Volatility Band Lookback Period: What 5 to 500 Bars Show

Every volatility band has a lookback setting, and most of the discussion about it stays qualitative. The short version, which what is rolling volatility covers properly, is that the window sets the estimate's memory: shorter windows react faster and are noisier, longer windows are steadier and slower to register a change.

That trade-off is true, but it does not say what the window does to the one number a band is actually claiming: how often the next close lands inside it. That question has a measured answer. The working paper behind the Behavioral Transform Model swept the window from 5 bars to 500 on the S&P 500 and scored every length the same way. This article walks through what came out, and why the obvious reading of the numbers turns out to be wrong.

What was measured

The setup is the paper's standard one. A band is built from the rolling mean and standard deviation of recent daily returns, projected one standard deviation either side of the prior close, using only data available at that close. Each next close is scored inside or outside. The sweep repeated this on S&P 500 daily data from 1927 to 2024 at eight window lengths: 5, 10, 20, 30, 60, 100, 200 and 500 bars, each scored on its own maximal valid sample. The full table is in Appendix A.8 of the working paper.

The raw numbers rise with the window

Read straight off the table, one-standard-deviation containment rose smoothly as the window lengthened: 58.6% at 5 bars, 64.6% at 10, 68.1% at 20, 69.4% at 30, 71.2% at 60, 72.3% at 100, 73.5% at 200 and 75.3% at 500. The 60-bar cell is the paper's headline figure. All of these are historical measurements on one index, and past behavior is not a guarantee of future results.

The tempting conclusion is that short windows produce bands that are too narrow and long windows produce better ones. The paper rejects both halves of that reading, and the reason is the target each window should be compared against.

Against the right target, short windows were not too narrow

A one-standard-deviation band is usually measured against the textbook 68.27%. That figure assumes the mean and standard deviation are known exactly. A band built from a handful of recent returns is estimating both, and the estimation error lowers the coverage a correctly built band should be expected to produce. The finite-window benchmark and why it replaces the textbook figure are explained in how to backtest a band indicator; what matters here is that the benchmark moves with the window.

The paper tabulated that benchmark at every length in the sweep: 58.7% at 5 bars, 63.5% at 10, 65.9% at 20, 66.7% at 30, 67.5% at 60, 67.8% at 100, 68.0% at 200 and 68.2% at 500. Set against it, the 5-bar band's 58.6% sat 0.1 points below its own target, essentially on it. The paper's reading is direct: the shortest window did not under-cover, and the earlier intuition that short windows under-cover was an artifact of comparing them with 68.27% rather than with their own null.

A 5-bar band is noisy, and it contains fewer closes. But it contained about as many as a band estimated from five returns should, historically. Low raw coverage at a short window is mostly the target falling, not the band failing.

What grows with the window is the excess

Subtracting each window's benchmark from its observed rate gives the excess: −0.1 points at 5 bars, then +1.1, +2.2, +2.7, +3.7 at 60 bars, +4.5, +5.5 and +7.1 at 500. The excess, not the raw rate, is what rises steadily with the window.

The paper attributes that excess to two effects that compound. The first is the shape of real returns: they cluster more tightly around the middle than a normal distribution does, with fatter tails, so a one-standard-deviation band catches a few more closes than the Gaussian target implies. That effect is present at every window. The second appears at long windows: a 200- or 500-bar estimate spans calm and turbulent stretches together and averages them, producing a band that is too wide on average for whichever regime the market is actually in. The paper calls this over-smoothing.

So the 75.3% at 500 bars is not a sign of a better band. Part of it is a band carrying volatility from a regime that has already ended. A higher raw number and a better-matched band are different things, and the sweep shows them separating as the window lengthens.

Why the canonical window is 60 bars

The paper states three reasons for fixing its window at 60 trading days: it lies within the calibration plateau of the sweep, it is short enough to react to a regime change within about one calendar quarter, and it is a round number near the middle of that plateau, which eases reproducibility. The appendix adds the reading behind the first reason: at 60 bars the excess is still mainly the fat-tailed excess rather than over-smoothing. The indicator's default window is 60 bars, and the rationale is summarised on the FAQ page.

None of this makes 60 the correct window for every purpose, and the paper does not claim it is. The window counts bars, not calendar time, so the same 60 bars span very different stretches on a weekly or monthly chart; how the paper scored that, and a shorter window matched to calendar time, is covered in daily vs weekly charts.

Where the evidence stops

The sweep is one instrument, the S&P 500, on daily bars, for the inner band only. The paper's cross-asset results were measured at the 60-bar window, not across the sweep, so nothing published shows how the eight window lengths behaved on currencies, commodities or crypto.

The sweep measures average calibration over nearly a century. It does not say how each window behaved at the onset of a fast volatility spike, where any backward-looking window lags in varying degrees; the published onset figure is for the 60-bar band only and is discussed in when volatility bands fail.

It also says nothing about whether any window length makes a band more useful for a particular way of trading. The paper validates the range's calibration, not the profitability of any use of it, at any window, and whether any use delivers value after costs is an open question.

The takeaway

Lengthening the lookback raised a band's raw containment from about 59% to about 75% on a century of S&P 500 data, historically. Read against the correct target for each window, that rise says something different: short windows sat on their own benchmark rather than under it, and what grew with the window was an excess over the target, driven by fat-tailed returns and, at long windows, by averaging across regimes. The window changes what a band is measuring and what it should be expected to contain. It does not turn a noisy band into an accurate one or a wide one into a better one.

Frequently asked questions

Is it a good idea to pick the lookback that gives the highest containment? On this evidence, no single raw number can settle it, because the highest containment in the sweep came from the longest window, where part of the excess was over-smoothing across regimes. Choosing a setting because it produced the best-looking result on the history being examined is also a form of fitting to that history, the problem set out in overfitting trading indicators. The sweep reports what each length measured; it does not rank them.

Does changing the lookback period make a band repaint? Changing any input recalculates the whole band on every historical bar, so past lines move to where the new setting would have drawn them. That is a recomputation under a different setting, not repainting. Under a fixed setting, the indicator's range and markers for a closed bar are fixed once that bar closes, as described on How It Works; only the live marker on the still-forming bar can update.

Do the window results apply to crypto, forex or single stocks? They were measured on the S&P 500 only. The paper's tests on other asset classes used the 60-bar window throughout, so whether the same shape across 5 to 500 bars holds elsewhere is not something the published record covers. The procedure is simple enough to repeat on any series of closes, which is the subject of the backtesting article linked above.

Why does a longer lookback make a band lag more in a crash? Each new day is one observation among the window's full count, so a single large move shifts a 500-bar estimate far less than a 20-bar one, and it takes longer for a turbulent stretch to dominate the window. That is the same property that makes long windows steady in calm markets. How much the 60-bar band's coverage fell at crisis onset is covered in when volatility bands fail.

The whole sweep, with the null for every window, is in the working paper, and the 60-bar band it describes is what BTM draws on a TradingView chart. If you'd like to see how that window reads the markets you follow, you can start a free 30-day trial.