Is Buying RSI Below 30 Profitable? 330 Settings Tested on USD/JPY

RSI compares how much price rose against how much it fell over the last N bars and turns that into a number between 0 and 100. Almost every description of it says the same thing: below 30 is oversold, above 70 is overbought. A second, quieter reading treats 50 as the dividing line: above it the buyers are ahead, below it the sellers are.

The first is a mean-reversion tool. The second is a trend-following one. The same line produces opposite rules, and the usual advice does not say which reading the settings belong to.

So both were measured side by side, on the same pair, the same years and the same costs. On USD/JPY in 2025, 150 fade settings run long and short separately (6 periods × 5 entry levels × 5 exit levels) plus 30 midline-cross settings (6 periods × 5 levels): 330 in all, on the fifteen-minute, hourly and four-hour charts, then repeated over 2024.

On the hourly chart the fade won 65.3% of its trades on average and survived both years in 31 of 150 settings. The midline cross won 23.7% and survived in 17 of 30. Not one cross setting won more than three trades in ten.

buy the oversold line (98)sell the overbought line (112)cross the midline (30)one-hour chart, 2025, 10+ trades

What RSI actually measures

Take the last N bars. Average the size of the up moves, average the size of the down moves, and express the up average as a percentage of the total. The only calculation parameter is period.

• Nothing but up bars gives 100
• Nothing but down bars gives 0
• Up and down in balance gives 50

So 50 is the point where gains and losses cancel, and distance from it measures how lopsided the recent bars have been. "Overbought" and "oversold" are an interpretation laid on top of that; the arithmetic only reports the imbalance.

A longer period does not reach 30

Longer periods create a practical problem. Averaging over more bars mixes the ups and downs together and pulls the number toward 50. The count below covers every hourly bar of 2025 on which RSI was defined:

everything the line touched all yearwhere eight bars in ten satvertical lines at 30 and 70; 6,226 hourly bars, 2025

A period-50 RSI never once closed below 30 in all of 2025. Its lowest reading was 31.2. "Use a longer period for cleaner oversold signals" describes something that does not happen: the signals do not get cleaner, they stop existing. On the four-hour chart both period 30 and period 50 stayed above 30 for the entire year.

The sweep shows the consequence directly: on the four-hour chart, 50 of the 150 long-fade settings never took a single trade, and 20 did nothing on the hourly chart either.

Putting RSI on a chart

Drawing it is the same everywhere. The decision that matters is whether the rule trades a level being reached or a cross through that level. This article tests the fade and the level-cross rule separately.

How this was measured

Enter the settings below and the same numbers come out.

The long fade and short fade are reported separately because a level condition is directional. "Buy below 30" and "sell above 70" produce entirely different results.

Mean reversion or trend following?

The higher-win-rate fade came through both years less often than the level cross.

(Mean win rate covers the 2025 settings that took at least one trade. The median is over every setting on that timeframe.)

The figure at the top plots the hourly rows one point each, leaving out settings that traded fewer than ten times: win rate across, net pips up. The red and purple fade settings spread across a 40–90% band; the blue cross settings sit in a narrow 17–30% band. The medians still come out at +830 pips for the cross, +11 for the long fade and −137 for the short.

Selling the overbought line is the weakest of the three, surviving both years in 9 of 150 hourly settings. USD/JPY rose 1,632 pips through 2024, and a rule that sells strength gets run over in a year like that.

Does a longer RSI period help?

Not uniformly: a longer period moves the fade and the level cross in opposite directions. The clearest result in the sweep is what the period does.

the oversold fade (25 settings each)the midline cross (5 settings each)share of settings profitable in both 2024 and 2025, hourly chart

Settings profitable in both 2024 and 2025, hourly chart:

The long fade had the most settings profitable in both years at period 30 (8 of 25), while the level cross had the most at period 9 (5 of 5). At period 50 those counts fell to 4 of 25 for the fade and 1 of 5 for the cross.

For the long fade, the profitable-setting count rose from 2 of 25 at period 7 to 8 of 25 at period 30, then fell to 4 of 25 at period 50. Longer periods made readings below 30 less frequent, and period 50 never fell below 30 on the 2025 hourly chart. This test does not identify why the count rose through period 30; it does show that the count halved when period increased from 30 to 50.

For the level cross, all five tested levels were profitable in both years at period 9. The count fell to 3, 2, 2, and 1 as period increased through 14, 21, 30, and 50. Period 9 traded 775 times a year on the hourly chart and period 50 traded 231 times, so fewer trades gave individual wins and losses more influence over annual net.

Levels behave more simply:

The fade's exit level survived best at 70 (9 of 30, against 6 at level 50 and 5 at 65). The back half of "buy at 30, sell at 70" holds up in this test even where the front half does not.

What the standard 14 / 30 / 70 did

The most-quoted settings, run both ways. The cross column uses the same period 14 with the level at 50.

Three of those nine rows came through both years: the long fade on M15 and H4, and the cross on H1.

The 2024 fifteen-minute row is the one worth staring at. 93 trades, 75.27% of them winners, +9.0 pips for the year. Three winners in every four, and what is left over after a year is nine pips.

On the hourly chart the standard fade fires 25 times a year. There is not much to say about a sample that size. A 60% win rate there means fifteen winners and ten losers.

The hourly cross took 638 trades for +1,272.0 pips, then 681 for +666.0 pips. The count is two orders of magnitude larger, and the year-to-year swing correspondingly smaller.

Why the higher win rate loses

The two sides of the trade are different sizes. Averaged over every 2025 setting:

The cross wins 3.5 times what it loses. At a 23.7% win rate, 0.237 × 71.5 − 0.763 × 20.5 still comes out positive. It rides a move as long as the line stays on one side of the level and reverses on the same bar when it does not, so a wrong call is cut early.

For the fade, the average win is only 0.73 times the average loss. It collects small profits when price returns to the exit level and takes a large loss when price does not come back. The holding times differ just as sharply: 188.9 hourly bars for the fade, about eight days, against 21.6 for the cross.

The same relationship holds on the other timeframes. The long fade ran +49.6 / −73.1 on M15 and +126.2 / −196.2 on H4; the cross ran +36.8 / −10.4 and +120.1 / −42.8.

Can you pick settings from last year?

The best fade of 2024 lost in 2025; the best cross fell but stayed profitable.

The period-14, 40 → 70 long fade made +1,139.8 pips in 2024 from 44 trades at a 77.27% win rate, then lost 798.8 pips in 2025. It also fell below the 2025 median of +11.2 pips.

The period-9, level-45 cross fell from +3,601.3 pips in 2024 to +1,254.9 pips in 2025, but remained above the 2025 median of +830.3 pips. Read the other way round, period 9 at level 40 made +1,815.4 pips over 672 trades in 2025 and +1,003.6 pips in 2024. These two cross settings were profitable in both years even though their annual net changed.

Filters, stops and targets

Two representative hourly settings with the usual additions bolted on.

The ADX filter left the cross worse off in both years. It does raise efficiency per trade (profit factor goes from 1.146 to 1.322 at ADX ≥ 25) but the total falls from +1,272.0 to +710.9 in 2025 and from +666.0 to +94.6 in 2024. Adding a trend-strength filter to the level-cross rule reduced annual net in both years.

The London and New York filter produced +1,669.9 pips in 2025 and −117.9 in 2024. Tokyo produced +284.3 pips in 2025 and +1,254.5 in 2024. Either session choice can appear superior when selected from only one year.

Only the widest pair beat the baseline in both years for the cross (+2,122.6 and +761.1). Tight stops helped in 2025 and turned 2024 negative. The 24-bar time exit produced the best single number in the table and the worst kind of evidence for it: 2024 came out below baseline.

Nothing improved the fade in both years.

Is the spread what loses the money?

For the level cross yes, for the fade no.

At zero spread the fade still loses 243.2 pips. Going from 0 to 2.0 pips costs it 50.0 pips: 25 trades times 2.0. Removing transaction cost therefore does not make the fade profitable; the trades themselves remain negative before spread.

The cross is the case the identity does describe. It pays 191.4 pips at 0.3 (638 × 0.3), 13% of its +1,463.4 gross, and 1,276.0 pips at 2.0, which leaves +187.4. The same arithmetic held exactly in the Bollinger Band and moving average cross runs.

Choosing RSI mean reversion because it wins often is not supported by this test. A separate RSI overbought and oversold duration test measures how often the line remains beyond 70 or 30 on the next bar. Williams %R also produced high win rates without correspondingly high annual net results. The Bollinger Band article compares a band-touch fade with a close breakout. QQE smooths RSI and Stochastic RSI runs a stochastic over it; both were measured the same way. Fibonacci retracements compare buying a price level with trading its break. CCI has no ceiling on its scale, and its fade lost in all six cells. If you want to check this on your own pair and your own window, the conditions can be assembled without writing code.

Notes

• One pair, two years. Nothing here promises the same behaviour on another pair or another year
• The fade holds until the exit level arrives, with no stop. A wrong call shows up at full size. Stops were measured separately
• Many fade settings trade too rarely to judge. The hourly median is 29 trades a year, and 50 of the 150 four-hour settings never traded at all. Win rates and profit factors on counts like those are wide numbers
• Entries and exits are taken at bar closes. Real fills differ
• Spread is a flat 0.3 pips throughout. Real spreads move with the session and around releases