The 895x Gap Behind Volmageddon. Credit Suisse’s $500M Question Still Isn’t Settled.
The vehicle that broke in 2018 is gone. The mechanism relocated into 0DTE options, now a majority of S&P 500 options volume.
A 3.8-standard-deviation day in February 2018 exposed an 895-times gap between what a Gaussian risk model treats as impossible and what Markov’s own inequality never ruled out. The specific vehicle that blew up that day has since been delevered, shuttered, or blocked by every major broker. The gap itself has not shrunk. It has relocated into a market that barely existed in 2018 and now carries a majority of all S&P 500 options volume.
On February 5, 2018, the S&P 500 fell 4.2%, a move the Bank for International Settlements measured at 3.8 standard deviations. Under the normal distribution most parametric risk models implicitly assume, a move that size happens about once every 55 years; under Chebyshev’s inequality, the distribution-free descendant of Markov’s inequality that assumes nothing about the shape of the return distribution at all, the same event carries a ceiling near 6.5%, roughly 895 times looser. That gap is not a historical curiosity. The specific products that failed that day (XIV, and the -1x version of SVXY) are gone or delevered, verified by the issuer’s own SEC filing within three weeks of the event. But the underlying condition that produced the gap, parametric models calibrated during a period of unusually low realized volatility, sitting under leveraged, mechanically-rebalanced, negative-convexity exposure, now shows up in a market that scarcely existed at the time and has since become one of the largest corners of US options trading.
The consensus is correct, as far as it goes
Ask any desk why nobody uses Markov’s inequality for risk management and the answer is fair: it is famously, almost comically loose. A 2023 paper working through the traditional Markov bound against the true tail of an exponential distribution found it overstating the tail probability by roughly 145 times at the 99.9th percentile, and by several orders of magnitude further into the tail. Stanford’s own introductory treatment of the inequality makes the same point with a coin-flip example: Markov’s bound puts the probability of 16 or more heads in 20 flips of a p = 0.2 coin at 25%, when the true probability is close to 1.4 x 10⁻⁸, a bound about 18 million times too wide. That is the price of a result that has to stay valid for every distribution consistent with a given mean, however strange, however fat-tailed. Nobody sizes a book off a number that far from the truth, and nobody should. The critique is correct.
Why the loose bound is the point
Markov’s inequality requires the variable to be non-negative: for X ≥ 0 and a > 0, P(X ≥ a) ≤ E[X]/a. Returns are not non-negative, which is the first objection a sophisticated reader should raise. The fix is not to wave the requirement away, it is to apply the inequality to the loss, L = max(0, -R), a non-negative quantity by construction. Markov’s inequality then bounds P(L ≥ a) directly, using nothing but the expected loss: no variance, no distributional shape, no model.
Add variance and the bound sharpens considerably. Squaring the deviation from the mean and applying Markov to (X - μ)² produces Chebyshev’s inequality, P(|X - μ| ≥ kσ) ≤ 1/k², a derivation both Stanford’s notes and a University of Washington concentration-inequalities course walk through identically. A one-sided refinement credited to Cantelli in 1928 sharpens this further for a purely one-directional loss event: P(X - μ ≤ -kσ) ≤ 1/(1 + k²).
What matters here is not the algebra, it is the decay rate. A Gaussian tail collapses exponentially in k², so by k = 4 or 5 a parametric model assigns a probability that is effectively zero. Chebyshev and Cantelli decay only polynomially, in 1/k², because they must stay valid for distributions that do not thin out that fast. At k = 2 the two views sit roughly an order of magnitude apart. By k = 4 they are separated by more than three orders of magnitude. The gap does not grow with the size of the move, it grows explosively with it, which means the two frameworks disagree most sharply in exactly the regime where a parametric model’s tail assumption is doing the most unverified work.
February 5, 2018, in the numbers
Short volatility had been one of the decade’s quietest trades. The VIX closed at an all-time low of 9.14 on November 3, 2017, and VelocityShares’ inverse VIX note, XIV, rose from $6.51 at the end of 2011 to $134.44 at the end of 2017, a 20x return, per a market commentary hosted by Cboe on its research-publications page; the document itself is authored by a third-party asset manager, DGV Solutions, and is self-labeled as commentary rather than research, so it is cited here only for objective price history, not for its own market views. Assets across the small group of leveraged and inverse VIX exchange-traded products reached roughly $3.5 billion by early February 2018, per a peer-reviewed study in the Financial Analysts Journal by Augustin, Cheng, and Van den Bergen, a figure consistent with BIS’s own independent estimate of roughly $4 billion at end-2017 once smaller products beyond XIV and SVXY are included.
On the day itself, the S&P 500’s 4.2% decline was the 3.8-standard-deviation move BIS later calculated, and the VIX moved from 17.31 to 37.32, a 20-point, 115.6% jump, the largest single-day percentage increase in the index’s history and more than double the prior record of 64.2% set in February 2007, again per the Cboe-hosted DGV commentary cited above; BIS corroborates the order of magnitude independently, describing the same move in its own prose as roughly a 20-point jump and the largest daily VIX increase since the 1987 stock market crash, without restating the precise decimal figures. A peer-reviewed analysis in the MDPI journal Risks put a number on the surprise: given the VIX’s typical -0.8 historical correlation with the S&P 500, a 4% equity decline should have produced a roughly 3.2-point rise in the VIX, not a 20-point one.
The mechanism was structural, not purely sentiment-driven. BIS’s own transaction-level analysis documents that both long and short volatility ETPs needed to buy VIX futures near the 4:15pm close to maintain their target exposure, a rebalancing collision that pushed 115,862 futures contracts, roughly a quarter of the day’s entire volume, through the market within a single minute at 16:08. The Augustin, Cheng, and Van den Bergen study adds the input that made this collision so dangerous in the first place: between 2007 and 2017, the S&P 500 VIX Short-Term Futures Index carried an average 90-day trailing volatility of 64.0% (the spot VIX itself averaged 117.2%), versus 17.4% for the S&P 500, and by late 2017 the S&P 500’s own trailing volatility had compressed further, to about 6.8%. Any model borrowing equity-scaled tail assumptions for this product class, calibrated during an unusually calm stretch, was miscalibrated before the shape of the tail even entered the picture.
The value of XIV fell 84% during the regular session and the product was terminated, triggering the acceleration clause in Credit Suisse’s own prospectus, which permitted termination once the note’s intraday indicative value fell to 20% or less of the prior day’s close. The closing indicative value on February 2 had been $108.3681. That 84% figure is BIS’s own number for the regular session specifically; by the time VIX futures finished spiking past 50 in after-hours trading, Six Figure Investing puts the full-day collapse at 97% for XIV, 91% for SVXY, and 87% for VMIN, the fuller unwind that the acceleration clause was ultimately settled against.
The other side of that trade has a name. Set Capital LLC v. Credit Suisse Group AG, a securities class action, alleges that Credit Suisse’s own hedging of its XIV exposure helped manufacture the liquidity squeeze that crashed the notes, clearing the way for the bank to redeem them at the crashed price. The Second Circuit revived the market-manipulation claims in 2021 after a district court had dismissed them, finding the allegations plausible enough to proceed; the investors’ own counsel puts Credit Suisse’s resulting profit at $475 million to $542 million, and a federal judge granted partial class certification on the manipulation claims in February 2025. The case has not gone to trial and the allegations are unproven. What is not in dispute is the shape of the trade: someone was structured to gain close to what retail noteholders were structured to lose, and it was the bank that sold them the note.
The gap, computed
Set the litigation aside and return to the number that started this. Take BIS’s 3.8 figure and run it through both frameworks. Under a normal distribution, the one-sided probability of a move that size or worse is about 0.0072%, or roughly one in 13,800 trading days. Under the Cantelli bound, which requires only the mean and variance and assumes nothing about shape, the ceiling on the same event is 6.48%, or roughly one in 15 observations. The Cantelli ceiling is about 895 times the Gaussian point estimate.
Run the same exercise one level down, on the VIX futures index itself, using Augustin, Cheng, and Van den Bergen’s own 64.0% realized-volatility figure. Converting that to a daily figure (dividing by the square root of 252 trading days, the standard annualization convention) gives a daily volatility near 4.0%. XIV’s 84% one-day fall implies, under a -1x daily-rebalanced product where a single day’s fund return approximates the negative of the index’s return, that the underlying futures index itself moved by a comparable magnitude that day, roughly 21 standard deviations by this measure. At that point the Gaussian framework does not just underestimate the tail, it stops producing a usable number: the implied one-sided probability is on the order of 10⁻⁹⁴ percent, a figure with no operational meaning. The Cantelli ceiling, built from the same inputs, still returns something a risk committee could act on: about 0.23%, or roughly one in 435. This is the sharper version of the same point. A distribution-free bound degrades gracefully as an event gets more extreme. A parametric one does not degrade, it fails silently, producing a confidently-stated number that has stopped meaning anything long before a risk manager would notice.
This is worth running as a live check, not only in hindsight. Cboe’s own VIX data shows the index closing at 16.45 on July 1, 2026, the day before this piece was finalized. Using implied rather than realized volatility as the input this time, match Volmageddon’s exact magnitude, a 4.2% one-day S&P 500 decline, against that single number, and the move works out to roughly 4.05 standard deviations: slightly more extreme than the 3.8 BIS calculated from realized volatility for the actual 2018 event. The Cantelli ceiling on a move that size is about 5.7%; the Gaussian estimate is about 0.0025%, a gap of roughly 2,270 times. The exact multiple will be different by the time this is read, since implied volatility moves daily. That is the point: the check costs nothing more exotic than the VIX print and can be rerun every morning, which makes it a live number rather than a 2018 artifact.
Neither of these ceilings is the true probability. What the gap measures, at any of these levels, is how much of a model’s stated confidence comes from something verified in the data (the mean and variance) versus something assumed on top of it (the tail’s shape). The only thing standing between “basically impossible” and “small but real” is a curvature assumption nobody in the chain re-verified before February 2018.
Why this is a decayed vehicle, not a decayed mechanism
An honest reading of this case has to ask whether the industry already fixed it. It largely did, for the specific products involved. ProShares’ own SEC filing, dated February 26, 2018, three weeks after the event, announced that SVXY’s target exposure would be cut from -1x to -0.5x and UVXY’s from 2x to 1.5x, effective the next trading day. XIV was terminated outright. VMIN, a smaller competitor, was wound down by November 2018 for lack of assets. Vanguard, Fidelity, and Merrill Edge each restricted retail access to leveraged and inverse volatility products within the following year. A simulation from Six Figure Investing, cited above, estimated that a repeat of the February 5 move against the post-delevered SVXY would produce roughly a 48% loss today, severe but survivable next to the 87% to 97% full-day losses realized in 2018. Current fund-data snapshots vary by provider and update lag: a same-day check of TradingView puts SVXY at $218.21 million, while Yahoo Finance and U.S. News, sourced to Morningstar, have shown figures ranging from roughly $190 million to $250 million across recent weeks, and Danelfin puts SVXY’s -1x successor SVIX in the neighborhood of $190 million to $211 million depending on the snapshot date. Combined, that is somewhere in the $400 to $460 million range, against a 2018 peak of roughly $3.5 to $4 billion across XIV, SVXY, and VMIN depending on the cutoff date and which products a given source includes, a reduction on the order of 87 to 90% under any reasonable reading of the range. Judged purely on the 2018 vehicle, this is a decayed setup: smaller, less leveraged, harder for a retail account to reach, and already the subject of a peer-reviewed forensic account with “crowded trades” in its own keyword list, on top of active federal litigation over who profited from it.
The mechanism is a different question, and whether it can recur in a new vehicle is not settled, it is actively and specifically contested. Distribution-free tail checking is not a crowded trade in the conventional sense, because it is not a return-generating signal that gets arbitraged away as more capital adopts it; running it does not consume anyone else’s ability to run it. So the relevant capacity question is not the AUM at which this stops working, it is how much capital currently sits in structurally similar exposure priced by models with the same blind spot. Zero-days-to-expiration options share the 2018 products’ core ingredients (leverage, mechanical hedging by market makers, and pricing built on realized volatility from unusually calm recent windows), and they have grown from 21.5% of total US listed options volume in 2024 to 24.1% in 2025, citing OCC and Cboe data, reaching more than half of all S&P 500 index options volume by the fourth quarter of 2024. Whether dealer hedging in that market can produce a Volmageddon-style collision, not merely share its ingredients, was tested in public on August 15, 2023, when the S&P 500’s decline accelerated by roughly 0.4% in twenty minutes. Goldman Sachs’s Scott Rubner attributed the acceleration to 0DTE-driven dealer hedging, as Bloomberg reported the next day. Cboe’s own gamma-exposure reconstruction of that same afternoon found market makers net long gamma, meaning their hedging should have dampened rather than amplified the move, until 3:30pm, by which point the index had already stabilized, and concluded the data does not support 0DTE hedging as that day’s driver. Two data-facing institutions read the same twenty minutes and reached opposite conclusions. Bank of America’s Nitin Saksena has separately argued the risk is overstated, while relaying an alarm raised by others that positioning could produce an event echoing Volmageddon; in the same reporting, J.P. Morgan’s Marko Kolanovic warned that 0DTE-driven swings could reach $30 billion in a single session. None of this proves the mechanism will recur at Volmageddon scale in this market. It shows the mechanism is contested by name, in public, by desks with access to the actual positioning data, inside a market that is now a majority of all S&P 500 options volume, on a scale the 2018 VIX-ETP complex, a few billion dollars in assets at its peak, never approached.
This isn’t a new idea. It’s an abandoned one.
Distribution-free, worst-case tail reasoning is not a novel proposal for portfolio construction, it is close to where modern portfolio theory began. A.D. Roy’s 1952 paper, “Safety First and the Holding of Assets,” proposed choosing a portfolio to minimize the probability that its return falls below a disaster level, d. Because Roy did not want to assume a specific return distribution, he bounded that probability using a one-sided form of Chebyshev’s inequality, which reduces the minimization problem to maximizing (μ - d)/σ, the same algebraic form, over a decade before it would be popularized as the Sharpe ratio.
Roy’s paper appeared the same year as Markowitz’s “Portfolio Selection.” Harry Markowitz himself, in a 1999 historical essay quoted in Mark Rubinstein’s 2002 retrospective in the Journal of Finance, described his own “father of modern portfolio theory” title as one Roy deserved to share equally. What separates the two founding contributions is not the ratio, which converges to the same form, it is the justification underneath it. Markowitz’s framework optimizes over an assumed joint return distribution. Roy arrived at an identical-looking ratio specifically to avoid needing one. Over the decades since, the industry kept the form (every risk-adjusted-return metric from Sharpe to Sortino to the information ratio is a variation on (numerator - benchmark)/dispersion) and let the distribution-free caution that originally justified interpreting a given number of standard deviations fall away, replaced by parametric or historically-simulated confidence intervals that require exactly the assumption Roy built his ratio to avoid needing.
Three objections, all worth taking seriously
The first: a 6.5% ceiling is too loose to act on, and nobody can post capital against a bound that wide without crippling the book. This is correct, and it is not the proposal here. Using the Cantelli ceiling as a replacement point estimate for VaR would produce a worse model, not a better one, for the same reason the consensus section above concedes. The actionable version is narrower: track the ratio between the parametric estimate and the distribution-free ceiling for any book carrying negative convexity, not the level of either number alone. Del Castillo’s 2023 paper makes the relevant point about why the ratio, not the raw bound, is the useful object, noting that avoidance of model risk is decisive when multiple competing models are present in a real-world situation. A ratio that stays in the single digits says the parametric model’s tail assumption is doing modest work. A ratio in the hundreds says the position’s entire safety margin rests on an unverified assumption about curvature.
The second: safety-first portfolio construction was tried and superseded by mean-variance and expected-utility approaches for good reason, because a criterion built on a worst-case bound is systematically overcautious and leaves return on the table in ordinary markets. That is also correct, and it is not an argument for reviving Roy’s allocation framework wholesale. Mean-variance optimization dominates safety-first sizing in normal regimes, and nothing here disputes that. The claim is narrower still: not that the criterion should return, but that the specific discipline behind it, checking a claimed tail probability against what the data alone can support, was discarded along with the criterion, and nothing in modern risk architecture replaced it.
The third, and the one a reader of the decay section above should raise directly: this whole case is 2018 news, already fixed by the very deleveraging and product terminations documented three sections up. Partially, and that partial concession is the point rather than a weakness in it. The specific instrument decayed. The condition that produced the gap, a parametric tail assumption unverified against the position’s own dispersion, is a property of how risk gets modeled, not of any single ticker, and the 0DTE evidence above, contested as it is, shows that condition currently sits under a larger, faster-growing, and actively disputed book of exposure than the one that broke in 2018. A technique with no crowding mechanism does not become less useful because the specific product it was first demonstrated on stopped trading.
What would change this view
Two things would weaken this argument considerably. First, public filings cannot show whether a given fund already runs a check like this. Quarterly 13F filings disclose long, US-listed equity positions only, as of quarter-end, on a 45-day reporting lag; they exclude derivatives, short positions, and non-US holdings entirely, which means the leveraged and short-vol exposure this piece is about would never appear in one regardless of which fund holds it. No standard SEC disclosure type, 13F included, reveals internal risk methodology. If distribution-free floors are already standard practice inside systematic multi-strategy risk committees, or inside the market-making desks now absorbing the bulk of 0DTE flow, the gap described here is being managed quietly rather than sitting unused, and the piece’s urgency, though not its math, would need revising.
Second, the 21-standard-deviation figure for the VIX futures index rests on a single realized-volatility input, an 11-year average from one peer-reviewed source, applied to XIV’s return as a proxy for the underlying index’s move. A different vol window, or a direct computation from the futures index’s own tick data rather than a leveraged product’s daily NAV, could move that specific number materially, though it would need to move by several full orders of magnitude to change the qualitative conclusion that a Gaussian framework produces an uninformative answer at this level. The equity-level 895x figure rests on firmer ground, a single BIS-calculated input with no proxy step, and is the number this piece’s actionable claim should be weighted toward. And the Goldman-versus-Cboe dispute over August 2023 cuts both ways by design: it is evidence the mechanism is live enough to argue about, not evidence it will replay at 2018 scale.
The actionable version, and its capacity bound
For any book carrying negative convexity or embedded short optionality, compute the Cantelli ceiling from nothing but the position’s own realized variance and post it next to whatever parametric VaR or expected-shortfall figure the desk already produces. Basel’s own post-crisis review moved capital requirements from VaR toward Expected Shortfall precisely because pre-crisis models systematically underpriced tail risk, and Expected Shortfall remains a parametric or historically-simulated measure, not a distribution-free one. Watch the ratio between the two numbers, not either level in isolation. This practice has no AUM ceiling of its own, because it is a governance check, not a factor that gets crowded, but it is only worth running on exposure large enough that a Volmageddon-scale ratio would matter to the book’s survival, which in practice means any negative-convexity sleeve above the low seven figures at a systematic shop, and any book at all inside a market-making or dealer operation now absorbing 0DTE flow at more than half of S&P 500 options volume. Run that check before the tail event, the way the data already allows, not after it.
📊 Want Deeper Quantitative Analysis?
This research took a long time: data collection, verification, and analysis. If you found value in this deep-dive, I publish exclusive quantitative research, trading strategies, and institutional-grade analysis on Patreon.
By joining, you’ll be supporting my work and motivating me to publish more content like this.
→ Join the Patreon community here
For more research like this, subscribe to The Mathematical Trader on YouTube, or connect with me on LinkedIn.