IV Rank and IV Percentile: Reading Volatility Against Its Own History

Why the raw number has to be normalised at all

Our guide to what implied volatility is establishes the point this page starts from: implied volatility is bounded and regime like, it rests near a floor in calm stretches and spikes when fear arrives, and a level that is a floor for one underlying is a peak for a calmer one. That guide names both normalisations in a sentence each and moves on, because its subject is the volatility number itself. This page is the arithmetic underneath that sentence.

The problem the normalisation solves is real. A steady large index and a jumpy mid-size name occupy entirely different volatility neighbourhoods, so the only defensible reference for today's reading is the same underlying's own recent history. Both measures do this, both return a figure between 0 and 100, both are usually quoted to no decimal places, and a great many people who quote one of them could not say which. That last part is the expensive bit, because the two numbers answer different questions and the answers separate exactly when the market has recently been interesting.

What follows is a computation rather than an argument. Both measures are implemented from their definitions, run on the same synthetic series, and reported side by side. Then the disagreement is engineered deliberately, so its magnitude and its persistence can be quantified instead of asserted. Then the two are tested for whether they forecast anything, and that result is published as it came out.

The two formulas, and the one convention nobody states

IV rank places today's reading on the line between the lowest and the highest reading of the lookback window. Writing the window as the trailing set of sessions ending on and including today, the rank is one hundred times the quantity today's IV minus the window minimum, divided by the window maximum minus the window minimum. It is a linear interpolation between two numbers. Every other session in the window is ignored entirely, and it is worth pausing on how strong that statement is: if the window holds 252 sessions, 250 of them have no influence on the answer.

IV percentile counts instead. It is one hundred times the number of sessions in the window whose IV was below today's, divided by the number of sessions in the window. Every session in the window contributes one vote, the extremes carry no more weight than any other observation, and the answer is a position within the distribution rather than a position within the range.

There is a convention buried in the percentile definition that reference material almost never states, and it changes the number. Are you counting sessions strictly below today's reading, or at or below it? Does the window include today, or end at the previous session? On a volatility series with repeated values near the floor these choices move the answer by several points, and two platforms making different choices will print different percentiles from identical data without either being wrong. The computation on this page counts sessions strictly below today, over a window that includes today, and says so because a measure quoted to a whole number needs its convention attached.

Notice that the failure modes are not symmetric versions of one another. Rank fails by being dragged around by two observations that may be months old and may never repeat. Percentile fails by discarding magnitude, so it cannot distinguish a session that is slightly above the pack from one that is far above it. Both failures are structural. Neither is a bug in anyone's implementation.

Both measures on one series, before anything is engineered

The first run is the ordinary case. Six years of simulated daily implied volatility, resting around 15 percent, dipping to 8.2 and reaching 39.9 at its worst session. Twenty-one of the 1,512 sessions printed above 30 percent. Both measures are computed on a one-year lookback, which leaves 1,261 sessions on which both are defined.

The two lines track each other loosely and separate constantly. Across those 1,261 sessions the average absolute gap between the two readings is 14.4 points and the median is 9.4. On 24.2 percent of sessions the gap exceeds 20 points. On 10.9 percent of sessions it exceeds 40. The correlation between them is 0.858, which is high enough that they are clearly measuring related things and low enough that quoting one when you meant the other is a real error.

A sharper way to put it: on 42.4 percent of sessions the two measures land in different quarters of the 0 to 100 scale. Nearly half the time, someone reading rank and someone reading percentile would place the same session in a different band, and the two would each believe they had stated a fact about the market.

The widest disagreement on this series arrives at session 626. Implied volatility that session was 17.69 percent, a shade above the series median of 15.27. The lookback window ran from a minimum of 8.21 to a maximum of 39.91, so IV rank returned 29.9. But 87.7 percent of the sessions in that same window sat below 17.69, so IV percentile returned 87.7. One reading says today is in the bottom third of the range. The other says today is above seven eighths of the year. Both are arithmetically correct.

Worth noting what did not happen. On this naturally generated series the full inversion, rank in the bottom quarter while percentile is in the top quarter, never occurred once in 1,261 sessions. The gap gets to 57.8 points but not to a clean contradiction. To produce that you need a specific history, and the next section builds one on purpose.

Engineering the disagreement: one spike is enough

Take a series and hold it as calm as a mean-reverting process will allow. Outside the event, this one never leaves the band from 13.05 to 17.76 percent and averages 15.11. Now insert a single violent session: implied volatility jumps from 15.2 to 61.7 percent and then decays back over the following weeks, which is how a fear event actually resolves. Nothing else about the series changes. Then compute both measures on the same one-year lookback and watch what each of them does with that one observation.

The spike itself is unremarkable. Both measures shoot to the top while volatility is genuinely high, which is what they are supposed to do. The interesting behaviour begins after the market has calmed down and the spike survives only as a number inside the lookback window.

Session 338 is the clearest case. Implied volatility that session was 17.54 percent, which is slightly above the calm-stretch average and utterly ordinary. The lookback window ran from 14.01 to 61.66, so IV rank computes as 17.54 minus 14.01, divided by 61.66 minus 14.01, which is 3.53 divided by 47.65, or 7.4. In the same window, 213 of the 252 sessions closed below 17.54, so IV percentile computes as 213 divided by 252, or 84.5. A gap of 77.1 points on a single session, from the same input data.

Read those two numbers as a person would. IV rank 7 says volatility is scraping along the bottom of its range and options are about as cheap as they get. IV percentile 85 says volatility is higher than it has been on more than four fifths of the past year. They are not describing different markets. They are describing the identical market, and only one observation from four months earlier separates them.

This is not a one-session curiosity. There are 54 sessions on this series where IV rank reads below 20 while IV percentile reads above 70. Over the settled stretch that runs from the spike's decay to the point where it leaves the window, implied volatility averaged 15.03 percent and IV rank averaged 3.1. Over the equally calm 160 sessions immediately after the spike had rolled out of the window, implied volatility averaged 15.30 percent, essentially the same market, and IV rank averaged 45.2. The same volatility, scored 3 in one stretch and 45 in the next, purely as a function of what was still inside the window.

How long does one spike distort the reading? Exactly as long as the lookback, and the exit is not the clean snap you might expect. Because the spike decays over weeks rather than vanishing, its tail rolls out of the window session by session, and IV rank ratchets upward the whole time without any help from the market. At session 552 the reading was 5.5 with implied volatility at 15.45 percent. Ninety sessions later the reading was 57.9 with implied volatility at 14.99 percent. Volatility went down by nearly half a point while the rank went up by 52. Every bit of that movement came from history leaving the window.

The most compact statement of the problem is a pair of sessions. On session 364 implied volatility was 16.37 percent and IV rank read 5.0. On session 651 implied volatility was 16.43 percent and IV rank read 100.0. Six hundredths of a volatility point separate the two markets. Ninety-five points separate the two readings. A measure that can do that is not describing today; it is describing the window, and the window is a choice.

Percentile is better behaved here, and the parent guide is right to call it the steadier gauge against a spike. Across the same settled stretch its average was 35.0 rather than 3.1, and it moves as the distribution moves rather than as the extremes move. That is a genuine advantage, and it is the strongest argument for percentile that exists. It is also narrower than it sounds, as the next section shows.

The lookback is an arbitrary choice, and it moves the answer

Both measures need a window, and the window is almost always inherited from whatever the platform defaults to. One year is conventional. There is no derivation behind it. Recomputing both measures at 63, 126, 252 and 504 sessions on the main series, over the 1,009 sessions where all four are defined, shows how much rides on a choice nobody makes deliberately.

Across those sessions, the four IV rank readings for the same session span an average of 26.0 points, with a median span of 23.5 and a maximum of 78.9. On 32.4 percent of sessions the span exceeds 30 points. On 7.3 percent of sessions the four windows produce a reading below 30 and a reading above 70 for the same market on the same day, which is a complete reversal of the verdict from nothing but the length of the lookback.

The marked session is worth stating in full because it is so clean. Session 1378 had an implied volatility of 11.81 percent, which sits in the bottom sixth of the entire six-year series: only 15.2 percent of all 1,512 sessions closed below it. On a three-month lookback, IV rank read 97.9. On a two-year lookback, it read 19.0. The three-month window happened to contain an unusually quiet stretch, so 11.81 was the top of that small range; the two-year window contained the spike, so 11.81 was near the bottom. The percentile readings tell the same story, 96.8 against 13.5. An absolutely low volatility reading was simultaneously the most expensive session of its quarter and one of the cheapest of its two years.

Here is the result that complicates the tidy conclusion of the previous section. IV percentile is more sensitive to the lookback choice than IV rank, not less. Its readings for the same session span 30.2 points on average against rank's 26.0, and the span exceeds 30 points on 47.9 percent of sessions against rank's 32.4. This is not a contradiction of the spike result, it is a different question. Percentile is robust to one extreme observation entering the window, because one vote out of 252 barely moves a count. It is exposed to the whole distribution shifting, which is precisely what happens when you change how much history you are counting. Rank has the opposite profile: fragile to extremes, comparatively indifferent to the bulk.

There is also a systematic drift with window length that has nothing to do with any particular session. The average IV rank across the series falls steadily as the window grows: 42.9 at three months, 40.8 at six, 33.5 at one year, 29.4 at two years. A longer window is more likely to contain a spike, a spike stretches the denominator, and every ordinary session gets pushed down the scale. So a trader who lengthens their lookback will observe their readings drift lower and may conclude that volatility has become cheap. Nothing about the market changed.

Why the shape of the distribution decides all of this

The behaviour above is not bad luck. It follows from the shape of a volatility series, and once you see the shape it is obvious that the two measures must behave differently.

The simulated series has a mean of 15.64 and a median of 15.27, so the mean sits above the median, which is the signature of a right-skewed distribution. Its measured skew is 1.38. The bulk of the observations crowd into a narrow band, and a small number of sessions sit far above it: the fifth percentile is 10.09, the ninety-fifth is 21.86, and the maximum is 39.91. The distance from the median up to the maximum is about three and a half times the distance from the median down to the minimum.

Now apply each formula to that shape. IV rank divides by the range, and on a right-skewed series the range is dominated by a maximum that almost never gets revisited. The denominator is therefore large relative to the everyday spread, and every ordinary session lands low. The measured consequence: across the series, IV rank read above 80 on only 5.7 percent of sessions. IV percentile read above 80 on 19.3 percent, which is close to the 20 percent an evenly spread measure would give. Rank's average reading is 33.5 against percentile's 45.7, a systematic 12-point offset between two numbers that are both meant to describe the same thing on the same scale.

Look at the occupancy bars in the figure and the mechanism is visible. Percentile is not perfectly even, because volatility clusters and a trailing window is not a random sample, but it is roughly flat: it spends time across the whole scale. Rank is heavily bottom loaded, with 664 of 1,261 sessions in the lowest three tenths of the scale and only 121 in the highest three tenths. Rank is not a broken measure. It is a faithful description of proximity to an extreme, and on a right-skewed series most days are a long way from the extreme.

This has a practical consequence for anyone applying a fixed threshold. A rule expressed as "above 70" means something quite different depending on which measure it is applied to, because one of them reaches 70 several times as often as the other. Thresholds imported from a source that used the other measure, which is most of what circulates, are not being applied as their author intended.

The honest limit: neither one is a forecast

Everything so far concerns whether the two measures describe the past accurately. The question that actually matters is different: does a high reading tell you anything about what volatility does next? The claim is everywhere, usually in the form that a high reading means volatility is likely to fall. It is testable on this series, so it was tested, and the result is published as it came out.

The test measures the change in implied volatility over the following 21 sessions, roughly one month, for every session on which the reading is defined. That is 1,240 observations. Before any number is quoted, the honest caveat has to come first: the series was built with mean reversion in it, so some relationship is guaranteed by construction. This test cannot tell you whether real implied volatility mean reverts. It can only tell you how much dispersion sits around a mean reversion that is known to be present, which turns out to be the more useful question anyway.

The middle of the scale behaves the way the folklore says. In the fifth and sixth rank deciles, implied volatility was lower a month later on 72.5 and 78.8 percent of sessions, and the average change was about two volatility points down. If you stopped reading there you would conclude the measure works.

The top of the scale does not behave that way at all. In the highest rank decile the average change over the next month was positive, at 3.43 volatility points, and volatility was lower a month later on 54.5 percent of those sessions, which is close to a coin toss. The relationship is not monotonic and the failure is at exactly the end everyone quotes. The reason is mechanical: IV rank reaches its maximum on the first session of a spike, not the last, and volatility clusters, so a reading of 100 frequently marks the beginning of a violent stretch rather than its end. Taking every session with a rank of 80 or above, the average change over the following month was an increase of 1.86 points, with a standard deviation of 9.80 and a worst case of a further 24.35 points up.

The dispersion is the real story. In the top rank decile the tenth to ninetieth percentile of outcomes runs from 5.63 points down to 18.37 points up. Knowing the reading narrows that range hardly at all. Put in the standard form: IV rank explains 2.0 percent of the variation in the next month's change, and IV percentile explains 3.6 percent. Restricting to the 60 non-overlapping observations, so the sample is not counting the same month sixty times, gives essentially the same answer.

And then the result that undercuts the entire exercise. The raw, un-normalised implied volatility level explains 22.6 percent of the next month's change, more than ten times either normalised measure. The correlation is 0.475 in magnitude against 0.142 for rank and 0.188 for percentile. Normalising to a 0 to 100 scale is precisely the step that discards the magnitude, and on a mean-reverting series the magnitude is where the information lives. A session at 40 percent volatility falls harder in absolute points than a session at 20 percent, and both of them can read 100.

That is measurable directly. Twelve sessions on this series scored 100 on IV rank. The implied volatility on those sessions ranged from 15.38 percent to 39.91 percent. The measure that was supposed to make the number readable had rendered a 2.6-fold difference in the underlying quantity as an identical score.

None of this makes the measures useless. It makes them what they are: compact descriptions of where today sits in a window of history, which is a genuinely useful thing to know and is not the same thing as a forecast. The reading tells you what you are being asked to pay relative to what you have recently been asked to pay. It does not tell you what happens next, and the arithmetic above is what that claim looks like when someone checks it.

What a reading changes, and what it does not

The level of implied volatility has a direct and unavoidable consequence for the price of an option, and that consequence is arithmetic rather than opinion. Take the same illustrative structure at three different sessions from the series: a 30-day at-the-money call and put together on an underlying at ₹1,000, priced on a standard option model at a 6.5 percent rate. This is the two-legged case because it isolates the volatility effect from any directional view.

At session 275, where IV rank read 0 and implied volatility was 11.69 percent, the pair cost ₹27.01. Breakevens sit at ₹972.99 and ₹1,027.01, so the underlying has to travel 2.70 percent in either direction before the buyer recovers cost, and the unprofitable zone between the breakevens is 5.40 percent wide. At session 1464, where IV rank read 100 and implied volatility was 20.96 percent, the same pair cost ₹47.99, breakevens at ₹952.01 and ₹1,047.99, a 4.80 percent move needed either way. At session 596, where IV rank also read 100 but implied volatility was 39.91 percent, the pair cost ₹91.11, breakevens at ₹908.89 and ₹1,091.11, a 9.11 percent move needed. All figures illustrative.

The second and third of those sessions carry the identical reading and one premium is 1.9 times the other. For the buyer of both legs the maximum loss is capped at the premium paid. For the writer of them the loss on the call leg is unbounded, and that asymmetry does not change with the reading. This is the behaviour, stated as behaviour: a high reading describes a market in which the same structure costs more and requires a larger move to break even, and in which the writer of that structure receives more for carrying a risk that has no upper limit. It is a statement about prices, not about what anyone should do.

The charge stack interacts with the reading in a way that is easy to miss. The costs that scale with the premium stay a constant share of it: securities transaction tax on options is 0.15 percent of premium on the sale side, raised from 0.1 percent by the Finance Act 2026, section 159, with effect from 1 April 2026, and stamp duty on an options transfer is 0.003 percent on the buy side under the Indian Stamp Act 1899, Schedule I, Article 56A(d). The costs that do not scale are the problem. Applying an illustrative flat fee of ₹20 per order plus 18 percent GST, across two legs, in and out, and an illustrative contract multiplier of 50, the total charge comes to about ₹96 against a premium of ₹1,350 at the low-volatility session, which is 7.14 percent of the premium, and about ₹101 against ₹4,555 at the high-volatility session, which is 2.23 percent. Illustrative throughout. Cheap options are not proportionally cheap to transact, because the flat part of the stack does not shrink when the premium does. Contract multipliers are periodically revised to keep contract value inside the prescribed band and the figure used here is illustrative only, not a current lot size.

One item could not be verified and is therefore not on this page. A current NSE equity-options exchange transaction charge could not be confirmed from a primary source at the time of writing, so no figure for it is quoted and the totals above exclude it. The real all-in cost is higher than the numbers shown.

None of this is a reason to take or avoid any position, and the context deserves stating once. SEBI found that about 93% of individual traders in equity derivatives made net losses over FY22 to FY24, with aggregate net losses exceeding Rs 1.8 lakh crore (SEBI, September 2024). A normalised volatility reading is a description of a price. It is not a solution to that finding, and treating a number between 0 and 100 as though it were is how the number does harm.

Reading a pair of values

Because the two measures disagree so often, the informative object is the pair rather than either number alone. The combinations below are not a scheme; they are the four cells of the scale with the share of sessions that actually landed in each, computed on the main series.

The practical instruction that falls out of the whole exercise is unglamorous. When a reading is quoted, ask which of the two measures produced it and over what window, because those two facts move the number by more than the market usually does. If the answer is not available, the number is not interpretable, and a number that cannot be interpreted is worse than no number because it feels like information.

The index-level version of the same idea has its own guide, on what India VIX measures, and the arithmetic here applies unchanged to any volatility series you can put in a window.

What the exercise leaves you with

Three things survive the computation. The first is that normalising a volatility reading is necessary, because a raw level genuinely is uninterpretable across underlyings, and both measures do that job. The second is that they do it differently enough to reach opposite verdicts on nearly half of all sessions, that the divergence is largest exactly after the events people most want to reason about, and that a single spike can hold IV rank near zero for a full lookback while the market underneath it is entirely ordinary. The third is the uncomfortable one: measured against what volatility actually did next, the normalisation destroyed most of the predictive content that the raw level had.

That last finding is not a reason to abandon the measures, and it would be a misreading to take it as one. It is a reason to hold them to what they are. They are descriptive statistics about a window of the past, computed to make one underlying comparable with another, and they succeed at that. They were never forecasts, the folklore that treats them as forecasts does not survive a decile table, and the gap between what a statistic measures and what people believe it measures is where most of the damage in this field gets done.

The habit worth taking from this page is smaller than any of the numbers on it: name the measure, name the window, and check whether a spike is sitting inside it. Three seconds of scepticism about a number between 0 and 100. That kind of care about what a statistic can and cannot support is the substance of the method we teach, and it generalises well beyond volatility.

Every figure comes from a single seeded simulation that reproduces identically on each run. The volatility series is generated by a mean-reverting process in logarithms with occasional upward Poisson jumps, which yields a bounded, right-skewed, spike-capable series. It is synthetic and is not a model of any specific security or index. IV rank and IV percentile are implemented directly from their definitions and both are checked against known answers on a strictly increasing ramp, a flat series and a hand-computed midpoint case before any result is reported. The percentile convention counts sessions strictly below today over a window that includes today.

The engineered series is the same process with its jump component switched off and one spike inserted by hand, so that the divergence it produces has a single identifiable cause. Option premiums are computed on a standard option pricing model at a 6.5 percent rate on a ₹1,000 underlying. The forward test regresses the change in implied volatility over the following 21 sessions against the reading, and is reported both on all overlapping observations and on the non-overlapping subsample.

All results are illustrative and simulated. They are not a track record, they are not a forecast, and they are not an indication of what any instrument would produce in a live account. The purpose of the exercise is to demonstrate properties of two calculations, which are properties of the arithmetic rather than of any particular market. Statutory charge rates cited are from primary sources with dates; where a rate could not be verified from a primary source it is not quoted.

• What implied volatility actually is →
• What India VIX measures →
• How to read an option chain →
• The eight ways a backtest can lie →