RSI in Crypto: What It Measures and Its Real Limits

What RSI measures

The relative strength index, introduced by J. Welles Wilder, is a momentum oscillator bounded between 0 and 100. It answers one question: over the lookback period, what proportion of the average price movement was upward?

That framing is not a simplification. It is the formula.

The widely used open-source TA-Lib implementation, whose source comments state it is based on the original work from Wilder and is intended to represent the original idea behind the classic RSI, notes that the conventional presentation

RSI = 100 minus 100 divided by one plus the ratio of average gain to average loss

is algebraically equivalent to

RSI = 100 times average gain, divided by the sum of average gain and average loss

and uses the second form for speed. The second form is the one worth carrying in your head, because it says plainly what RSI is: the percentage of recent average movement that went up.

An RSI of 70 means 70% of the average movement over the lookback was upward. An RSI of 50 means gains and losses averaged the same size. An RSI of 30 means only 30% of the movement was upward.

Why the 30 and 70 thresholds fail in trends

This is where most RSI usage goes wrong, and the arithmetic is exact rather than a matter of opinion.

Start from the plain-English form. RSI equals 100 times average gain divided by the sum of average gain and average loss. Set that equal to 70 and solve.

• RSI 70 requires average gain divided by total average movement to equal 0.7.
• Which means average gain must be 2.33 times average loss.
• RSI 30 requires the mirror image: average gain equals about 0.43 times average loss.

Working it through

Now take a market in an ordinary uptrend. Over 14 periods it rises on 10 of them and falls on 4, and for simplicity every move is the same size, 1%.

• Average gain: 10 periods times 1%, divided by 14, equals 0.714%.
• Average loss: 4 periods times 1%, divided by 14, equals 0.286%.
• Ratio of average gain to average loss: 2.5.
• RSI: 100 times 0.714 divided by 1.0, equals 71.4.

So a market that goes up on 71% of days, with gains and losses of identical size, prints an RSI above 70 and is labelled overbought.

Why this happens. The 70 threshold is not a measure of how far price has travelled or how stretched it is. It is a statement that up moves have outweighed down moves by a factor of about 2.33 over the lookback. That is the definition of a trend. An indicator that flags "overbought" whenever a trend exists will flag it continuously for the entire duration of the trend.

Reading Figure 1

This is Bitcoin from September 2024 to February 2025. The window was chosen, and it is worth saying why: it is a stretch where the indicator did the thing this section describes, so it shows the failure clearly rather than averaging it away. Read it in three passes.

First pass: when did it say sell, and what happened next. RSI first closed above 70 on 16 October 2024, with price at about $67,600. It went on reading overbought through most of the following six weeks, including a run of nineteen consecutive days in November. Anyone acting on the label sold at $67,600 and then watched the position they exited rise to $106,144 by 21 January, a gain of 57% taken entirely from the wrong side of the trade.

Second pass: look at where the shading actually is. It clusters where price is rising fastest. That is not a coincidence and it is not an indictment of the calculation — it is what the arithmetic above guarantees. A ratio of average gain to average loss above 2.33 is a trend, so the indicator is faithfully reporting the presence of one. The error is entirely in the word attached to it.

Third pass, and the one people skip: look at the actual top. Price peaked on 21 January 2025. RSI that day read 64.5 — below the threshold, not overbought, no signal of any kind. The turn that a trader most wanted to catch is the one the indicator was silent for. So the same figure contains both failure modes: a sell instruction that stayed wrong for three months, and no instruction at all at the moment one would have been worth having.

What to take from it. The shaded stretches are not warnings that were early. They are a description of trend strength, correctly computed, mislabelled by convention. Read them as "this move has been persistent" and the figure becomes informative; read them as "this move is exhausted" and it costs you the 57%.

What this actually tells you. "Overbought" is a misleading label. A high RSI is evidence that a trend is present and persistent, which is the opposite of evidence that it is about to end. Selling every RSI 70 print in a sustained advance means selling repeatedly into strength, and the position is wrong for as long as the trend lasts. The same applies inverted at 30 in a downtrend.

The actionable version. Before applying 30 and 70 to any asset, measure how much time that asset actually spent outside those bands over the past year. If it spent a substantial share of days above 70, then 70 is not an extreme for that asset and the threshold is telling you nothing. Two practical adaptations follow: either shift the bands to reflect the asset's own distribution, or abandon the threshold interpretation entirely and use RSI as a continuous momentum reading where the level describes trend strength rather than exhaustion. In a strong uptrend, the level that matters is usually the one RSI holds above on pullbacks, not 70.

RSI is scale invariant and that matters in crypto

Here is a property of the formula that is rarely stated and matters more in crypto than anywhere else.

RSI is a ratio of average gain to total average movement. If you multiply every price change in the lookback by the same factor, both the numerator and the denominator scale by that factor and the reading does not change at all.

What this actually tells you

RSI carries no information about volatility. The two rows above describe markets with a tenfold difference in daily movement and identical readings. A trader who sizes a position from an RSI level is treating those two situations as equivalent, and they are not remotely equivalent in risk terms.

This is a bigger problem in crypto than in equities. Crypto assets routinely shift between volatility regimes that differ by a factor of several, within the same asset, within weeks. An equity index moves between narrower volatility bands, so a scale-invariant indicator loses less by ignoring magnitude. In crypto it discards the variable that most determines the outcome of a position.

It also explains why RSI settings transfer badly between assets. A 14-period RSI on a large-cap and on a low-liquidity token produce numbers on the same 0 to 100 scale that mean entirely different things in risk terms, because the underlying move sizes differ by an order of magnitude while the reading does not.

The actionable version. Never let RSI influence position size. Pair it with a measure that is not scale invariant, such as average true range or realised volatility, and let that measure determine size while RSI at most informs direction. The two answer different questions and only one of them knows how much you can lose.

Why your RSI does not match someone else's

Three sources of disagreement, all avoidable once you know about them.

Smoothing method. Wilder's smoothing weights the newest observation at one divided by the period. Some platforms offer a simple moving average or a standard exponential average instead, where a 14-period exponential average would weight the newest observation at 2 divided by 15, roughly 13.3%. That is nearly double, and it produces visibly different readings and different threshold crossings.

Candle boundaries. Because crypto trades continuously, a daily RSI depends on where your chart anchors the day, exactly as candlestick patterns do. The same market data produces different daily closes on different timezone anchors, and RSI is computed from closes. This is covered in detail in how to read a crypto candlestick chart.

Data source. Each exchange has its own closes. RSI computed from one venue's data will differ from another's, and on thin pairs the difference can be large enough to move a reading across a threshold.

RSI divergence and what it is worth

Divergence is the observation that price made a higher high while RSI made a lower high, or that price made a lower low while RSI made a higher low. It is widely treated as the most reliable RSI signal. Three structural problems make that hard to sustain.

It is defined retrospectively. A divergence requires two peaks or two troughs. The second one is only identifiable as a peak once price has turned away from it, which is after the move you would have wanted to trade. In real time you cannot know whether you are looking at a completed second peak or the middle of a continuing advance.

Which peaks you choose is a free parameter. On any chart there are many local highs. Selecting a different pair produces a different answer, and there is no standard rule for which pair is correct. An indicator whose reading depends on a subjective selection step cannot be evaluated the way a mechanical rule can, and it is exactly the kind of free parameter that makes backtests unreliable.

Divergence is common in trends and can persist. Because RSI is bounded at 100 while price is not, a sustained advance necessarily produces lower RSI highs eventually, even as price continues higher. That is arithmetic, not a warning. Divergence in a strong trend can persist through several further legs.

What is left of it. Divergence is a compact way of noticing that the rate of change is slowing while the level continues to rise. That observation is real and sometimes useful as context. It is not a timing signal, and any claim about its reliability should be treated the same way as any other pattern claim: ask for the threshold definitions, the peak-selection rule, and the false-positive rate, as described in how to read crypto market signals.

What RSI cannot tell you

• It cannot tell you about volatility. The reading is scale invariant, so identical values describe markets with completely different risk.
• It cannot tell you a trend is ending. High readings are evidence a trend exists and has persisted.
• It cannot lead price. Every input is a price change that has already occurred.
• It cannot react quickly. Wilder's smoothing weights the newest period at one over the lookback, about 7.1% for a 14-period RSI, so a single dramatic candle barely moves it.
• It cannot be compared across platforms without checking settings. Smoothing method, period, exchange and timezone anchor all change the number.
• It cannot tell you position size. Nothing in the formula contains information about how much you can lose.

How to use RSI honestly

For the position-sizing framework that RSI cannot supply, see managing risk in crypto, and for where stops belong relative to volatility, crypto stop-loss strategy.

Conclusion

RSI has an unusually clear definition once you use the algebraically equivalent form: it is the percentage of recent average price movement that went upward. Everything worth knowing about the indicator follows from that one sentence.

The threshold problem is arithmetic rather than opinion. A reading of 70 requires average gains to be 2.33 times average losses over the lookback, which is simply what a trend looks like. A market rising on 10 of 14 days with symmetric move sizes prints 71.4 and is labelled overbought while doing nothing unusual. Selling that label in a sustained advance means selling into strength repeatedly, and staying wrong for the duration of the trend. Before using 30 and 70 on any asset, check how much time it actually spent outside those bands, because for many crypto assets those levels are ordinary rather than extreme.

The scale-invariance property is the limit that matters most in crypto and is almost never mentioned. Multiply every price change by ten and the reading is identical, so a calm advance and a violent one can print the same number. Crypto assets shift between volatility regimes far more than equity indices do, which means an indicator blind to magnitude is discarding the variable that determines how much a position can lose. RSI should never influence position size for this reason.

Two operational points close it out. Wilder's smoothing weights the newest period at one over the lookback, about 7.1% at the default setting, so RSI is slow by construction and a sharp reversal in the reading takes many periods. And a large share of disagreements about RSI levels are configuration differences: smoothing method, period, exchange and timezone anchor all change the number. Used as a continuous description of momentum persistence, checked against the asset's own history, and paired with a volatility measure for sizing, RSI is a reasonable tool. Used as an overbought and oversold trigger, it is a trend indicator being read as a reversal indicator.

Frequently asked questions

Sources and further reading

Data behind Figure 1:

• Bitcoin daily closes for 17 September 2024 to 10 February 2025, Binance BTC/USDT spot. RSI is computed in the figure from those closes using Wilder's smoothing over 14 periods.

Primary sources:

• TA-Lib — RSI implementation source, ta_RSI.c

Related CoinBeaver articles:

• How to read crypto market signals
• How to read a crypto candlestick chart
• What trading volume tells you
• Managing risk in crypto
• Crypto stop-loss strategy
• Why most crypto traders lose money

This article is educational and is not financial advice. The 14-period example with ten gains and four losses, and the calm and violent advance comparison, are constructed illustrations chosen to show how the formula behaves, not observations of any real market. Figure 1 is the opposite: real Bitcoin closes, but a deliberately selected window, picked because the indicator behaved there in the way the section describes. One window demonstrates that a failure mode exists; it measures nothing about how often it occurs, and it should not be read as a rate. No historical win rate for RSI thresholds or divergence is quoted, because any such figure depends on the smoothing method, the period, the exchange data source, the timezone anchor, and in the case of divergence on an undefined peak-selection rule. Confirm your platform's RSI settings before comparing a reading with anyone else's.