Term Structure Beginner Realised vol over a moving window Backward-looking

Rolling Volatility

Three different window lengths give three different answers, and all of them are correct.

Quick answer: Rolling volatility is the realised volatility of an underlying computed over a moving window of the most recent days, re-estimated each day as the window slides forward — so its level depends as much on the window length you chose as on what the market actually did.

In simple words

Rolling volatility answers what has volatility been lately, where lately is a window you pick — the last 10 days, the last 20, the last 60. Each day you drop the oldest return, add the newest, and recompute the annualised standard deviation. The catch is that the answer depends heavily on the window. A 10-day window reacts almost immediately to a sudden shock and spikes the highest, because a single violent day is a large fraction of just ten. A 60-day window barely notices the same shock, because one day is a small fraction of sixty, but it stays elevated for months afterwards. Both are correct answers to what has volatility been.

So the window length is not a technical detail you can leave on default; it is the actual question you are asking. A short window asks what is volatility right now and gives a jumpy, responsive answer. A long window asks what has volatility been over the season and gives a smooth, slow one. Neither is more true than the other — they are measuring different things — and a lot of confusion about volatility comes from two people using different windows and arguing about whose number is right.

Not to be confused with: Implied volatility, which is forward-looking and extracted from option prices. Rolling volatility is backward-looking and computed from past returns of the underlying itself — no options involved. People compare a rolling historical volatility against an implied volatility to judge whether options are cheap or dear, but the two are different objects: one is a measurement of a past that happened, the other a price for a future that has not.

The picture

Same market, three windows, three stories

10-day, 20-day and 60-day rolling realised volatility of NIFTY through a single shock, spot 24,000.

The three lines respond to the identical price series completely differently: the 10-day spikes highest and fastest, the 60-day barely lifts but stays raised for months. This proves the window length is not a detail — it manufactures the shape of the volatility you see, and the eventual fall in each line is partly the shock leaving the window rather than the market calming.

Professional explanation

The window length manufactures the shape

Rolling volatility feels like a measurement, but it is a measurement filtered through a choice, and the choice does most of the work. A short window weights the recent past heavily by giving each of a small number of days a large share, so it moves fast and swings wide. A long window spreads the weight thinly across many days, so it moves slowly and stays in a narrow range. Feed the identical price history into a 10-day, a 20-day and a 60-day rolling volatility and you get three lines that peak at different heights, at different times, and fall at different rates. None is wrong. Each is the correct realised volatility for its window. But it means the sentence realised volatility is 18% is incomplete until you say over what window, and a chart of rolling volatility is as much a picture of the analyst's window choice as of the market.

The window drop-out effect — a fake decline built into the estimator

This is the artefact everyone falls for. A single large move enters the window on the day it happens and stays inside it for exactly window more days, contributing its full squared return to every estimate in that span. Then, on one specific day, it falls out of the back of the window and the estimate drops sharply — not because the market did anything that day, but because the estimator stopped counting an old day. Traders see the drop and read it as volatility declining, as the market calming, as a signal to re-risk. It is none of those things. It is the shock leaving the window. The decline is a property of the ruler, not of the thing being measured, and it happens at a predictable moment — exactly window days after the shock — which is often precisely when a trader who sized positions on the elevated reading is being told, falsely, that the danger has passed.

EWMA: exponential weights and no cliff edge

The fix for the drop-out cliff is to stop using a hard window with a sharp back edge. Exponentially weighted moving average volatility replaces the equal-weight box with weights that decay smoothly: today's squared return gets weight (1 − λ), yesterday's gets λ(1 − λ), the day before λ²(1 − λ), and so on, so an old observation's influence fades gradually to nothing instead of being counted fully and then dropped at once. Because no day ever falls off a cliff, EWMA has no drop-out artefact — when volatility falls in an EWMA series, the market actually calmed. The parameter λ controls how fast the weights decay: a higher λ means a longer memory and a smoother series, a lower λ means faster reaction. RiskMetrics popularised λ = 0.94 for daily data, which corresponds to an effective memory of roughly 30-odd days and a half-life of about eleven days — recent enough to react, long enough to be stable — and it became a de facto standard precisely because it removes the cliff while keeping responsiveness.

Why rolling-volatility position sizing fails at the worst moment

Here is the sentence a marketing department would cut. A great deal of risk management sizes positions inversely to a rolling volatility: when measured volatility is low, take more risk; when it is high, take less. The window drop-out effect makes this rule systematically wrong at exactly the wrong time. After a shock, the rolling volatility stays elevated for the whole window length, forcing the position smaller and keeping it small even as the market may already have calmed — so you are under-risked through the recovery. Then, on the drop-out day, the reading falls off its cliff and the rule tells you to re-risk, right at the arbitrary moment the old shock exited the window, which has nothing to do with current conditions and may be just before the next one. A sizing rule driven by a backward-looking window does not just lag the market; it lags it by a fixed, mechanical amount and then jumps, and the jump is an artefact being treated as a signal.

The window drop-out cliff

A single large day held inside a 20-day window, then falling out of it exactly 20 days later.

The reading stays elevated for exactly the window length after the shock, then drops sharply on the day the shock exits the window — even though nothing happened in the market that day. This cliff is the window drop-out effect: the apparent decline in volatility is an artefact of the estimator, not a change in the market, and it is what EWMA is designed to remove.

Formula

Rolling realised volatility over an n-day window

RV_n(t) = √( (252 / n) × Σ r_{t−i}² ), i = 0 … n−1

Computed each day t over the trailing n days, assuming zero mean return and annualising on 252 trading days. As t advances, the oldest return r_{t−n} drops out and the newest r_t enters — that drop-out is the source of the cliff-edge artefact. The estimate is a backward-looking average, so it tells you about the window that has just passed, not the day ahead.

RV_n(t)Rolling realised volatility as of day t over an n-day window, annualised as a decimal (0.14 = 14%).
nThe window length in trading days — the choice that determines how fast the estimate reacts and how long a shock persists in it.
r_{t−i}The log return i days before t, ln(P_{t−i} / P_{t−i−1}), where P is the closing level.
Σ r_{t−i}²The sum of squared daily log returns across the n days in the trailing window; the zero-mean assumption means no sample mean is subtracted.
252Trading days per year, the annualisation factor.

EWMA volatility — the drop-out-free alternative

σ²_t = λ × σ²_{t−1} + (1 − λ) × r_{t−1}²

The exponentially weighted moving average variance, updated recursively each day. λ is the decay factor: RiskMetrics uses λ = 0.94 for daily data, giving a half-life of about eleven days and an effective window near 30 days. Because weights decay smoothly rather than dropping at a window edge, EWMA has no cliff-edge artefact, so a fall in an EWMA series reflects a genuine calming of the market.

How to compute and read rolling volatility without being fooled

Choose your window length deliberately, and state it. A 10-day window asks what is volatility now; a 60-day asks what has it been this season — decide which question you mean.
Compute daily log returns of the underlying and, each day, take the annualised standard deviation over the trailing n days assuming zero mean.
Plot the series and mark the dates of any large single-day moves — you will need them to spot artefacts.
When the series drops sharply, check whether a big day exited the window exactly n days earlier. If it did, the drop is a drop-out artefact, not a calming market.
If you are using the volatility to size positions or trigger decisions, switch to EWMA so that declines reflect genuine changes rather than the window edge.
Pick λ for EWMA to match your horizon: 0.94 is the RiskMetrics daily default; a higher λ smooths more and reacts slower, a lower λ reacts faster and is noisier.
Never compare a rolling volatility of one window against an implied volatility of a different tenor and conclude options are cheap — match the horizons first.

Practical example

NIFTY worked example

NIFTY at 24,000 has a single violent session — a 4% move — in an otherwise ordinary stretch. Look at how three windows report it. In a 10-day window that one day contributes roughly 0.04² / 10 to the mean squared return, which annualises to about √(0.04²/10 × 252) ≈ 20% on its own; in a 20-day window the same day contributes half as much and shows about 14%; in a 60-day window it barely lifts the reading to about 8%. Same day, three very different numbers. Now watch time pass with no further large moves. The 10-day reading stays near 20% for ten sessions, then on the eleventh session — when the 4% day falls out of the trailing ten — it drops abruptly, even though that eleventh day was quiet. A trader reading the 10-day line sees volatility collapse and concludes the market has calmed. Nothing calmed. The shock simply exited the window. Switch to EWMA at λ = 0.94 and the same episode shows a smooth rise and a gradual decay with no cliff, because the 4% day's weight fades instead of being dropped.

BANKNIFTY worked example

BANKNIFTY shows why the window choice interacts with the underlying's own character. BANKNIFTY realises more volatility than NIFTY and moves in sharper bursts around bank results and rate decisions, so its rolling volatility is both higher and spikier. Suppose BANKNIFTY at 52,000 has a 5% results-day move. In a 10-day window that single day pushes the annualised reading toward 25%, and it holds there for ten sessions before dropping off a cliff on the eleventh — and because BANKNIFTY has more of these bursts than NIFTY, its short-window rolling volatility is a sequence of these cliffs, each an artefact. A risk model that sizes BANKNIFTY positions off a 10-day rolling volatility is therefore making mechanical up-and-down sizing decisions on the drop-out schedule of past bursts. The longer 60-day window smooths the bursts but then lags badly, staying elevated long after a burst has passed. On a spiky underlying, EWMA earns its keep more clearly than on a placid one.

Lot sizes used above (NIFTY 75, BANKNIFTY 30) are those in force at the time of writing; NSE revises them periodically. Figures exclude brokerage, STT, exchange charges, stamp duty and GST. Examples are teaching scenarios built on round numbers — they are not historical quotes, not backtests and not trade calls.

Risk note. Volatility-targeting and inverse-volatility position sizing are widely marketed as disciplined risk control, and they are built on rolling volatility. The uncomfortable point is that a rolling window makes them systematically mistimed: they keep you under-risked through a recovery while the shock sits in the window, then tell you to re-risk on the arbitrary drop-out day when the old shock exits — which can be right before the next shock. Sizing off a backward-looking window does not remove risk; it relocates the mistake to a predictable, mechanical moment, and leverage applied on a false all-clear is how a calm-looking book blows up.

Advantages & limitations

What it is good for

It is simple and fully transparent — just an annualised standard deviation of recent returns — so anyone can compute it and see exactly what went into it.
It is model-free. It uses only observed prices of the underlying, with no option-pricing assumptions, which makes it a clean benchmark against which model-dependent implied volatilities can be judged.
The window length is a tunable question. A short window gives a responsive now reading and a long window a stable season reading, so you can deliberately match the estimate to the horizon you care about.
It is the natural input to a volatility cone and to realised-versus-implied comparisons, because it produces exactly the historical realised volatilities those tools are built from.
It makes regime persistence visible. A long-window rolling volatility that stays elevated for months captures volatility clustering — the tendency of turbulent periods to persist — in a single readable line.

Where it breaks down

It is backward-looking by construction, so it describes the window that has just passed and can say nothing about the day ahead — a shock always arrives before the estimate rises.
The equal-weight window creates the drop-out artefact: a shock is counted at full weight for exactly the window length and then dropped at a cliff, manufacturing a fake decline unrelated to the market.
Its level is inseparable from the window choice, so the same market produces materially different numbers at 10, 20 and 60 days, and no single one is the realised volatility.
It reacts only after a move, and on a short window it is so jumpy that a single outlier dominates the reading, while on a long window it lags so badly that it stays elevated long after conditions change.
It assumes a zero mean and uses close-to-close returns, so it ignores intraday range and can misstate volatility for an underlying that trends strongly or gaps between sessions.

Common mistakes

Reading the drop-out cliff as a calming market. When a rolling volatility falls sharply, check whether a big day just exited the window exactly window days ago — if so, nothing calmed, the estimator dropped an old observation.
Leaving the window length on a platform default and treating the output as the realised volatility. The window is the question; an unstated window makes the number uninterpretable.
Sizing positions inversely to a short-window rolling volatility. It keeps you under-risked through a recovery and then re-risks on the arbitrary drop-out day, mistiming risk at exactly the wrong moment.
Comparing a 10-day rolling volatility against a 30-day implied volatility to decide options are cheap. Different horizons have different natural levels; match the tenors before you compare.
Using a long window on a fast-moving underlying and concluding volatility is low because the reading lags. A 60-day window can still be reporting last quarter's calm well into a new turbulent regime.
Trusting a short-window reading dominated by a single outlier. On a 10-day window one 4% day can be most of the estimate, so the number reflects that one day far more than the general environment.

Professional usage

Risk desks and systematic strategies use rolling volatility as the raw realised-volatility input, but the sophisticated ones almost always move to EWMA or a GARCH model for anything that drives sizing, precisely because the equal-weight drop-out artefact corrupts inverse-volatility rules. RiskMetrics' λ = 0.94 became an industry default for daily value-at-risk because it gives a smoothly decaying memory that reacts to shocks without the cliff. A volatility trader compares a rolling realised volatility against the implied volatility of the matching tenor to gauge the volatility risk premium, and delta-hedges a position so that the profit and loss depends on the gap between realised and implied — which makes an accurate, artefact-free realised estimate the difference between measuring an edge and measuring a window.

Portfolio and overlay managers use rolling volatility to scale exposure to a volatility target, but the good ones explicitly correct for the drop-out effect — blending windows, using exponential weights, or capping how fast the estimate can fall — so that a mechanical re-risk is not triggered by an old shock leaving the window. Quant researchers use rolling realised volatility as the label their forecasting models are trained to predict, and as the benchmark those forecasts are scored against, which is why the choice of window and estimator is treated as a modelling decision in its own right rather than a preprocessing detail. In every professional use the recurring theme is that the estimator's artefacts are as important as the market's behaviour, and pretending otherwise builds the artefact into the decision.

Key takeaways

Rolling volatility is realised volatility over a moving window, and the window length is the actual question — a 10-day, 20-day and 60-day window give three different, equally correct answers to the same market.
The window drop-out effect makes a shock stay in the estimate at full weight for exactly the window length, then fall off a cliff, manufacturing an apparent decline in volatility that is an artefact, not a calming market.
EWMA fixes the cliff by decaying weights smoothly instead of dropping them at a window edge, so a fall in an EWMA series reflects a genuine change; RiskMetrics uses λ = 0.94 for daily data.
Position sizing off a rolling window is systematically mistimed — under-risked through a recovery, then re-risked on the arbitrary drop-out day — so the artefact becomes a trading mistake at the worst moment.
It is backward-looking and model-free, which makes it the clean input to volatility cones and realised-versus-implied comparisons, but useless as a forecast of the day ahead.

Rolling volatility looks like a plain measurement and is really a measurement plus a choice, and the choice — the window length — does most of the talking. Pick it deliberately, state it, and remember that when the line falls it may be the market calming or it may be an old shock walking out of the back of the window on a schedule that has nothing to do with today. When the number is going to drive risk, drop the hard window for EWMA so that a decline means what you think it means. The most dangerous volatility reading is the one that fell for a reason you did not check.

Frequently asked questions

What is rolling volatility?

Rolling volatility is the realised volatility of an underlying computed over a moving window of the most recent days and re-estimated each day as the window slides forward. Its level depends heavily on the window length you choose, not just on what the market did.

Why does the window length matter so much?

Because it determines how a shock is weighted. A short window gives each of a few days a large share, so it reacts fast and swings wide; a long window spreads weight thinly, so it moves slowly and stays narrow. The same market gives three different numbers at 10, 20 and 60 days.

What is the window drop-out effect?

It is the sharp fall in a rolling volatility that happens when a past shock exits the trailing window — exactly window days after it occurred. The estimate drops because the estimator stopped counting an old day, not because the market calmed; it is an artefact of the equal-weight window.

Why does rolling volatility stay high after the market has calmed?

Because a shock remains inside the trailing window at full weight for exactly the window length. Until it drops out the back, every estimate keeps counting it, so the reading stays elevated long after the market itself has settled — the estimator has a fixed-length memory.

What is EWMA volatility?

EWMA is exponentially weighted moving average volatility, where each past squared return gets a weight that decays smoothly by a factor λ each day. Because no observation falls off a cliff, EWMA has no drop-out artefact, so a fall in the series reflects a genuine calming of the market.

Why does RiskMetrics use λ = 0.94?

Because for daily data λ = 0.94 gives an effective memory of roughly 30-odd days and a half-life near eleven days — recent enough to react to shocks, long enough to stay stable. J.P. Morgan's RiskMetrics popularised it for value-at-risk, and it became a de facto industry default.

How does EWMA fix the drop-out cliff?

By replacing the hard window edge with weights that fade gradually to nothing. An old shock's influence decays smoothly rather than being counted fully and then dropped on one day, so there is no cliff and no fake decline — declines in an EWMA series correspond to real changes.

Which window length should I use for rolling volatility?

It depends on the question. A 10-day window answers what is volatility now with a jumpy, responsive reading; a 60-day answers what has it been this season with a smooth, slow one. Match the window to your horizon and always state which you used.

Is rolling volatility the same as implied volatility?

No. Rolling volatility is backward-looking, computed from past returns of the underlying with no options involved; implied volatility is forward-looking, extracted from option prices. One measures a past that happened, the other prices a future that has not.

Why is inverse-volatility position sizing risky?

Because it runs on a rolling window that is systematically mistimed. It keeps you under-risked through a recovery while the shock sits in the window, then tells you to re-risk on the arbitrary drop-out day when the old shock exits — which can be just before the next one.

Does a short window always give a higher volatility reading?

Not always higher, but always more responsive and more extreme in both directions. A short window spikes highest right after a shock because one day is a large fraction of few, but between shocks it can read lower than a long window that is still carrying an old move.

How is rolling volatility calculated?

Take the daily log returns of the underlying, and each day compute the annualised standard deviation over the trailing n days, usually assuming a zero mean and annualising on 252 trading days. As the window slides, the oldest return drops out and the newest enters.

Why does volatility clustering show up in rolling volatility?

Because turbulent days tend to follow turbulent days, so a window that spans a stormy period stays elevated across it. A long-window rolling volatility that remains high for months is a direct picture of clustering — the persistence of volatile regimes — in a single line.

Can rolling volatility predict future volatility?

Not directly. It is backward-looking, so it describes the window that just passed, and a shock always arrives before the estimate rises. Because volatility clusters, a high recent reading tends to be followed by more high readings, but the estimator itself forecasts nothing.

What is the effective window of EWMA at λ = 0.94?

Roughly 30-odd days, with a half-life of about eleven days — the point at which an observation's weight has fallen to half. That makes λ = 0.94 comparable to a month-long window but without the hard edge, so it reacts to shocks without a drop-out cliff.

Why do NIFTY and BANKNIFTY rolling volatilities differ?

Because BANKNIFTY realises more volatility and moves in sharper bursts around bank results and rate decisions, so its rolling volatility is both higher and spikier. On a short window that produces a sequence of drop-out cliffs, one for each past burst leaving the window.

Should I compare rolling volatility with implied volatility?

Only at matching tenors. Comparing a 10-day rolling volatility against a 30-day implied and concluding options are cheap compares two different horizons; match the window to the option tenor first, then the gap between them reflects the volatility risk premium.

What does a zero-mean assumption do to the estimate?

It means the calculation does not subtract a sample mean before squaring returns, treating the average daily return as zero. Over short windows this is close to true for an index and simplifies the formula, but for a strongly trending underlying it can overstate the volatility slightly.

Why can a single day dominate a short-window reading?

Because on a 10-day window one day is a tenth of the sample, so a 4% move contributes a large share of the total squared return. The resulting reading reflects that one outlier far more than the general environment, which is why short-window volatility is so jumpy.

Is EWMA always better than a simple rolling window?

It is better when declines must be trustworthy — for sizing or risk triggers — because it removes the drop-out artefact. For a transparent, easily explained historical measure, a simple window is fine; the choice depends on whether an artefact-free decline matters for your decision.

How does rolling volatility feed a volatility cone?

The cone is built from the historical distribution of rolling realised volatility at each window length. Each point on the cone is a percentile of the rolling volatilities for a fixed window, so rolling volatility is the raw material the cone is assembled from.

Voice search & related questions

Natural-language questions people ask about rolling volatility.

Why do my 10-day and 60-day volatility charts look so different?

Because the window length changes what you are measuring. The 10-day reacts fast and spikes on any single big day, while the 60-day smooths everything and lags. Same prices, two different questions — what is volatility right now versus what has it been over the season — so two different lines. Neither is wrong.

My volatility suddenly dropped but nothing happened — why?

Almost certainly the window drop-out effect. A big move from exactly window days ago just fell out of the back of your trailing window, so the estimate dropped even though today was quiet. Check the date of the last big move; if it was window days ago, the fall is an artefact, not a calming market.

What is the point of EWMA over a normal window?

It gets rid of the cliff. A normal window counts an old shock at full weight and then drops it on one day, giving a fake decline. EWMA fades old data out gradually, so when the volatility falls, the market actually calmed — which is exactly what you need if the number is driving your risk.

Why is 0.94 the magic number for EWMA?

It is not magic, it is a sensible default. J.P. Morgan's RiskMetrics chose λ = 0.94 for daily data because it remembers about a month of returns with a half-life around eleven days — fast enough to react to a shock, slow enough not to jump at every wiggle. It stuck as an industry convention.

Is it safe to add risk when my volatility measure drops?

Not on trust. If you are using a rolling window, check the drop is not just an old shock leaving the window — re-risking on that arbitrary day can land you right before the next move. Use EWMA so a decline means the market calmed, and remember nothing here is investment advice.

Which volatility window is the correct one?

There is no single correct one — the window is the question, not a setting. Short windows tell you about now and are jumpy; long windows tell you about the season and lag. Pick the one that matches the horizon you care about, and always say which window your number came from.

Sources & references

Last reviewed 10 July 2026. Educational content only — not investment advice.