When Volatility Bands Fail: The Crisis-Onset Problem

Every volatility band looks good in a calm market. The honest question is what happens in the worst week — when volatility doesn't drift higher but jumps, overnight, to a level recent history has never seen. For a band built from recent price behavior, the answer is structural: it can only widen after the new volatility starts showing up in its data — so at the onset it runs too narrow, precisely when being too narrow costs the most.

This article explains why that happens, what the failure actually looks like when you measure it, and why we think publishing this number matters more than the flattering ones. As always, everything here is descriptive — an expected range describes where price has tended to trade; it does not predict direction, and nothing here is a trade signal.

Every rolling band is built from the past

A volatility band — any volatility band — works by measuring how much price has recently been moving and drawing a range scaled to that measurement. Oisigma's indicator samples a rolling window of recent returns; other designs use a moving average of price dispersion or an exponentially weighted scheme. The construction details differ, but they share one structural property: the width of today's band is a function of yesterday's behavior.

Most of the time, that's exactly what you want. Volatility is persistent — calm days cluster with calm days, volatile days with volatile days — which is why a range built from the recent past stays well calibrated over long stretches. But persistence is an average property, and averages have exceptions. The exception that matters most is the fast crisis.

What failure looks like, measured

When a shock arrives — a crash day, a pandemic headline, a sudden macro break — realized volatility can multiply in a session or two. A rolling window can't know that yet. If the window is sixty bars long, fifty-eight of those bars still describe the old, calmer market. The band widens as the new days enter the window, but for the first days of the transition it is provably too narrow: it describes the market that existed last month, not the one that exists this morning.

Oisigma's working paper measured what that lag costs. Across the tested history, the inner band contained the next close about 71% of the time overall, and roughly as consistently decade by decade — but on the most extreme crisis-onset days, containment fell to about 65%. These figures are historical, measured in research; past behavior is not a guarantee of future results. The full context, including the decade table and the 40-instrument results, is on the Proof page.

Two things are worth noticing about that number. First, the failure is real but bounded: even on the days engineered to be hardest for the method, containment degraded by around six percentage points rather than collapsing. Second — and this is the subtle part — the failure is invisible at coarser resolution. Decade-level containment held through the Great Depression, 1987, 2008, and COVID, because onset days are rare and the band re-adapts within days. The weakness lives specifically at the onset: the first days of a volatility jump, before the window catches up.

Doesn't a shorter window fix it?

The obvious fix is to shorten the window so the band reacts faster. It helps at the onset — and costs you everywhere else. A short window doesn't just react quickly to real regime changes; it reacts quickly to noise. The band whipsaws in ordinary markets, widening after every large day and tightening after every quiet one, which degrades the very stability that makes a range readable. Window length is a trade-off between responsiveness and stability, not a free parameter with a correct answer. No setting removes the structural fact that a backward-looking measure lags a forward jump.

What about more sophisticated machinery? Models from institutional risk management — GARCH-family and exponentially weighted approaches — put extra weight on the most recent observations, so by construction they react to a volatility shock faster than a plain rolling window does. Faster is not immune, though: even the quickest-reacting model can only respond after the first extreme days have entered its data. And on the average calibration task, the working paper found GARCH, EWMA, and the plain rolling window statistically indistinguishable. How much faster reaction helps at the onset specifically is an empirical question we haven't measured for those designs — the onset figure above is ours, measured on our own band, and it's the only one we'll quote.

The other honest limit: deep tails

Crisis onset is one of two published limits worth knowing. The other is that the outer band is slightly optimistic in the deep tails: the very largest moves are somewhat more common than the band's nominal coverage implies. That is the fat-tails property of market returns showing up exactly where you'd expect it to — in the extremes. (For what fat tails are and why they matter, see Is this price move normal?, which covers the statistics in more depth.) A calibrated range is accurate on average, not in every moment, and the deep tail is where "not every moment" concentrates.

Why we publish the failure mode

Here is the uncomfortable part for the indicator category as a whole: the onset lag is structural to any band estimated from realized history, though how much each design suffers is an empirical question — and we have only measured our own. What you will rarely find is a vendor who has measured it on their own tool and published the number. That's a claim you can verify yourself — look for a published crisis-onset containment figure, or any published containment figure, in the documentation of the band tools you use.

We publish ours because the alternative is worse. A range whose limits are undocumented invites exactly the wrong confidence at exactly the wrong time: a trader leaning on a band during a fast crisis, unaware that this is the one regime where the band is known to run narrow. Stated plainly: the band is at its least reliable on the days when markets are at their most dangerous. Knowing that is worth more than a screenshot of the band behaving well.

Reading a band during a crisis, descriptively

None of this makes a band useless in volatile markets — it changes what a breach means. In ordinary conditions, a close outside the range marks a statistically unusual day. In the first days of a fast crisis, breaches cluster: the band is still sized to the old regime, so consecutive closes can land outside it while the window adapts. Descriptively, that cluster is itself information — it's what regime change looks like through the lens of a calibrated range. Some users treat a run of consecutive breaches as a cue that recent history has stopped describing current conditions, and that wider uncertainty applies to everything else on their chart too. That is context for your own judgment, not a signal, and it's the frame in which the model's markers are meant to be read.

The takeaway

Volatility bands fail at crisis onset because they are built from the recent past, and a fast crisis is precisely a break with the recent past. The lag is structural; its size is an empirical question each vendor can only answer for their own tool. For ours, the measured record shows a bounded step down — around six percentage points off the long-run containment rate on the most extreme onset days, historically — concentrated in the first days before the window adapts. A tool can't remove that limit. It can only measure it and tell you, or not.

If you'd rather use a range whose limits are published alongside its strengths, Oisigma's BTM indicator is available for TradingView with a free 30-day trial (then $15/month, cancel anytime). Watch how the band adapts on your own charts — including on the days it's slowest to. Start your free trial →