The claim
The VIX has spent six months inside a narrow band and the reflex reading has a name. Coiled spring. Compressed. Loaded. The image is a physical one: energy going in, and energy that has to come out.
It is a claim about conditional probability wearing a metaphor's clothes. Given a compressed VIX, are big volatility spikes more likely over the next month than they ordinarily are? That has an answer, and the answer is in every session since 2003.
We counted. Across 579 compressed days, big spikes followed less often than they followed an average day at the same VIX level. Not more.
What "compressed" means here, exactly
For every session, take the highest high and the lowest low of the trailing 120 sessions, roughly six months, and divide the distance between them by the average close over those same 120 sessions. That gives a band width as a percentage of the level the index was trading at, which is the only way to compare 2008 with 2017.
A day counts as compressed when that band width sits in the bottom tenth of all readings since September 2003. The threshold works out to 58.82%.
Then, for every day, the measurement looking forward: the highest high the VIX reached over the following 30 sessions, as a percentage above that day's close. Call it the forward peak. It asks the question the coiled-spring claim actually makes, which is about how far up the index got, not where it finished.
Why the measurement starts in September 2003
This is the detail that decides whether any of the rest is worth reading.
The dataset is 9,274 daily VIX bars running from 3 January 1990. The first 3,457 of them carry a high, low and close that are all the same number. They are close-only records with no real intraday range behind them.
Feed those into a high-low range statistic and it will report the early 1990s as the calmest stretch in the index's history, because a range built from bars with no range is zero. Every percentile computed off that distribution would be wrong, and wrong in the direction that flatters this exact test by making today look less unusual.
So the reference sample starts at 19 September 2003, the first date followed by a sustained run of real intraday ranges. That leaves 5,817 usable bars, of which 5,786 have a complete 30-session window ahead of them. Thirty-seven per cent of the available history is discarded on purpose.
Where the VIX actually is
At the close of 30 September 2026, the last complete session, the VIX was 16.33. Its 120-session band runs from 13.80 to 23.34 against a mean close of 16.90, which is a band width of 56.44% of level. That is the 7.3rd percentile of everything since 2003.
It has been in the bottom decile for three sessions, starting 28 September. Before this stretch, the last session at or tighter was 19 October 2023, which is 2.95 years ago.
The containment is the part you can see without any statistics. Every close for 122 straight sessions, starting 8 April 2026, has sat inside that 13.80 to 23.34 band. On a closing basis the extremes over that run are 14.22 and 22.23. The session before it, 7 April, closed at 25.77.
One thing this is not. It is not the narrowest six-month range since 2023, and anyone writing that has picked a measure without saying so. On range as a percentage of level, today is tighter than anything since October 2023. On raw VIX points, today's 9.54 is wider than 18 April 2024, which printed 8.05. The two measures disagree because the VIX was trading higher in April 2024, near 18, so the same points bought a smaller share of the level. Range-as-a-share is the right measure for comparing across regimes, and it is the one used throughout. But "tightest since 2023" is only true on one of them, so this piece says one of its tightest stretches instead.
The count
Across every session since September 2003 with a full window in front of it:
| Median forward peak | Reached +50% or more | |
|---|---|---|
| After compression (n = 579) | +21.8% | 17.1% |
| Every day (n = 5,786) | +28.7% | 26.0% |
Compressed days were followed by smaller peaks and fewer big ones. The distribution shows where the difference sits, and it is not in the extremes:
| Forward peak | After compression | Every day |
|---|---|---|
| +100% or more | 4.7% | 7.2% |
| +50% to +100% | 12.4% | 18.9% |
| +25% to +50% | 26.6% | 29.7% |
| +10% to +25% | 33.0% | 26.8% |
| Under +10% | 23.3% | 17.4% |
Compression does not remove the tail. It thins it. Nearly one compressed day in twenty was still followed by the VIX doubling inside six weeks. What changes is the middle: compressed days land in the +10% to +25% bucket far more often, and in the +50% to +100% bucket far less.
The control, which is the part that matters
There is an obvious objection to everything above, and it is the right one. Compression is not random. The VIX compresses when it is low, and when the VIX is low a +50% move is a smaller absolute distance to travel. Any comparison against all days is partly comparing quiet markets with loud ones.
So hold the level still. Take only days where the VIX closed between 14 and 19, a band around today's 16.33, and run the same test inside it.
| VIX between 14 and 19 | Median forward peak | Reached +50% or more |
|---|---|---|
| After compression (n = 236) | +21.9% | 6.8% |
| Every day (n = 2,103) | +30.3% | 27.5% |
At the same starting level, big spikes followed compression about a quarter as often as they followed an ordinary day. That is the opposite of the coiled-spring claim, and it is the single number this piece would defend.
Where that result holds, and where it does not
The honest version requires splitting the two columns, because they do not behave the same way under stress.
The spike rate holds up. We re-ran the comparison across four period cuts and four level bands, sixteen combinations. The compressed share reaching +50% is below the matched baseline in fifteen of them. The one exception is the 2010-onward sample with the level band widened to 12 to 21, where it reads 43.3% against 33.0%. On the tightest band tested, 15 to 18, the compressed rate sits below baseline in every period: 1.2% against 28.9% post-2003, 5.9% against 33.1% from 2005, 7.1% against 33.1% from 2010, 12.5% against 38.1% from 2015.
The median does not. 184 of the 236 level-matched compressed days fall in 2003 and 2004, and 156 of those are a single unbroken stretch from 9 October 2003 to 15 June 2004. That is not fifteen separate pieces of evidence. It is mostly one long post-dot-com grind, and it is doing most of the work in the median column.
Drop those two years and the medians read +32.7% for compressed days against +32.4% for all days at the same level. There is no gap left. From 2010 onward it is +34.3% against +32.8%, which is a gap pointing the wrong way. The sample shrinks to 52 and 39 days respectively, so neither sub-sample proves anything on its own, which is the point: the median comparison is not resting on evidence strong enough to survive a reasonable cut of the data, and the +21.8% against +28.7% headline should be read as a fact about 2003 to 2026 taken whole rather than a property of compressed markets.
If one sentence has to carry this: at a matched VIX level, compression has not been followed by more big spikes in any slice of the record we tested. How much smaller the typical move was depends on which decade you include.
It is not going quiet. It is chopping.
The coiled-spring image has a second problem, and this one does not need a base rate at all.
A spring stores energy because it stops moving. If the VIX were doing that, its day-to-day changes would be shrinking alongside its range. They are not. The 20-session standard deviation of the VIX's daily point changes sits at the 45th percentile of post-2003 readings. On the full history back to 1990 it reads 47th, and on log changes rather than point changes it reads 55th. Every version of the measure lands in the middle.
So the index is confined to a band at the 7th percentile while moving day to day at the 45th. It is covering ordinary ground and keeps hitting the same ceiling and the same floor. That is a market with an active disagreement about risk and a resolution to it that keeps arriving, not a market holding its breath.
Nothing is being stored. The two measurements are describing different things, and the metaphor has quietly assumed they are the same thing.
Fifteen episodes, in thirteen different years
One reasonable worry about 579 days is that they are all the same market. They are not, entirely. Requiring six months between one qualifying day and the next, the record holds 15 prior compression episodes plus the one running now, spread across 13 separate calendar years from 2003 to 2024. They happened under four Fed chairs, before and after the financial crisis, and before and after the 2018 volatility-product unwind.
The episode table below is description. It is where and when, nothing more.
The result we computed and then threw away
We also ran the forward test on those 15 episodes directly, one observation each, to sidestep the overlap problem. It produced a clean-looking number pointing the opposite way to everything above.
We are not publishing it, and why is more useful than the number would have been.
It flips sign on a knob with no right setting. Whether two compressed stretches count as one episode or two depends on how far apart you require them to be. Set that gap at 30 or 60 sessions and the episode median comes in below the all-days baseline. Set it at 90, 120, 180 or 250 and it comes in above. The data did not change. Only an arbitrary choice did, and there is no principled argument for 120 over 60.
It is not monotone in the threshold either. Tightening or loosening the compression cutoff from the 4th to the 20th percentile moves the episode median around in no consistent direction, and it crosses the baseline somewhere between the 12th and the 15th percentile. A real effect gets stronger as you tighten the condition. This one wanders.
And n is 15. With 15 observations the median is literally one of them, the eighth. The rate at which episodes reached +50% rests on seven cases, three of which were +111%, +146% and +288%. The last is late July 2015, three weeks before the August 2015 flash event. No base rate built from 15 draws can price a tail like that, and one that tries will mostly be reporting which singular events happened to land inside the window.
Fifteen is enough to say compression has occurred in many different market regimes. It is not enough to say what follows it. Those are different questions and the same 15 episodes cannot answer both.
Why the 579 figure needs its own warning
The day-level test has the opposite problem and it needs stating plainly rather than buried.
579 compressed days are not 579 independent observations. They come from roughly 15 episodes, and days inside an episode share almost all of their forward window with each other. Two consecutive compressed days are asking about nearly the same 30 sessions. The effective sample size is far below 579 and closer to the number of episodes.
This is why no confidence interval appears anywhere in this piece. Computing one from 579 would assume independence that is not there, and it would produce a tight band around a number that deserves a wide one. The honest statement is directional: across the compressed days in the record, big spikes were less common, and that direction survived every level band and period cut we tried but one.
The same test, applied to a different kind of claim
The Claim Checker on this site took 124 posted trading rules and tested each one against simply holding the asset. Two passed. Neither beat holding.
Those were rules: buy here, sell there, do this when that happens. The coiled-spring claim is not a rule. Nobody is telling you to trade it. It is a belief about conditional probability, and it gets repeated precisely because it has no entry price attached and therefore never gets scored.
That is what makes it worth counting. A trading rule at least announces itself as testable. A claim shaped like an image does not, and so it survives on how well it sounds rather than on how often it has been right. The method is the same in both cases. State the condition, find every historical instance of it, and look at what followed without deciding in advance what you want to find.
What remains unknown
Whether the surviving result is causal or compositional. Compression at a given VIX level may simply be a marker for market conditions that are themselves calmer, in ways this measurement does not see. The test establishes that the two go together in the record. It does not establish that the compression is doing anything.
What the modern sample would say with more of it. From 2010 onward there are 39 level-matched compressed days. The direction of the spike-rate result holds there. The sample is far too small to call that confirmation, and the only cure is time.
Why 2003 and 2004 behaved so differently. Those two years supply most of the median gap and we have no explanation for it beyond naming the period. An explanation would either strengthen the result or dissolve it, and we do not have one.
Whether a 30-session window is the right one. It was chosen before any result was computed, which is the part that matters, but it was not chosen for a reason that would survive challenge. A shorter or longer horizon is a different question and was not run.
What a second vendor's VIX would show. Every bar here comes from one source, TradingView's TVC:VIX daily series, cross-checked against the live in-page quote but not against a second independent vendor. The pre-2003 close-only problem was found by inspection rather than by comparison, and a second source might reveal others.
Educational commentary only. Nothing here is advice, a forecast, or a recommendation.