band, CCI and Williams %R won 50% to 54% of their hourly trades and lost 32% to 77% in total.
Why hourly signals cannot pay their costs
We also measured how much each buy signal predicted the next 24 hours. The best signals, RSI crossing above 70 and a Keltner breakout, were followed by average moves of +0.42% and +0.30%, against −0.13% for buying at random. Those edges were real, statistically, but about the size of a 0.5% round trip, so trading them could not pay. A signal can be right and still not be worth trading.
A later, larger test was kinder to hourly bars
In October 2026 we ran a much larger study with Upbit's actual fee of 0.05% per trade plus one price tick, on 12 large coins from January 2024, with slower settings than the textbook defaults. There, 74% of 1-hour setups made money and 810 passed our two-period test. The differences explain each other: lower costs, large coins only, slower settings, and a period that included the 2024 rally. The lesson is the same in both: costs, coin choice and speed decide more than the indicator. Results by timeframe.
What the research literature says
- A study of 7,846 technical rules on a century of US stock data found that even low trading costs wiped out their in-sample performance, and that good rules could not be picked in advance (Bajgrowicz and Scaillet, Journal of Financial Economics, 2012).
- About 15,000 rules on cryptocurrencies had no out-of-sample predictive power for Bitcoin: the best rule of 2017 lost in 2018 (Hudson and Urquhart, Annals of Operations Research, 2021).
- After adjusting for data snooping and trading frictions, significant excess returns from technical rules on crypto appeared only rarely (Anghel, Finance Research Letters, 2021).
See every rule on every bar length in the indicator atlas.