How Accurate Are Bollinger Bands? A 97-Year Audit

Ask the internet how accurate Bollinger Bands® are and you'll get answers like "65–70% when used correctly" or a video promising 93.5%. Those numbers describe trading strategies built on top of the bands — entry rules, exits, filters — which means they measure the strategy, not the indicator. They also tend to arrive without a test you could rerun yourself.

There's a more useful way to pose the question, because a band indicator makes a claim you can actually check. This article checks it, against 97 years of daily S&P 500 data. The result: at the canonical settings, the band's real-world containment runs meaningfully below what its construction implies — and the size of the gap depends heavily on which settings you use, which is why any Bollinger accuracy figure quoted without its settings is close to meaningless.

What "accurate" should mean for a band

A band indicator isn't a buy/sell system, so a win rate is the wrong yardstick. What a band actually asserts is statistical: price should stay inside me a certain percentage of the time. Testing that assertion is called a containment test — you count, bar by bar through history, how often the next close landed inside the band. We've written a full guide to comparing band indicators on the containment metric if you want the method in depth; here we just apply it.

The advantage of this yardstick is that it's objective and rerunnable. No strategy assumptions, no cherry-picked chart windows — just a claim and a count.

The claim built into the default settings

Bollinger Bands® at their canonical settings — a 20-period simple moving average with bands at ±2 standard deviations of price, usually written (20, 2) — carry an implicit statistical claim. Under the textbook assumption those settings gesture at, a ±2 standard deviation interval should contain roughly 95% of observations. That's the number many traders have in their heads when they treat a touch of the band as "stretched" or read %B as a probability.

So the checkable question becomes: does the (20, 2) band actually contain ~95% of closes?

What 97 years of data show

Our working paper ran this audit on daily S&P 500 closes from 1928 to 2024 — about 24,000 trading days — and across a 40-instrument universe spanning equities, FX, commodities, rates, and crypto. Two results matter for this question, and each comes with its settings attached:

At the canonical (20, 2) settings, standard Bollinger Bands® contained the next close about 83% of the time — versus the ~95% the ±2SD construction nominally implies. A 12-percentage-point shortfall on the band's own terms, persistent across the sample. These figures are historical calibration measurements; past behavior is not a guarantee of future results.

At matched settings, the gap is starker. When the paper matched Bollinger Bands® to the same window and a ±1 standard deviation width used by a return-space band (so both bands make the same nominal claim on the same data), the price-based Bollinger construction contained only ~42% of closes. That ~42% figure applies only to the ±1SD matched configuration — it is not the (20, 2) number, and quoting it without the settings would be misleading in the other direction.

The point of showing both: "how accurate are Bollinger Bands" has no single answer. It depends entirely on the settings, and the honest way to report any band's accuracy is with the configuration stated. The full audit tables are on our Proof page.

Why the gap exists

The shortfall isn't random noise; it has a mechanism, and it's the construction. Bollinger Bands® measure a moving average and standard deviation of raw price levels. In a trending market, the 20-period average of price lags behind where price currently is, so the band's center drifts off-center — price rides one edge while the far edge trails uselessly behind. At the same time, the standard deviation of trending prices inflates with the trend itself, widening the band for the wrong reason. Off-center plus mis-scaled adds up to more closes escaping than the ±2SD label suggests. (The drift mechanism deserves its own article, and it will get one — this paragraph is the short version.)

Fat-tailed returns account for part of any band's shortfall versus Gaussian arithmetic — real markets produce more extreme days than the textbook assumes — but the head-to-head result shows construction is the larger culprit: a band measuring the same volatility in return space, tested on the same data at the same nominal width, held ~94% at 2σ. The comparison is published, on the containment metric only, in the working paper.

Does this mean Bollinger Bands are useless?

No — and the audit doesn't say that. The shortfall matters if you read the band as a probability statement: if a touch of the outer band means "rare event" to you, it's worth knowing that at (20, 2), historically about one close in six landed outside, not one in twenty. Traders who use the bands as a relative framing of high and low, without attaching probabilities, are making a different and more modest use of them.

Our position is narrower than "Bollinger Bands® are broken." It's this: any band's accuracy is a testable claim, most vendors never publish the test, and we think traders deserve the numbers. We publish ours — misses included, like the model running too narrow at fast crisis onsets — on the Proof page, and the audit above is exactly the kind of test we invite on our own band. If you're weighing a replacement, the Bollinger Bands® alternative page covers that comparison directly.

The takeaway

Skip the win-rate folklore. Bollinger Bands'® accuracy, measured the way the indicator's own construction invites — containment against 97 years of data — is ~83% at the canonical (20, 2) settings versus ~95% nominal, and ~42% in the matched ±1SD configuration. Historical figures, settings stated, rerunnable from the published methodology; past behavior is not a guarantee of future results.

If you'd like to see what a band looks like when its containment claim is published and audited before you're asked to trust it, the Behavioral Transform Model draws a calibrated expected range on any TradingView chart — descriptive context, no signals. You can start a free 30-day trial and run your own eyes over the calibration on the markets you actually trade.