Does Bitcoin Bounce After RSI Drops Below 30? The Test, Fees Included

It is the most taught rule in trading: when RSI drops below 30, the asset is "oversold", so it will bounce. It is in every tutorial, every free bot and most public Pine strategies. On October 13, 2026 we asked the question plainly: after RSI(14) goes below 30, what does Bitcoin actually do 1 hour, 6 hours and 24 hours later?

Short answer: no tradable bounce. In 2019-2023, Bitcoin tended to keep falling after the signal. In 2024-2026, the period the study had never seen, even that is no longer shown. After fees, the average is negative everywhere.

Research on past data, not investment advice.

The protocol: written before the first number

An event study asks a simpler question than a backtest: no stop, no target, no position size to tune. It only looks at what price does after a trigger. It is the real test of a signal, before building a strategy on it.

• Data: Binance BTCUSDT one-hour candles, January 1, 2019 to October 1, 2026, 67,861 candles (public klines from data.binance.vision).
• Trigger: RSI(14) < 30, read on the closed candle. Candle i only reads candles 0 to i.
• Measure: return from the signal close to the close 1, 6 and 24 candles later.
• Non-overlapping episodes: a new signal only counts once the previous one has finished its horizon. Otherwise a single crash would count twenty times.
• Fees: 20 basis points round trip (0.20%).
• Reserved period: everything from January 1, 2024 onward is set aside. Discovery happens on 2019-2023; the decision is taken on 2024-2026. No close from the reserved period enters a discovery number.
• Trials counted: the study is registered as one trial in the same counter as the Lab's judge, before any computation. That day it was trial no. 11 for this researcher. The confirmation threshold tightens with the number of trials: α = 0.076%, shared across the 3 horizons, i.e. 0.025% per horizon.

Two baselines are used. The first is chance: the same episodes shifted at random inside the same period (999 shifts). The second is the most ordinary explanation: momentum. A candle that has just dropped hard has its own aftermath, RSI or not. The "controlled effect" removes that momentum.

The numbers

The trigger fired 3,213 times over the period. Once episodes are separated:

The sentence the Lab wrote for the 24-hour horizon:

"At 24 hours: in discovery, 331 episodes, controlled effect −0.537% (random bars, same control: 0.010%), p = 0.005; reserved period, 203 episodes, controlled effect −0.213%, p = 0.081 (threshold 0.025%) → mirage."

How to read this verdict

1. In 2019-2023 the effect exists, but in the wrong direction. After RSI went below 30, Bitcoin on average kept falling: −0.54% over 24 hours once momentum is removed, with p = 0.005. "Oversold" did not announce a bounce; it announced more of the drop.
2. In 2024-2026 that effect is not confirmed. The sign stays negative at 24 hours, but p = 0.081 is far from the 0.025% threshold. At 1 and 6 hours nothing is left (p = 0.43 and 0.42).
3. With fees, everything is negative. Even the best gross cell (+0.027% at 6 hours, reserved period) becomes −0.173% once the 0.20% round trip is paid.

The study verdict is therefore mirage at all three horizons: an effect visible in discovery that does not hold on the unseen period, and that does not pay the fees anyway.

Why so many backtests say the opposite

If you have seen winning "RSI 30" backtests, they often have one or more of these three weaknesses:

• Trials are not counted. Testing RSI 25, 28, 30, 32, then 7, 14, 21 periods, then keeping the best is 12 trials presented as one. The best of 12 looks good by construction.
• No period is set aside. The setting is chosen and judged on the same data.
• Fees are missing. A 0.03% edge per signal is gone at the first real round trip.

The study above tunes nothing: one threshold (30), one RSI period (14), three horizons declared in advance. It was rerun in English: same numbers, same trial (the counter does not move, but one more opening of the reserved period is recorded).

Same method, another signal: aggressive market buying

The same day, the same method was applied to Binance taker flow: when market buys or sells exceed 3 standard deviations of their weekly average, does price follow over 24 hours?

• Discovery 2019-2023: 537 episodes, controlled effect +0.69%, p = 0.003.
• Reserved period 2024-2026: 290 episodes, controlled effect +0.22%, p = 0.065 against a 0.091% threshold. Mean net of fees: −0.10%.

Verdict: mirage as well. The installed Lab reproduced these numbers to the decimal compared with the first run the day before. The burst of aggressive buying had a follow-through in 2019-2023; on the unseen period it is no longer shown and no longer covers fees.

What this study does not say

• It does not say RSI is useless. It says that this trigger, alone, on BTC 1h, does not announce a bounce that pays.
• It does not test RSI in an uptrend, a divergence, or another asset. Each variant would be a new counted trial, with a stricter threshold.
• The reserved period starts in 2024: it is now opened. A future study on that window can no longer claim to discover it.

For an overview of indicators, see RSI, MACD, Bollinger: which ones actually help. On how counted trials work, see what 350 autoresearch sessions on noise show.

Run the test yourself

• The free Windows Lab: the event study runs on your PC, on your data, with the reserved period and the trial counter. An AI (Claude, Codex, DeepSeek through opencode) can launch it over MCP.
• The free Quick Check: paste your "RSI 30" Pine strategy and see how it does against buy-and-hold, fees included.

Key takeaways

• After RSI(14) went below 30, BTC 1h kept falling on average in 2019-2023 (−0.54% at 24 hours, momentum removed).
• In 2024-2026, a period the study never saw, the effect is no longer shown (p = 0.081, threshold 0.025%).
• With 0.20% fees deducted, the mean is negative at all three horizons and in both periods.
• Verdict: mirage. The study was a counted trial, declared before the first number.

Research on past data, not investment advice. No past result guarantees a future result.