Simple vs Sophisticated Volatility Models, Tested

Read almost any tutorial on volatility modeling and you'll meet the same ladder. On the bottom rung sits the rolling standard deviation — simple, equal-weighted, allegedly naive. Above it sits EWMA, which weights recent days more heavily. At the top sits GARCH, fitted, persistent, and mean-reverting — the family institutions use. The implication is that each rung improves on the one below, so serious work should climb.

The ladder is real. What the tutorials rarely say is that it ranks the models on one particular task — and that on a different, equally practical task, the ladder flattens. This article is about that second task: keeping a band on a chart honestly calibrated, one bar ahead. On that job, there is a published head-to-head result, and it's not what the ladder predicts.

Everything here is descriptive. A volatility model estimates how large a market's moves have recently been; nothing in this article predicts direction or constitutes a trade signal.

Three models, one question

All three models answer the same question — how volatile is this market right now? — and differ mainly in how they weight the past.

Rolling standard deviation treats the last n returns equally: every day in the window counts the same, and a day that leaves the window stops counting entirely. It has no parameters to fit beyond the window length. Our primer on what rolling volatility is covers the mechanics; the short version is that it's a plain average of recent squared deviations.

EWMA — the exponentially weighted moving average, best known from the RiskMetrics methodology — replaces equal weights with exponentially decaying ones. Yesterday matters more than last week, which matters more than last month. Nothing ever fully leaves the window; old days just fade. One decay parameter controls how fast.

GARCH goes further: it fits the weighting to the data. A GARCH model estimates, from history, how strongly volatility clusters and how quickly it reverts toward a long-run level. After a shock, a GARCH forecast rises and then decays back — it has an opinion about the future path of volatility, not just its current level.

Reading the list top to bottom, each model uses strictly more information than the last. That's the intuition behind the ladder, and for some jobs it's correct.

The winner depends on the task

Ask each model to do institutional risk work — forecast volatility over the next month, price the decay of a shock, feed a value-at-risk engine across horizons — and the differences bite. GARCH's mean reversion gives it a term structure: it can say "elevated now, calmer in three weeks," which a flat rolling estimate cannot. This is why the sophisticated end of the ladder earns its keep in risk management, and none of what follows disputes that.

But a band on a chart asks something narrower: given how this market has been moving, how wide is a typical next bar? One step ahead, over thousands of bars, judged by whether the stated coverage matches the realized coverage. That judgment has a name — calibration — and a natural score: the containment rate, the fraction of bars whose close lands inside the band (our comparison guide explains the test in full).

One-step-ahead calibration is a forgiving task for a simple estimator. Volatility is persistent — calm days cluster with calm days — so "tomorrow will be about as volatile as the recent past" is already most of the answer. The sophistication that separates the models at long horizons has little room to matter one bar out.

What the published head-to-head found

Oisigma's working paper ran exactly this comparison. The same return-space band construction was scored on the same data with three different variance estimators inside it: the plain rolling estimate that the Behavioral Transform Model uses, a GARCH model, and RiskMetrics-style EWMA. On the one-day-ahead calibration task, the three were statistically indistinguishable. As the Proof page summarizes it: sophistication buys essentially nothing here; the construction is what matters.

The construction point is the sharp end of the finding. Scored by the same causal question, a standard price-space Bollinger Bands® setup at its (20, 2) default covered about 83% of next closes against roughly 94% for the return-space band at the same two-sigma width — and the paper is explicit that this gap comes from the return-space construction, not from the choice of variance estimator. Swapping the estimator among rolling, GARCH, and EWMA moved the result within statistical noise; swapping the measurement space moved it by double digits. (Why measuring returns rather than price levels matters this much is its own topic — the return space vs price space deep dive covers it.)

For context, the calibration record being matched is the one published on the Proof page: historically, the inner band contained the next close about 71% of the time and the outer band about 94%, measured across 97 years of S&P 500 daily data and 40 instruments in five asset classes — past behavior is not a guarantee of future results.

One more detail from the paper is worth knowing, because it shows GARCH earning its keep as a research tool in the same study. The measured ~71% sits slightly above the correct Gaussian finite-sample benchmark, and the paper found that a heavy-tailed GARCH simulation reproduces that excess while a Gaussian one does not. Sophistication didn't improve the band — but it helped explain, in research, why real markets behave the way they do.

Where the equivalence stops

The published result is about average calibration, and it's scoped to the three estimators actually compared. It is not a claim that the models are interchangeable everywhere.

The clearest case is a fast volatility shock. EWMA and GARCH weight recent observations more heavily, so by construction they react faster than an equal-weighted rolling window when volatility jumps. The paper measured the onset cost for Oisigma's own band — containment fell to about 65% on the most extreme crisis-onset days, a limit we publish rather than bury — but it did not measure the corresponding number for the other estimators, so no comparative claim is available there. The crisis-onset article covers that failure mode in depth.

There's a practical trade on the other side of the ledger. A rolling estimate is auditable at a glance: sixty returns, one average, no fitted parameters, nothing that changes when a model is re-estimated. Anyone can recompute it and get the same number — which is precisely what makes a calibration claim about it easy to check. A fitted model's output depends on estimation choices that a chart user can't see. When the fitted machinery adds nothing measurable to the task at hand, the simpler object is easier to verify, and easier to trust for that reason.

The takeaway, scoped honestly

So: is GARCH better than rolling volatility? For multi-horizon forecasting and risk-engine work, the ladder stands. For one-day-ahead band calibration — the task a chart band actually performs — the published evidence found no measurable difference among the three estimators tested, and found instead that the choice people rarely discuss, measurement space, is where the calibration was won or lost.

That's why the Behavioral Transform Model keeps its volatility estimate deliberately plain. Not because sophistication is bad, but because on this task it was tested and bought nothing — and a simple construction whose calibration record is published is easier for anyone to check than a fitted one.

Frequently asked questions

Is GARCH better than a simple rolling standard deviation? It depends entirely on the task. For multi-horizon forecasting and risk-engine work, GARCH's mean reversion gives it a term structure a flat rolling estimate cannot produce. On the one-day-ahead band calibration task, the published head-to-head found the three estimators tested statistically indistinguishable.

Does EWMA react faster than a rolling window when volatility spikes? Yes — by construction. EWMA and GARCH weight recent observations more heavily, so they register a shock sooner than an equal-weighted window does. The paper measured the crisis-onset cost for Oisigma's own band only, so no comparative onset figure for the other estimators is available from it.

Why doesn't the Behavioral Transform Model use GARCH? Because on this specific task the fitted machinery was tested and added nothing measurable, and a plain rolling estimate is auditable at a glance — no fitted parameters, nothing that shifts when a model is re-estimated. Anyone can recompute it and arrive at the same number.

Does this mean sophisticated volatility models aren't worth using? No, and that isn't what was tested. The equivalence result covers average one-day calibration for three estimators inside the same return-space construction. It says nothing about longer horizons, value-at-risk work, or volatility term structure, where the sophisticated end of the ladder earns its keep.

If you'd like to see what a deliberately simple, calibrated band looks like on your own charts, BTM runs on TradingView with a free 30-day trial — $15/month after, cancel anytime. The published record is on the Proof page; the band on your own symbols is the rest of the argument.