Does Combining RSI and Bollinger Bands Help? 720 Backtests

"Buy when RSI is below 30 and price touches the lower Bollinger band." Stacking two indicators like this is the most common way to build a rule out of two.

What the operation does is not in doubt. Combining creates no signals. It takes the bars where RSI is below 30 and discards the ones that did not touch the band.

On the hourly chart in 2025, RSI closed below 30 on 356 bars. 216 of them, 60.67%, also touched the lower band at two standard deviations. The other 40% are thrown away, and that is what "filtering" means here.

Throwing bars away moves the win rate on its own, because a smaller sample sits further from its own average. So "the win rate went up when I combined them" only means something if it went up more than discarding the same number of bars at random would have.

That is the measurement. Of 108 pairings, 21 fall outside the 95% range of random discards, and 19 of those 21 fall below it.

long (RSI below 25, 30, 35)short (RSI above 75, 70, 65)USD/JPY hourly 2025, 18 pairings

Settings for RSI and the Bollinger band

RSI runs between 0 and 100 in a pane under the chart. The Bollinger band draws a moving average with lines a multiple of the standard deviation above and below it, over the price itself.

Both are states rather than events. RSI below 30 holds for as long as it holds; it does not fire once.

Using both at once

Switch both on and set the logic to require all conditions. Only bars where both point the same way remain. RSI below 30 and a lower-band touch both produce long signals only, so the pairing is long-only; the short pairing is built from RSI above 70 and an upper-band touch.

How this was measured

All three use the same ten-bar exit. The comparison is between three entries, and different exits would turn it into a comparison of exits.

The random range is built like this: if 356 bars are below the RSI level and 216 of them also touch the band, then 216 bars are drawn from the 356 at random, a thousand times, and the resulting win rates are sorted to take the 2.5th and 97.5th percentiles. The generator runs from a fixed seed, so the published range is reproducible.

Against the same number of bars picked at random

Hourly, long side. The pool is bars below the RSI level; the pairing is the subset that also touched the band.

All nine 2025 pairings land inside the range. Three of the 2024 pairings land outside it, and all three land below (47.97% against 48.78-52.85%, 40.48% against 44.05-57.14%, 45.41% against 45.95-55.14%).

Notice how the range widens. Keeping 94.64% of the pool gives a range 3.78 points wide; keeping 37.50% gives one 21.43 points wide. The harder the filter, the more the win rate is expected to move on its own.

Across three timeframes and both sides there are 108 pairings. 21 land outside the range: 19 below and 2 above. Requiring the same timeframe, side and setting to land outside in the same direction in both years leaves 4 of 54, all of them on the hourly short side and all of them below.

What the band condition selects

If the pairing lands low, the band must be selecting a particular kind of bar. Splitting the RSI pool by whether the bar touched the band:

Run-up is how far price had already travelled into the signal over five bars, in ATR: for a long, how far it had already fallen.

The touched half had run 1.80 to 2.82 times as far, on bars 1.43 to 2.03 times as large. The mean RSI of the two halves differs by less than 2.1 points.

So the band is not selecting bars that are more oversold. At the same RSI reading, it selects the ones falling faster. In 8 of the 12 cells the touched half had the lower rate ten bars on. On the hourly chart in 2025 the short side reads 32.97% against 54.79%, a gap of 21.82 points.

The 4-hour long pool in 2025 is left out of the table: only 11 bars there did not touch.

Did the win rate rise?

Hourly, the RSI level alone against the pairing:

Five of the 18 improved, by between 0.02 and 3.67 points, and all five are long. All nine short pairings fell, the largest by 10.39 points.

How much was discarded does not predict the change. Keeping 95.1% gives -1.50; keeping 53.4% gives -5.98.

Trading all 720 runs

Entering at the close and exiting ten bars later, hourly long side:

Seven of the 15 cleared both years. The rest of the grid:

Only the hourly long side and the 4-hour short side produced anything. In the other four cells nothing survived, alone or paired.

The two cases where combining rescued a losing level

Five of the seven survivors were already profitable in both years using the RSI level alone. Only two turned a level that lost a year into one that did not.

Read alone, that looks like the filter earning its place. But both sit inside the random range: the 2024 pool below RSI 35 won 48.78%, the 321 bars kept at 2σ won 47.35%, and the random range is 45.17 to 52.65%. Discarding the same number of bars at random reaches the same place.

The short fade does not get fixed by combining

The short side lost almost everywhere, alone and paired. On the hourly chart in 2025 all 15 finished negative, the shallowest being RSI above 65 at -324.6 pips.

This series has measured the short fade breaking in a rising market several times, and 2024 rose from 140.87 to 157.20. But 2025 opened at 157.23 and closed at 156.67, and the short side still lost in all 15 hourly runs. Direction alone does not account for it.

Combining does not soften it either. Hourly 2025, RSI above 70 alone returns -1,040.3 pips; adding the 2σ touch gives -1,158.0, and the win rate falls from 42.47% to 32.08%.

Cost

Spread paid matched trade count times 0.3 pips in all 18 hourly runs.

Combining does cut the bill. RSI below 35 takes 145 trades and pays 43.5 pips; adding the 2.5σ touch takes 87 and pays 26.1. The 17.4 pips saved cost 58 trades.

Related articles

• Is buying RSI below 30 profitable?: the pool used here, tested on its own across 330 settings
• Bollinger Bands: mean reversion or breakout?: the band on its own, across 315 settings
• Candlestick pattern win rates: another case where the answer only appears once a control is placed beside it
• Is the engulfing candle profitable?: filters added to an entry, measured the same way
• Do Trend Filters Remove False Signals? 1,596 Backtests: the random range built here, applied across all 38 published settings

Notes

The data is USD/JPY from 1 January 2024 to 31 December 2025. Fills are at the close, the spread is fixed at 0.3 pips, and the lot is 0.1. RSI uses a period of 14 and the band a period of 20. The random range comes from drawing the pairing's own number of bars out of the pool a thousand times and taking the 2.5th and 97.5th percentiles of the resulting win rates; the generator uses a fixed seed. Where the pairing keeps the whole pool there is nothing to draw, so the pool's own rate is used. Because the horizons run out to 20 bars, that many bars at the end of each year are excluded from the counts.