RSI and Bollinger: two ways to buy a stretched price

Mean reversion starts from an observation: prices wander around a level, and a sharp move away from it is often partly undone. A mean-reversion strategy buys after a fall that looks too far and sells once the price has come back. The whole difficulty is telling a stretch that will snap back from the start of a real fall. A9 offers two versions, and they answer that question differently.

Bollinger reversion: a band around the average

Bollinger bands draw a moving average and two lines around it, set a number of standard deviations above and below. When prices are calm the bands are narrow; when they swing, the bands widen. A close outside the band is unusual by the price's own recent standard.

A9's Bollinger reversion, with the defaults on hourly bars:

• The middle line is the 40-bar average. The bands sit 2.5 standard deviations either side.
• When a bar closes below the lower band, the bot buys.
• When the price closes back at or above the middle line, it sells.
• It holds for at least 10 bars before it will exit, so that noise right after the entry does not cost a round trip.
• It is long only by default. Shorting above the upper band is available, but in A9's tests every short-enabled setting was among the worst, because selling a stretched rally often means selling a breakout.

RSI reversion: a short RSI on a long filter

The relative strength index (RSI) compares recent gains with recent losses on a scale from 0 to 100. Readings near 0 mean almost every recent bar fell. A9's version uses a very short RSI, two bars, which reacts to a single sharp drop.

With the defaults on hourly bars:

• It buys when RSI(2) falls to 10 or below.
• It sells when RSI recovers to 55.
• It buys only while the price is above its own 100-bar average. That is the filter.
• Shorting is on: it sells short when RSI rises to 70 or above and covers when it falls back to 45. The short side has no trend filter. A bot with shorting on trades the perpetual in your futures account, for both sides.

Why the filter matters

An unfiltered reverter buys every sharp drop, including the ones that are the start of a long fall. The trend filter changes the bet: it buys dips inside an uptrend and sits out a falling market entirely. In A9's tests on six symbols since 2022, the filtered setting had a positive median in each of four years, while the classic unfiltered version, used as the control, lost money overall. The filter is the strategy, not a detail of it.

How the two differ in practice

• Bollinger waits for an unusual close relative to recent volatility, and exits at the average. It trades less often and tends to hold through a slow recovery.
• RSI reversion reacts to one or two sharp bars and exits early at 55. It trades more often, for smaller gains each time, and the trend filter keeps it out of downtrends.
• Bollinger's protection against a falling market is the minimum hold and the Fear & Greed floor on crypto; RSI's is the trend filter.

On tokenized stocks both read only bars from the US regular session, 09:30 to 16:00 New York time, so that thin overnight prices do not create false signals.

When mean reversion loses

• A one-way fall. The price that looked stretched keeps going. The bot holds one entry and does not add, but the middle line it exits at moves down with the price, so the exit can come well below the entry.
• A strong rally, for the short side. Shorting into strength is how RSI reversion loses on a breakout.
• Fees on small moves. RSI reversion's exits at 55 often capture a fraction of a percent. Every trade pays the round-trip fee, and the backtest charges it.

The typical record is many small wins and occasional larger losses. A high win rate is expected and says little about the size of the loss you have not yet seen.

Reading the backtest

Compare the return with buy and hold over the same window, then read the maximum drawdown. Check how many trades the window held: a handful of trades is not enough to judge a strategy that is meant to trade often. If you used Auto-tune, read what the nearby settings earned; a result that collapses when a setting moves by a fifth leaned on luck.