Guide · Indian markets · Research method · Updated 2026-09-20

A volatility zone is a compression of the record it was fitted to, and the record moves

This page computes rather than quotes. Every figure came out of the exchange session files named in the Sources block, with the window, the rule and the arithmetic stated so a reader can redo it and disagree with a number instead of a vibe. That matters here because volatility zone pages are unusually full of confident round numbers whose origin nobody can reconstruct.

India VIX is a price for variance, and prices are not forecasts

Start with what the number is, because three misreadings of it survive almost every treatment of the topic. India VIX is the option market's price for variance on the broad fifty share index over the coming thirty calendar days, restated as an annualised standard deviation in per cent. The exchange builds it from the order book of near and next maturity index option contracts: the best bid and best ask at each strike, weighted by the inverse square of the strike level, summed across the out of the money chain, then interpolated between the two maturities to a constant thirty day horizon.

The order book detail is not a technicality. Many strikes do not trade in a given interval, so a last traded price would be stale and a computation built on traded prices would freeze exactly when it mattered. Using the resting quotes keeps the index continuous, at a cost: when spreads widen, part of the move in the index is the widening rather than a genuine repricing of risk. In thin conditions the number carries liquidity as well as expectation, and nothing in the published level separates the two.

The first misreading is that the level is a percentage move. It is not. A reading is an annualised standard deviation, so the operational figure requires a division. At the reading of 11.39 on 2026-09-18, the implied daily standard deviation is 0.718 per cent and the implied typical daily move, the mean absolute deviation of a normal, is about 0.572 per cent. Note which day count each conversion uses: scaling to the thirty day horizon multiplies by the square root of thirty over three hundred and sixty five, because calendar days are the convention inside the index formula, while a daily figure divides by the square root of about two hundred and fifty two sessions, because sessions are what a daily move happens in. Mixing the two is how a reader ends up a fifth out.

The second misreading is that the index is realised volatility. It is not, and the record says how far apart the two run. Comparing each session's reading with the volatility the broad index actually went on to deliver over the following 21 sessions, across 2,837 comparisons, the index exceeded what turned up on 82.3 per cent of occasions, with a median ratio of 1.32. In horizon terms the median implied move was 4.35 per cent against 3.43 per cent delivered.

That persistent gap is a risk premium, not an error. Somebody is being paid to carry variance, and the price of carrying it sits above the average amount delivered. Read the last two columns together, because they disagree in a useful way. In points of movement the gap widens with the level, from 0.90 per cent in the quiet zone to 1.29 per cent in the elevated one. As a ratio it barely moves, 1.31 to 1.33 with no clean ordering. So the premium is roughly proportional to the level rather than a fixed number of points, which is the opposite of how a flat buffer would behave.

The third misreading is that the index looks forward. It looks both ways at once, and more firmly backward. Rank correlation between the reading and the trailing 21 session realised volatility of the broad index is 0.746; against the forward window it is 0.657, on 2,691 overlapping observations. So the reading agrees more closely with what just happened than with what happens next. It is still the better instrument of the two, because the trailing window on its own reaches only 0.540 against the forward one, so the option market adds something. The honest description is a price that leans heavily on recent experience and contains a modest amount of genuine anticipation. A change in the reading is also mostly the same session's move seen from the other side: rank correlation between the daily change in the index and the same session's index return is -0.515.

The distribution, measured over the longest window the files support

The archive of daily index close files on hand runs from January 2013, but the India VIX row first appears on 2014-05-14, so that is where the series starts here. It runs to 2026-09-18, giving 3,048 sessions, about 12.4 years. Guards run before any figure is computed and the build refuses to write if one fails: every full calendar year from 2015 to 2025 must hold at least 235 sessions, no gap between consecutive sessions may exceed seven calendar days, and a broad index close must exist on every session carrying a volatility index close. The largest gap the record actually contains is 6 calendar days. A missing block would masquerade as a step change in the level and would corrupt every percentile around it.

Read the shape before the summary statistics, because the shape is what invalidates most of them. The mass sits in a narrow band and the right tail is long and thin. Lowest close in the record is 9.15 on 2025-12-26; highest is 83.61 on 2020-03-24. Half of all sessions fall between 13.16 and 18.19, a span of 5.03 points, while the top one per cent spans from 39.70 to 83.61.

The recent low end is where this record dates most live pages. The lowest close anywhere before 2025 was 10.14 on 2023-07-28. Since then the series has printed 15 closes below 10.00, all of them between 2025-09-18 and 2026-01-07, with the record low of 9.15. Anything written before late 2025 that describes a floor near 10 or 11 is describing a range the series has since left, and the cheapest way to check that claim is to filter the same public session files.

The round numbers describe a period, not the market

Almost every treatment of this topic hands over a pair of thresholds: below fifteen is calm, above twenty is high, above thirty is panic. The numbers are not absurd, and they are not derived from anything. Put them against the record and see what they actually cut.

A threshold at 20 leaves 16.4 per cent of sessions above it, so a system with a separate parameter set for high volatility would use that set on about one session in six and would be fitting it on 499 observations spread over a decade. A threshold at 30 leaves 66 sessions, which is not a regime, it is a handful of weeks. Meanwhile a threshold at 15 splits the record almost exactly in half, at 47.6 per cent below, so calling everything above it something other than ordinary describes half of all trading as unusual. The earlier internal draft of this topic used 14 and 19 and claimed those were the 50th and 75th percentiles of its own window. On the real 2020 to 2024 distribution the 50th percentile is 16.16 and the 75th is 20.68. The numbers were not derived from the data they were attributed to.

Deriving the cuts from the distribution instead of choosing them

A defensible threshold has one property: someone else with the same data reaches the same number without knowing what you wanted the answer to be. That rules out any cut chosen because it made a backtest look better, and it rules out round numbers. It leaves a rule stated in advance.

The rule used here is the plainest available. Take every India VIX daily close in the stated window, sort it, and cut at the 33rd and 67th percentiles by linear interpolation. That gives 13.76 and 17.00, and by construction each zone holds a third of the record: 1,017, 1,015 and 1,016 sessions. Nothing was tuned. The zones are named quiet, ordinary and elevated, deliberately not calm, normal and stressed, because a third of all sessions is not stress and a label that overstates what it found will be believed too far.

Why percentiles rather than the mean and standard deviation, which is the other obvious rule? Because the standard deviation of this series is not a stable scale. Set the cuts at the mean minus half a standard deviation and the mean plus one, and the record splits 30 / 61 / 9 per cent instead of into thirds, because a long right tail drags the mean up and inflates the spread. Then remove 58 sessions, 1.90 per cent of the record, and refit both rules.

The mean and standard deviation upper cut moves from 22.42 to 19.82, 11.6 per cent, and the share of the full record it calls elevated goes from 9 per cent to 17 per cent. The percentile cut moves from 17.00 to 16.84, 1.0 per cent, and coverage stays within a point and a half of a third. A percentile counts sessions and a moment weights them by how extreme they were, which is why one rule survives the episode and the other is largely a report of it.

One genuine extension survives measurement. Split the elevated third at the 95th percentile of 24.66 and the two halves do not behave alike: mean absolute next session move is 0.837 per cent on the 858 elevated sessions below that level and 1.706 per cent on the 153 above it, a factor of 2.04 inside a single zone. A fourth zone earns its place there because the data separates, not because four sounds more thorough than three. The cost is stated in the table above: adding it takes the episode count from 325 to 378 and leaves the tail zone with 5 per cent of the record to fit anything on.

Why a discrete label rather than a continuous adjustment

If the level carries information about the size of the next move, the obvious design is to scale a parameter smoothly with the level and skip the zones entirely. That design is defensible and this page does not argue against it in principle. It argues that the data will not support the resolution it implies.

The broad ordering is there and the fine ordering is not. Of the nine steps from one decile to the next, 1 goes the wrong way, and most of the increase is concentrated in the top decile at 1.353 per cent against 0.820 per cent below it. Adjacent deciles are separated by roughly a twentieth of a percentage point of movement, on three hundred sessions each, which is not a gap this record can resolve.

The second cost is sample. A ten zone scheme changes label so often that the record breaks into 1,074 episodes with a median length of 2 sessions, and the label repeats next session only 64.8 per cent of the time. Five zones give 562 episodes at 81.6 per cent. Three give 325 at 89.4 per cent. A parameter set has to be fitted on episodes, not on sessions, because sessions inside one episode are close to the same observation repeated. Coarse is not a simplification here. It is the resolution the sample actually has.

Persistence is what makes the label usable

A classifier is only worth attaching a decision to if the class you are in tells you something about the class you will be in. Otherwise every parameter change arrives just as the condition it was chosen for ends.

Two mechanisms in that table. The diagonal is heavy: 89.4 per cent of sessions keep the previous label, and the extreme zones hold harder than the middle at 91.5 and 92.6 per cent against 84.2 per cent. The corners are empty: a jump from quiet straight to elevated in a single session happened 2 times in 1,017 quiet sessions, and the reverse jump from elevated straight to quiet happened 0 times in 1,016. The ordinary zone is not a resting state, it is the corridor everything has to pass through, which is why it is the least persistent of the three and holds the most episodes at 161 against 88 and 76.

Persistence decays in a way worth knowing before anybody builds a rule on it. The label holds 89.4 per cent of the time one session out, 77.0 per cent five sessions out, 69.7 per cent ten out and 60.7 per cent at 21 sessions, which is about the horizon the index itself covers. Over the horizon the reading is priced for, the zone you are in today is barely better than the unconditional frequency of a third. So a zone is a description of now, not a month ahead forecast, and using it to set something that cannot be changed for a month throws away the only property that made it usable.

The mean and the median disagree sharply about episode length, and the disagreement is the finding. The median episode is 3 sessions in all three zones while the means run 6.3 to 13.4. Most episodes are brief and a few are enormous, the longest 311 sessions elevated from 2020-02-26 to 2021-05-28. Quoting the mean suggests a comfortable fortnight of stability; quoting the median suggests noise. Both are true of different episodes, and a rule that assumes either one alone will be wrong most of the time.

The label flickers at the boundary, and confirmation fixes most of it

Where does the brevity come from? Not from the market changing state every three sessions. From a reading sitting on a threshold and crossing it back and forth.

41.8 per cent of all episodes last two sessions or fewer and together hold only 6.1 per cent of the record. Those are not regime changes. And 31 per cent of all sessions sit within five per cent of one of the two thresholds, so this is not a rare corner of the data. Right at the lower cut, the next session's mean absolute move is 0.540 per cent for the 274 sessions just below and 0.575 per cent for the 246 just above. The label changes; the behaviour does not.

Requiring two consecutive sessions to agree removes 41 per cent of the switches at a cost of disagreeing with the raw label on 2.2 per cent of sessions. Three sessions removes 60 per cent for 4.5 per cent. Median episode length goes from 3 sessions to 12, and coverage barely shifts, so the confirmed label is still splitting the record into thirds. The switches removed are the ones with no content, which is the rare case where a rule gets strictly better by acting later.

What the zone orders, and what it does not

A classifier is worth its complexity only if the classes differ in something you can act on. Two questions are worth putting to the same labels, and this record answers one of them and refuses the other.

The size question is settled by this record. Mean absolute next session move runs 0.490, 0.611 and 0.968 per cent, a factor of 1.98 end to end, with intervals that do not overlap. The overnight gap separates harder still: the chance of the index opening more than half a per cent away from the previous close goes from 12.2 per cent after a quiet session to 42.0 per cent after an elevated one, a factor of 3.4. Anyone carrying a position overnight is exposed to a different distribution of opens in each zone, and that, rather than anything about the closing move, is the most usable thing in the table.

The direction question returns nothing. Mean signed next session move is +0.0205, +0.0347 and +0.0688 per cent, every interval crosses zero, and the three test statistics are 1.03, 1.37 and 1.54. The share of up sessions runs 54.1, 52.3 and 55.2 per cent, with the highest in the elevated zone, which is the opposite of the folk reading and is also not a finding. It is a null result, reported because every zone page implies otherwise without ever measuring it.

Fit the cuts on one period, apply them to another

Here is the limit that belongs on the page rather than in a footnote. The whole method above fits the thresholds on the record it then describes. Split the record instead and the picture changes.

The cuts themselves look close, 14.22 and 16.73 against 13.37 and 17.54. The consequence is not close. Applying the other half's cuts relabels 20.0 and 14.7 per cent of sessions, and coverage stops being thirds: the second half under the first half's cuts becomes 44 / 18 / 38 per cent, so the ordinary zone shrinks to under a fifth of sessions while quiet takes over two fifths. A parameter set fitted for a third of sessions is then being applied to a much smaller or larger slice than it was fitted on.

The last column is the whole problem in one place. A reading of 16 is quiet under the 2023 fit and elevated under the 2026 fit, and nothing about the rule changed between them. The upper cut ranges from 14.13 to 21.54. That range, not a decimal place, is the error bar on any threshold anyone publishes for this series.

An expanding window fit answers the obvious objection, that the full record cuts used above could not have been known early on. Refit the cuts every session on all earlier readings only, after a burn in of 500 sessions, and 2,548 sessions from 2016-06-08 get a label a person could actually have held at the time. Those labels differ from the full record fit on 261 of them, 10.2 per cent, and coverage comes out 44 / 25 / 31 per cent rather than thirds. The cuts start at 15.48 and 17.50 and end at 13.76 and 17.01. Every zone share quoted anywhere on this page, including in the table below, is a property of a fit that used the whole record.

Read down the elevated column. 2020 spent 85 per cent of its sessions there and 2017 spent 0 per cent, not one session. A classifier that needs a minimum number of observations in each zone before its parameters mean anything will wait years in some periods and be flooded in others, and the waiting is not visible from the pooled coverage figure of a third.

One quarter wrote most of the tail

The failure mode generic pages omit is not that thresholds drift. It is that a threshold fitted over a window containing one extreme episode is substantially a description of that episode, and the page reporting it usually presents it as a description of the market.

Measure it directly. The 58 sessions from 2020-03-01 to 2020-05-31 are 1.90 per cent of the record. They carry 56.8 per cent of its total squared deviation from the mean, every one of the 30 closes above 40, and 80 per cent of the 66 closes above 30. Remove them and excess kurtosis falls from 31.6 to 1.5, skewness from 4.34 to 1.11, the standard deviation from 5.94 to 3.90 and the 99th percentile from 39.70 to 28.57. The mean barely moves, from 16.48 to 15.92, and the median hardly at all.

So a page reporting the mean and the median of this series is reporting the market. A page reporting the standard deviation, the skewness or the kurtosis is reporting one quarter of 2020 with the rest of the decade as background. The two look identical in a table of summary statistics, and that is the trap. It is the same trap behind every published threshold sitting somewhere in the twenties: fitted on a window containing that quarter, those levels are where the episode put them.

This is also the answer to the tempting shortcut of defining the top zone by how unusual it was. A cut at two standard deviations above the mean of this series lands at 28.36, above which sit 85 sessions, 2.8 per cent of the record. On a symmetric distribution that cut would take about two and a half per cent. The discrepancy is the skew, and it means the familiar sigma vocabulary silently changes meaning on this series. What a standard deviation understates in a fat tailed series is the neighbouring piece.

What the classifier is actually for

Nothing above establishes that three zones are the right number, that these cuts are the right cuts, or that the conditional differences will hold. This record cannot settle those questions: 3,048 sessions hold 325 episodes, 76 of them elevated, and a claim about behaviour in the elevated zone rests closer to that number than to the 1,016 sessions it contains.

What the classifier delivers is narrower and real. It converts a continuous number nobody has intuition about into a state with a measured record attached, and the measured record says specific things. The size of the next move scales by a factor of 1.98 across the zones and the chance of a material overnight gap by 3.4. Direction does not move at all. A position sized on the pooled distribution is sized for a blend of three states, and in the elevated zone the same position carries roughly 3.4 times the overnight gap exposure without a single contract being touched.

It also tells you what to distrust. A published threshold is a statement about a window; ask which one. A zone share is a property of a fit; ask whether the fit used data from after the period it labels. A tail statistic on this series is a report on one quarter; ask what happens when that quarter is removed. The current reading of 11.39 on 2026-09-18 sits in the quiet zone under the full record cuts, and under the trailing three year refit for 2026 it is quiet as well, which is the agreeable case. The disagreeable cases are the ones near a boundary, and there are a lot of them.

The habit underneath all of it is the transferable part: state the rule before looking at the result, publish the window, and measure what your own classifier does to the record rather than describing what it ought to do. Reading a performance figure as an average across states is the next step once a label exists, the instability of a fitted correlation is the same lesson on a different statistic, and publishing a result somebody else can check is what turns a threshold into a claim rather than an opinion.