What it is
The RSI (Relative Strength Index) compares how much price has risen and how much it has fallen on average over the last 14 candles, and sums it up as a number from 0 to 100. If nearly everything has been gains, it gets close to 100; if nearly everything has been losses, close to 0.
Textbooks call an RSI below 30 oversold and above 70 overbought, and read it like this: price has fallen (or risen) too much and a bounce is due. We measured it.
How we measured it
1-hour BTC, ETH and SOL candles since 1 January 2021, a 14-candle RSI and rules fixed before looking at the results. At each signal we place two bands at ±1.5 times the average hourly volatility (ATR) and check which one price touches first within the next 48 hours. We compare it with chance: the same measure at any random hour, which goes up first 49.6% of the time. The margin (±) is the 95% confidence interval.
Oversold and overbought
| Signal | Cases | Textbook way | Chance |
|---|---|---|---|
| RSI enters oversold (crosses below 30) | 2,217 | 47.9% ±2.1 | 49.6% |
| RSI enters overbought (crosses above 70) | 2,313 | 48.2% ±2.0 | 50.4% |
| RSI leaves oversold (back above 30) | 2,221 | 46.1% ±2.1 | 49.6% |
| RSI leaves overbought (back below 70) | 2,310 | 46.6% ±2.0 | 50.4% |
How to read it:
- Entering oversold was not a bounce signal: price went towards the bounce first 48.0% of the time (±1.5), a little less than chance. An extreme RSI shows up when price is moving hard, and the most common outcome was that the move carried on a bit further.
- Waiting for it to leave the extreme, the textbook "confirmation", did not help either: 46.4% (±1.5). It is what departs most from chance, and it does so the other way. It happened in all six years (from 44.8% to 47.6%) and in all three coins, so it does not look like luck.
- A likely explanation: when the RSI leaves oversold, price has already bounced a bit, and often that bounce was a pause within the fall.
The useful part: the RSI does not tell you where price will go, it tells you how far and how strongly it has run. Read that way, an RSI of 25 is not "buy", it is "be careful about standing in front of this move".
Divergences
There is a bullish divergence when price makes a lower low than the previous one but the RSI makes a higher low: price falls, but with less strength. The bearish one is the reverse, with highs. Textbooks read it as a warning of a turn.
The rules we used, fixed before looking: two consecutive swing lows (or highs), each the extreme of the 3 candles on either side, between 5 and 60 hours apart, and the RSI of the first one below 40 (or above 60). The signal is known 3 hours after the second swing, when it is confirmed, and that is where you decide.
| Signal | Cases | Textbook way | Chance |
|---|---|---|---|
| Plain swing | 30,233 | 49.7% ±0.6 | 50.0% |
| Bullish divergence | 982 | 50.3% ±3.1 | 49.6% |
| Bearish divergence | 1,058 | 51.8% ±3.0 | 50.4% |
| All divergences | 2,040 | 51.0% ±2.2 | 50.0% |
Overall, divergence sits one point above chance, within the margin. But the interesting part is by year:
| Year | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|---|
| Divergences | 50.5% | 46.6% | 57.3% | 57.3% | 48.9% | 46.3% |
| Margin | ±5.1 | ±5.2 | ±5.9 | ±5.1 | ±5.1 | ±5.8 |
In 2023 and 2024 divergence clearly departed from chance; in 2022 and 2026 it fell below it. The same signal behaved differently depending on the market. That is exactly what you train: recognising what context you are in before trusting a signal.
How to use the RSI
- As a measure of strength, not direction. An extreme RSI says the move is strong; standing against it for that reason alone did worse than chance.
- Together with the rest of the chart. In another study we combined a holding level, an extreme RSI, trend and volume at once: 51.4% (±4.8) went in favour. Even stacking signals there is no certainty; that is why every trade has its stop and its size.
- Watching the context. What worked one year did not work the next: train with moments from different markets.