MEASURED · 23 SEPTEMBER 2026

Does RSI work on gold?

Yes — and this is the first indicator we have tested that survives the statistics. It just does not work the way it is taught. Across twelve years of XAUUSD, higher RSI readings preceded higher returns, which is the opposite of what “overbought” is supposed to mean.

Why we did not test the 30/70 rule

Nearly every RSI study asks whether buying below 30 and selling above 70 works. That tests folklore rather than the indicator. Wilder picked those numbers; they are conventions, not measurements. If RSI carries information, the useful question is where on the 0-100 scale it sits — and the answer could be nowhere, or at 12, or only at the extremes.

So we mapped the whole curve. Every ten-point band gets the same treatment at three horizons on three timeframes, and the thresholds come out as a result instead of going in as an assumption. We tested the threshold approach directly once before, on the stochastic oscillator; this asks whether the idea of a threshold is right at all.

The control is the whole article

Gold roughly tripled over this sample. The average forward return from any bar is positive, so “price rose after RSI fell below 30” is a statement about the bull market. Every figure below is an edge: the return in that band minus the return from a randomly chosen bar over the same horizon. Zero means the band told you nothing that being in the market did not.

Before it ran, the RSI implementation was checked against an independent one (identical to ten decimal places), confirmed to return 100 on a monotonic rise and 0 on a fall, and tested for lookahead by verifying a reading does not change when later bars are added.

The shape of it

One day ahead, on 15-minute bars — the largest sample here. Base rate 0.0585%.

Read the edge column from the middle outwards. From about RSI 30 upward it climbs steadily: 30–40 is negative, 60–70 positive, 70–80 more so, and 80–90 is the strongest band on the table. That is a gradient, not a threshold — and it slopes the wrong way for the rule everyone is taught.

It is not chance

Every previous study in this series ended with “the hits match what chance produces”. This one does not.

31 hits where 3.8 were expected is roughly eight times the chance rate, and 10 bands clear a threshold corrected for all 75 tests at once. Whatever is going on, it is not the multiple-comparisons illusion that explained our seasonality and golden cross results.

Which way they point

7 of the 10 point to momentum — high RSI followed by strength, low RSI by weakness. Only 3 behave the way the textbook says, and all of them sit in the deep-oversold bands over horizons of a week or more. The single largest result in the study is one of those: RSI 20–30 preceded a 0.99% excess return over the following month, on 184 daily observations.

So both stories are in the data, at different depths and different speeds. What is not in the data is the version printed in every textbook: that a reading above 70 means price is due to fall.

Before you trade any of it

Gold’s median spread is $0.20 on a median price of $1,827 — 0.0113%. Any edge smaller than that is inside the cost of capturing it. 9 of the 10 clear that floor. The one-hour result does not, and it is the one with the most observations behind it, which is a fair summary of how this usually goes.

Clearing the spread is necessary, not sufficient. None of this was tested as a strategy with entries, exits, position sizing or slippage — it is a measurement of what followed a state, which is a much weaker claim than “this makes money”. Our risk-reward study and ATR stop study both found that how you manage a trade changes the shape of results, not their sign.

The confound we cannot remove

RSI is calculated from recent price changes. A high reading is a description of recent strength, so discovering that high readings precede further gains is uncomfortably close to discovering that gold trended — which it did, roughly tripling over the sample.

The base rate controls for the overall drift. It does not control for trend persistence, and no arrangement of this test can, because the indicator and the thing being predicted are both functions of the same returns. The honest reading is therefore narrower than the headline: gold’s momentum persisted over this period, and RSI is a momentum measure, so RSI appears to predict. That is a fact about gold in 2014–2026, not a hidden property of the indicator, and it is the reason we have not built anything on it.

The number that would have made a better headline

RSI above 90 preceded a -0.14% one-hour move with a p-value of 0.006. “Extreme overbought crashes gold” is a far better story than anything above, and it rests on 31 observations in twelve years. The 0–10 band has 34. Both fail the corrected threshold, both are excluded from every conclusion here, and both are exactly the kind of number that gets screenshotted — the same trap as the 4-hour Fibonacci level that ranked first of 131 on seven observations.

Limits

One instrument, one broker, 2014-01-14 to 2026-09-23, in a period gold spent mostly rising. A momentum result from a trending sample is the least surprising thing this could have produced, and it would be worth far more if it held in a flat or falling decade. RSI(14) only; other periods would shift the bands.

Forward windows overlap heavily, which inflates significance in a naive test — the bootstrap control draws with the same structure, but overlapping data still means fewer independent observations than the row counts suggest, particularly on the daily chart where some bands hold under 200. The full table for all three timeframes is published with our other studies, including the base rates and p-values that make each row checkable.

Common questions