Bollinger Bands and RSI: A Defined Forex Mean-Reversion Framework

Mean reversion needs a reason to expect a return

The hypothesis is that a short-term price stretch sometimes reverses toward a recent reference level. Bollinger Bands describe distance from a moving average using recent dispersion; RSI describes the balance of recent gains and losses. Both are calculated from price, so their agreement is not two independent pieces of evidence.

A low RSI or a lower-band touch is not automatically a buy. Prices can follow a band through a persistent trend, and closes outside the bands can occur as a continuation develops. John Bollinger’s own rules caution against treating band tags as standalone signals. Official Bollinger Band rules.

The framework below waits for a completed return inside the bands and applies a defined low-slope filter. These choices make the hypothesis testable; they do not establish that it has a profitable edge.

Fix the indicator calculations

Use 20 completed daily closing prices. Their simple average is the middle band. Calculate the population standard deviation of those same 20 prices, dividing squared deviations by 20, and place the outer bands two standard deviations above and below the mean. Set any indicator shift to zero. MetaQuotes band calculation.

Use RSI(14) with Wilder-style average gains and losses: initialise with 14 changes and subsequently apply (13 × prior average + new value) ÷ 14. RSI = 100 − 100 ÷ (1 + average gain/average loss). Define all-gain, all-loss and completely flat cases consistently; use 50 for the flat case in this reference implementation. RSI documentation.

A two-standard-deviation band is not a promise of 95% future containment. Price observations are not independent normally distributed draws, and the rolling mean and dispersion change over time.

State the confirmation and regime filter precisely

1. Use one daily-session convention and evaluate only completed bars.
2. A long candidate requires yesterday’s close below yesterday’s lower band with yesterday’s RSI below 30, followed by today’s close back between today’s bands.
3. A short candidate requires yesterday’s close above yesterday’s upper band with yesterday’s RSI above 70, followed by today’s close back between today’s bands.
4. On the confirmation day, require |SMA50 today − SMA50 ten bars ago| ÷ ATR14 today ≤ 0.5. Use Wilder-style ATR and skip invalid or zero denominators.
5. Enter at the next defined session’s tradable ask for a long or bid for a short. Permit one position per pair.

The slope filter, thresholds and periods are illustrative research choices. Fix them before testing. A low-slope reading can still occur just before a breakout; it does not identify a permanently stable range. Do not select a threshold by inspecting which losing historical trades it removes.

Define the stop, target and maximum holding time

At entry, freeze the target at the confirmation day’s SMA20. Skip the trade if the executable entry has already reached or passed that target in the intended direction. For a long, place the stop at the lower of entry minus 1.5 ATR14 and the lower of the two setup-bar lows minus 0.1 ATR14. For a short, use the higher of entry plus 1.5 ATR14 and the two-bar high plus 0.1 ATR14.

The first protective stop or target fill exits the trade. If neither occurs during ten completed daily holding bars, exit at the first tradable quote after the tenth bar. The target and initial stop remain fixed; no averaging down, added positions or discretionary widening is allowed in this baseline.

Size so the estimated loss, including execution costs, fits an illustrative 0.25% equity budget. Cap aggregate initial stop risk at 0.75%, scale simultaneous candidates proportionally and round down. These numbers define the example, not universal safe limits.

Worked example with a fixed target

Suppose EUR/USD closes below its lower band with RSI at 27. The next day it closes back inside; the slope ratio is 0.3 and therefore passes the declared filter. The confirmation-day SMA20 is 1.1000, ATR14 is 0.0040 and the lowest setup-bar low is 1.0900.

The following session’s ask is 1.0940. Entry minus 1.5 ATR is 1.0880; the setup low minus 0.1 ATR is 1.0896. The rule chooses the lower level, 1.0880, producing a 60-pip stop. The frozen 1.1000 target is also 60 pips away.

For USD 10,000 equity and a USD 25 risk budget, assuming USD 10 per pip per lot and USD 20 estimated costs per lot, volume is 25 ÷ 620 = 0.04032 lots, rounded down to 0.04. A target fill earns about USD 24 before costs; a stop loses about USD 24 before costs. The apparent 1:1 chart ratio becomes less favourable after charges.

Read the research without overstating it

Coakley, Marzano and Nankervis examined a large universe of technical FX rules, including Bollinger and RSI families. Their paper reports that data-snooping controls reduced the number of robust rules, while some indicator-based rules remained profitable within their sample. How profitable are FX technical trading rules?

That is more nuanced than either claiming technical analysis always works or claiming every tested rule failed. It also does not validate the combined confirmation, slope filter, ATR stop and time exit defined here. This exact framework needs its own evaluation under executable retail costs.

Keep every attempted variant in a research log. Trying many combinations of periods, RSI thresholds and filters and reporting only the best makes accidental patterns look persuasive. The final evaluation sample should remain unavailable for selecting those choices.

Test execution and ambiguous candles conservatively

Begin by checking a handful of signals against independently calculated indicators. Confirm that yesterday’s outside close is compared with yesterday’s band, not a band recalculated using today’s price. Ensure the trade cannot execute before the confirmation candle is complete.

Use bid/ask history where possible. If both stop and target fall inside one daily candle, the candle alone cannot establish which was reached first. Use appropriate intrabar data or state a conservative assumption rather than recording every ambiguous candle as a win. Apply gaps and stop slippage.

Include spreads, commissions, financing, conversion and missed trades. Report net expectancy, drawdown, average win/loss, time exposure and results in trending versus quieter periods. Assess sensitivity to modestly wider spreads and nearby parameter settings. Reject a conclusion that depends on a few favourable fills, an exact threshold or one selected currency pair.

Know when the model is vulnerable

Mean reversion is exposed when a new policy expectation changes the price level the market considers reasonable. A currency that appears cheap relative to its 20-day average can continue falling as investors reassess rates, growth or risk. The moving reference is descriptive, not fair value.

Entry confirmation may reduce some premature fades, but it can also create a late entry with little remaining distance to the target. A fixed target keeps this baseline auditable while sacrificing adaptation to a changing average. The ten-bar time exit limits waiting, but can close immediately before a later reversal.

Use demo to verify rule adherence and execution assumptions. Define event, data-quality and drawdown suspension rules before any live experiment. If a stop is reached, accept the recorded outcome. Buying progressively more because the next band reading looks even more extreme changes the strategy into an uncontrolled averaging process.