What RSI actually measures

The Relative Strength Index was created by J. Welles Wilder Jr and introduced in his 1978 book New Concepts in Technical Trading Systems, alongside a June 1978 article in Commodities magazine (Wilder, 1978). It has been on essentially every trading screen since.

First, clear up the name, because it misleads constantly. "Relative strength" here does not mean the asset's strength relative to another asset or an index. That is a different concept entirely. RSI compares an asset only to its own recent past.

The formula is compact:

• RS = average gain / average loss over the lookback period

• RSI = 100 minus 100 divided by (1 + RS)

Wilder's default lookback is 14 periods, the output is bounded between 0 and 100, and 50 is the neutral centreline where average gains and average losses are equal. Those three facts trace cleanly to his original text. One detail gets lost in most explanations: Wilder smoothed those averages with his own recursive method, using a smoothing constant of 1/N, so each new period contributes 1/14 and the previous average carries 13/14. That is mathematically distinct both from a simple average and from a standard exponential moving average, which uses 2/(N+1). Different platforms occasionally get this wrong, which is why RSI values sometimes differ slightly between charts.

Run the formula backwards

Here is the part almost nobody does. If RSI is just a transform of the gain-to-loss ratio, you can invert it and ask what a given reading actually asserts.

An RSI of 70 is not a verdict, a ceiling, or a prediction. It is a compressed way of saying that buyers have dominated the last fourteen periods by better than two to one. Nothing in that statement says the domination is about to stop. In a powerful trend the ratio simply stays lopsided, and RSI stays high, sometimes for weeks.

This is the standard criticism of RSI, and it is worth being straight about its status. It is taught everywhere, but this research pass surfaced no primary source that established it empirically and survived scrutiny. The arithmetic above is the sturdier ground: an RSI of 70 corresponds to average gains of roughly 2.33 times average losses, so for as long as a trend keeps that ratio intact, the indicator will sit above 70. Treat "overbought can persist" as something that follows from the formula rather than as an established research finding.

The 70/30 thresholds themselves deserve a similar note. They are near-universal convention, and many traders shift them to 80/20 in strong trends. Whether Wilder personally endorsed exactly 70 and 30 in his original text is murkier than the confident attributions suggest, and sources genuinely disagree on it. The levels are useful convention. They are not physical constants.

What the testing actually shows

The honest summary is that rigorous evidence on RSI trading rules is thinner than its popularity implies, and mixed where it exists.

The most directly relevant test comes from crypto. A 2023 study swept RSI(14) overbought and oversold rules across BTC/USDT and ETH/USDT, trying thresholds from 70 to 90 on the upside and 20 to 40 on the downside, then applied the statistical machinery designed to catch exactly the trap Part 14 described (Chen, Wang & Yang, 2023). In sample, the rules looked superb: annualised returns of 66.8% on Bitcoin and 68.8% on Ethereum. Out of sample, from December 2021 to October 2023, and after White's Reality Check, Hansen's SPA test and realistic transaction costs, the profits were no longer statistically significant. Across their wider study of four rule families, covering RSI alongside moving-average crossovers, Bollinger Bands and MACD, the authors read the overall pattern as broadly consistent with market efficiency in crypto.

That correction machinery is not an obscure preference. Hsu and Kuan applied White's Reality Check and Hansen's SPA test across a universe of nearly 40,000 trading rules, the same work Part 14 drew on, and that pairing became the benchmark for correcting data-snooping bias in technical-rule research (Hsu & Kuan, 2005). A rule that cannot clear it has not been shown to work; it has been shown to look good among the many rules that were tried.

Fairness demands the other side. A separate 2023 study by Zatwarnicki, Zatwarnicki and Stolarski, published in MDPI's Sensors, tested RSI-based automated trading on Bitcoin, Ethereum and other altcoins and reported decent profitability exceeding buy-and-hold (Zatwarnicki, Zatwarnicki & Stolarski, 2023), though that profitability figure was only checkable here against a secondary summary rather than the paper itself. It sits in genuine tension with the one above. It also reports no Sharpe ratio, no other risk adjustment and no transaction-cost accounting, which is a large gap: a strategy that trades often and ignores fees is not being measured against the thing that actually kills such strategies. Neither study is the last word.

The statistics that do not survive checking

Two of the most repeated RSI claims dissolve when you chase the source.

RSI(2) mean reversion. Short-lookback RSI systems in the Larry Connors style circulate with impressively precise numbers: a 64.33% win rate, 17.84% annual return, a Sharpe of 1.10, a profit factor of 1.45 across twenty years of Nasdaq 100 data. That exact figure set did not hold up under verification, and the underlying backtest carries no out-of-sample test, no walk-forward validation and no data-snooping correction. It runs a Monte Carlo simulation of trade outcomes, which sounds rigorous and is a different thing entirely: it resamples the trades the rule already produced rather than asking whether the rule works on data it was never fitted to.

RSI divergence. Price making a higher high while RSI makes a lower one is one of the most widely taught setups in trading, and one of the least tested. The most substantial empirical attempt found is Bulkowski's, which examined 994 stocks between January 1995 and September 2010, identifying 19,294 peaks and valleys with defined spacing rules and a standardised entry (Bulkowski divergence test). The methodology is disclosed and serious: peaks and valleys at least eight days apart and between three weeks and two and a half months apart, with entries on day nine at the open. The specific success-rate percentage widely quoted from it, however, could not be independently confirmed, and the source is Bulkowski's own site rather than a peer-reviewed journal. That is a limitation of what this survey surfaced rather than a verdict on the wider literature: no peer-reviewed large-scale test of divergence, and no independently verified success rate for this one, turned up in this pass.

Reputation versus evidence

Using RSI honestly

• Read it as a momentum gauge, not a trigger. RSI answers "how one-sided has recent trading been." That is a genuine question worth answering, and it is not the same as "what happens next."

• Stop fading strength in a trend. Shorting because RSI hit 70 means betting against the exact condition the number is reporting. If you use levels at all in a trending market, 80 and 20 at least acknowledge the problem.

• Use the 50 line for context. Persistently above 50 means gains have been outweighing losses; persistently below means the reverse. This is a duller and more defensible use than the extremes.

• Treat divergence as a question, not an answer. It is a reason to look harder at structure and volume, not a standalone entry, and the evidence behind it is thin.

• Distrust precise backtest statistics. A win rate quoted to two decimal places, with no out-of-sample test behind it, is marketing wearing the costume of research.

Common mistakes

• Selling into strength on a 70 print - the reading describes the trend rather than warning about it.

• Thinking "relative strength" means versus the market - RSI compares an asset only to its own recent past.

• Trusting cross-platform RSI values blindly - Wilder's 1/N smoothing is not a standard EMA, and implementations vary.

• Trading divergence mechanically - it can persist for a long time while price keeps going.

• Assuming an in-sample backtest is evidence - the crypto RSI rules looked like 66% a year until they were tested properly.

RSI on ApeX Omni

ApeX Omni's charts carry RSI alongside the price action this series has been teaching you to read. The sensible use on a perpetuals venue is as context rather than command: let RSI tell you how one-sided momentum has been, then take the actual decision from structure, levels and volume, which are independent of it. The stakes are higher here than on a spot chart. Fading an "overbought" reading in a strong trend is a losing habit at 1x and an account-ending one at 20x, because the trend can keep paying the other side long after the oscillator has pinned. If you do trade against momentum, put the stop where the structure says and trigger it on the mark price, exactly as Part 12 argued. Trade the chart on ApeX Omni, and let RSI inform the read rather than issue the order.

The bottom line

RSI is a clean, well-designed measurement wrapped in a badly chosen word. "Overbought" sounds like a verdict on value, when the arithmetic says only that buyers have recently outweighed sellers by roughly two to one over fourteen periods. Nothing about that ratio is unsustainable, which is why strong trends park the indicator at the top of its range and leave it there while everyone waiting for the reversal pays for the privilege. The tested evidence for trading RSI mechanically is weak, the most-quoted statistics behind its popular systems do not survive a source check, and the one thing you can verify for yourself is the formula. Use it to describe momentum. Do not ask it to time a top.

Next in this series: MACD, two lagging averages subtracted from each other, and what the crossover everyone trades is really telling you.

Frequently asked questions

What does RSI measure? It measures the ratio of average gains to average losses over a lookback period, usually 14, expressed on a 0 to 100 scale. It compares an asset to its own recent price behaviour, not to another asset or index.

Does RSI above 70 mean I should sell? No. An RSI of 70 corresponds to average gains being about 2.33 times average losses, which is a description of a strong uptrend. The reading can stay above 70 for as long as that imbalance continues, so it is momentum information rather than a sell signal.

Who invented RSI and when? J. Welles Wilder Jr, introduced in his 1978 book New Concepts in Technical Trading Systems and in Commodities magazine that June.

Is the RSI(2) mean-reversion strategy proven? Not on the available evidence. The precise performance statistics circulated for it did not survive verification, and the backtest they come from includes no out-of-sample testing, walk-forward validation or data-snooping correction.

Does RSI divergence work? It is far less tested than its popularity suggests. The largest test found here is not peer-reviewed, and the specific success rate usually quoted from it could not be confirmed, so treat divergence as a prompt to investigate rather than a proven signal.

Why does RSI show different values on different platforms? Because Wilder's original smoothing uses a constant of 1/N, which is not the same as a simple average or a standard exponential moving average. Platforms that implement it differently produce slightly different readings.

Explore more from this series: Part 12: Order Types | Part 13: Do Indicators Actually Work? | Part 14: Moving Averages | Part 16: MACD Explained

This article is for educational and informational purposes only and is not financial, investment, or legal advice. Do your own research before making any trading decision.