Does RSI Overbought/Oversold Work? A 37-Market Audit

"Overbought" and "oversold" are two of the most-used words in technical analysis, and the Relative Strength Index is where most people get them: above 70 the instrument is stretched and due to fall back, below 30 it is stretched the other way and due to recover. That is a reversion claim anchored to a fixed level, and unlike most things said about indicators it can be tested as stated.

We registered the test before running it, ran it on 37 markets, and it came back null: no reversion at the shipped levels, the opposite sign at longer horizons, and a rival built from nothing but the recent return that did at least as well everywhere. This article sets out what was tested, what came back, and the one thing an RSI zone entry did reliably mark.

Read this limit first

The run used daily bars, the shipped RSI(14) with 70/30 levels, and the first bar on which RSI entered a zone. It says nothing about intraday charts, about divergence patterns, about RSI applied to other series, or about other lengths. Oversold entries are scarce on instruments that drift upward, so single-side cells are noisy on some names, and two of the 37 instruments dropped out of the 70/30 cells for want of 25 events on a side. The universe is 37 instruments of daily data across equities, FX, rates, credit, metals, energy, volatility and crypto — 207,790 bars.

What the reading claims, and how it was scored

The overbought claim is that price falls back; the oversold claim is that it recovers. Put together, the forward return after an oversold entry should exceed the forward return after an overbought entry. That difference was measured on the bar RSI first crossed into each zone, with the forward return taken from that bar's close and scaled by the instrument's own recent volatility so that different markets pool fairly. Positive means reversion; the primary horizon was five bars, with one, ten and twenty bars as the curve.

RSI is a bounded transform of the ratio of smoothed gains to smoothed losses over 14 bars — a re-expression of the recent return. So the credit test, registered in advance, compared it with the plain 14-bar return standardised by volatility, thresholded on each instrument so that it fired on exactly as many bars as RSI's zone entries did. If a bounded oscillator with a named level cannot beat that, the transform contributed nothing. A one-bar abnormal-move flag was included as a second reference for the volatility comparison.

The null was a calendar rotation of every series by one common offset of at least a year, with zone entries re-derived from the shifted series, for all 340 admissible offsets; per-cell p-values and a family-wise correction across each grid follow from it. Why that null rather than a bootstrap decides the outcome is explained in the first post in this series.

The result: no reversion at the shipped levels

After 6,328 zone entries at 70/30, the oversold-minus-overbought difference in scaled 5-bar return was −0.05 volatility units, interval [−0.33, +0.20], rotation p 0.61. At one bar it was −0.01. Sixteen of 35 instruments were positive. At the looser 65/35 levels the five-bar figure was −0.04. Whether measured on the entry bar or on every bar spent inside a zone, the reading was followed by nothing that resembled reversion at one to five bars, historically, on this panel.

At ten and twenty bars the sign went the other way: −0.45 at ten bars (per-cell p 0.023) and −0.58 at twenty. After an oversold entry the instrument continued lower on average; after an overbought entry it continued higher. Across the 12-cell grid the family-wise p was 0.23, so those cells are reported as robustness observations and not as a result — but they are the opposite of what the reading claims, and a reader who uses 70/30 as a reversal cue should know which way the longer-horizon sign pointed.

Hit rates did not move either. Overbought entries were followed by a negative five-bar return no more often than any other bar; oversold entries by a positive one 1.7 points more often, not clearing the null. Past behavior is not a guarantee of future results, and none of this is a forecast in either direction.

The credit test

The standardised 14-bar return, at the same event rate, did at least as well as RSI on every cell. At five bars its figure was +0.14 against RSI's −0.05; the difference, RSI minus the bare return, was −0.14 with a rotation p of 0.21, and the eight-cell family had a p of 0.51. The pre-registered kill condition was met in full: on 37 daily instruments the RSI overbought/oversold reading carried no reversion information beyond a volatility-standardised 14-bar return at the same event rate, and no reversion at all at the shipped levels.

What a zone entry did mark

Forward 10-bar volatility on zone-entry bars was 1.17 times that on other bars, and 1.11 times after adjusting for the volatility regime and trend, above one on 32 of 37 instruments. The bare 14-bar return at the same rate gave 1.16, and a one-bar abnormal-move flag gave 1.29, above one on all 37. So an RSI zone entry was a mild volatility event — volatility clustering seen through a 14-bar smoother — and a weaker one than a single large day. That is a description of what the bars were, not a claim the indicator makes.

The level is not portable

Registered as a descriptive endpoint and worth reporting on its own: at 70/30 the entry rate ran from 2.3 to 3.9 per 100 bars across instruments, which looks portable until the two sides are separated. The overbought share of entries had a median of 62% and ranged from 22% to 82%, exceeding 60% on 20 of 37 instruments. An instrument that drifts upward spends its time being "overbought"; one that drifts down is perpetually "oversold". At 80/20 the rate ran from 0.2 to 1.4 per 100 bars, and the pooled reversion figure flipped sign between five and ten bars. A fixed level on a bounded oscillator does not mean the same thing on different instruments, and the direction of the asymmetry is simply the instrument's drift. The same point, made for band width rather than an oscillator level, is why a threshold has to be calibrated per instrument to support any cross-market statement at all — a theme that runs through this site's comparison of band constructions.

The takeaway

Tested as stated, on 37 daily instruments at the shipped settings, the RSI overbought/oversold reading was not followed by reversion at one to five bars, ran the opposite way at ten to twenty, and did no better than the plain volatility-scaled 14-bar return it is built from. What a zone entry did mark was mildly elevated volatility, less than a single large day marks. And the 70/30 level itself fired lopsidedly, with the lopsidedness tracking each instrument's drift. None of this says RSI is useless for every purpose; it says the specific claim the shipped reading makes did not hold on this sample, and that the bare return would have told a reader the same thing with fewer moving parts. The reason for publishing a null this plain is the same as for every result in this series, and it is set out in why we publish our methodology.

Frequently asked questions

Does RSI work better on intraday charts? Untested here. Every figure above is from daily bars, and the limit is stated first because a large share of RSI use is intraday. Whether the shipped reading behaves differently on hourly or 5-minute bars is a separate question that would need its own pre-registered run; nothing in this audit answers it either way.

What about RSI divergence? Not tested. The audit covered the overbought/oversold reading as shipped — the first bar RSI enters the 70 or 30 zone — because that is the claim the levels themselves make. Divergence between RSI and price is a different construction with a different implied claim, and it was neither specified nor run.

Is 80/20 better than 70/30? Not on this sample, and the question has less content than it looks. At 80/20 only 1,001 entries occurred across all 37 instruments, too few for per-instrument statements, so the level was reported only as a pooled figure: mildly positive at one and five bars, negative at ten and twenty. Tightening the level changed how often the reading fired, by a factor of seven across instruments, more than it changed what followed.

Is RSI a volatility indicator, then? Only weakly. Zone entries were followed by about 11% higher standardised 10-bar volatility on 32 of 37 instruments, which is real, but a one-bar abnormal-move flag at the same rate was followed by about 29% higher, on all 37. A 14-bar smoother dilutes the single large day that carries most of the information; if the question is whether a move was unusually large, the smoother is the wrong instrument for it.

If you would like to see what a calibrated expected range shows on the same bars — whether a move was large or ordinary for that instrument, with its calibration record and its limits published on the Proof page — the Behavioral Transform Model draws one on any TradingView chart. You can start a free 30-day trial and compare it with the oscillator you already use.