Overbought has never meant too high. Measured on 47 Bybit perpetuals and nearly three years of daily candles: what actually follows an entry into extreme territory, and why the trend regime changes the meaning of the same number.
It is the most misused indicator in technical analysis. Not because it is bad, but because its name misleads: overbought has never meant too high.
The previous articles dealt with price references — the EMAs 13, 25 and 32 for an asset's state, the long averages for its structure. The Stochastic RSI does not measure a price. It measures a position within a recent window, which makes it an instrument of another nature — and calls for another reading.
The RSI measures the ratio between gains and losses over the last fourteen candles. The Stochastic RSI applies the stochastic question to that RSI: where does the current value sit between its lowest and highest of the last fourteen readings? The result is scaled from 0 to 100, then smoothed.
This double transformation has a consequence that is systematically forgotten: the Stochastic RSI is normalised by its own window. It reaches 100 as soon as the RSI touches its maximum of the last fourteen values — whether that maximum corresponds to a 40 % surge or a 2 % drift. The indicator knows nothing about amplitude. It knows only rank.
In other words, it answers “where are we in the current move?”, and never “has this move gone too far?”. That second question belongs to stretch, which scales the distance travelled by the asset's own volatility. Two distinct questions, two distinct instruments.
The faulty reasoning fits in one sentence: the indicator is above 80, so it is expensive, so I sell. It sounds reasonable. It is wrong, and it is costly — because a normalised indicator stays pinned to its upper end for the whole duration of a strong trend.
That is a mechanical property, not a market anomaly: as long as each new candle prints an RSI close to its recent high, the ratio stays close to 100. An asset can therefore read 95 for three weeks and double in the meantime. Selling on that basis alone means getting out precisely where the move is working.
What remains is to find out by how much. Rather than assert it, we can measure it.
Measured on the 47 most traded Bybit perpetuals with at least 320 daily candles, from 10 December 2023 to 4 September 2026. The calculation is the screener's own, in 14·14·3·3. The trend regime is defined by where price sits relative to its EMA 200. We record the average return 10 and 20 days after each event, compared with that of an ordinary day in the same regime — the control, without which an isolated figure means nothing.
First, above the EMA 200:
| Situation, price above the EMA 200 | Cases | 10 d | 20 d |
|---|---|---|---|
| Ordinary day (control) | 8,903 | +1.21 % | +2.01 % |
| Entering overbought (> 80) | 459 | +1.73 % | +2.11 % |
| Leaving overbought (back < 80) | 445 | +2.29 % | +4.90 % |
| Entering oversold (< 20) | 415 | −0.39 % | +0.11 % |
The result is unambiguous. Entering overbought territory in an uptrend announces no decline: what follows is better than an ordinary day. And crossing back below 80 — the “sell signal” par excellence — is followed by +4.90 % over twenty days, close to double the control. It is the best configuration in the whole measurement, and precisely the opposite of what the common reading predicts.
Symmetrically, the “bargain” is not one: entering oversold above the EMA 200 gives −0.39 % over ten days, against +1.21 % for an ordinary day. Buying because the indicator has just dropped below 20 means stepping in front of a move that is still under way.
Below the EMA 200, the same thresholds change meaning:
| Situation, price below the EMA 200 | Cases | 10 d | 20 d |
|---|---|---|---|
| Ordinary day (control) | 20,367 | +1.05 % | +2.42 % |
| Entering overbought (> 80) | 1,013 | +0.26 % | +1.11 % |
| Entering oversold (< 20) | 1,039 | +1.78 % | +3.28 % |
Here alone the indicator behaves as the textbook says: overbought below the EMA 200 is followed by +0.26 % against +1.05 % for the control, and oversold pays more than average. The very same value — 82, say — therefore precedes +1.73 % above the EMA 200 and +0.26 % below it. The threshold is identical; what changes the outcome is elsewhere.
Two caveats are essential, without which the table above would read as a recipe.
The median is negative in fifteen of the sixteen situations measured, control included. The positive averages come from a minority of large-amplitude moves, not from regularity. That is this market's signature, and it forbids translating “+4.90 % on average” into “I will make money”.
The proportion of winning cases barely moves. It stays between 39 % and 51 % across the sixteen situations, against 42 % to 46 % for an ordinary day. This is the most important finding of the whole measurement: the Stochastic RSI does not improve the frequency of correct decisions. It shifts the amplitude of what follows, not the probability of being right.
An instrument that does not change your chances of being right is not a directional signal. It is something else.
Direction is decided elsewhere — by structure and trend regime. Once that decision is made, a separate question remains: now, or later? That is where, and only where, the Stochastic RSI works.
Its practical value comes down to three uses:
In all three cases, the indicator comes in after selection, never in its place.
The Stoch RSI column gives %K and %D for each of the four timeframes — 4 h, 1 day, 3 days, 1 week — in 14·14·3·3. The useful reading combines three elements, never the value alone:
An asset whose averages are favourably arranged and whose daily Stoch RSI is low but turning up describes a resumption after a pause. That is the configuration the filter highlights — not because it wins more often, but because it offers a cheaper entry point on a structure already validated.
The thresholds are not sacred. 80 and 20 are conventions. Nothing in the measurement gives them special status; they serve as common markers, not as triggers.
The setting changes the readings. The screener calculates in 14·14·3·3. Other schools use a shorter RSI, more responsive, which reaches the extremes more often. That is neither better nor worse: it is a different trade-off between responsiveness and false alarms. What matters is not to compare readings obtained with different settings.
A past measurement is not a promise. The figures above cover a specific window, shaped by the market conditions of that period. They establish that the common reading is wrong on this sample; they do not establish that the opposite reading would be profitable.
On 4 hours, noise dominates. The shorter the timeframe, the more extremes follow one another without consequence. The measurement here covers the daily; nothing permits transposing it as is.
What remains is modest and solid: an indicator that does not say where price is going, but where you stand in what it has just done. Used to date a decision already taken, it is useful. Used to take that decision, it costs money — and the measurement puts a number on how much.
This is not investment advice. This article sets out technical-analysis concepts for educational purposes. Past performance is no guide to future performance, and trading crypto-assets carries a risk of losing the entire capital committed. Legal information.