Strategy Framework
Volatility Didn't Fail. The Sizing Rule Did.
A reported volatility-fund drawdown is a lesson in left-tail survival, not a reason to declare short-volatility dead.
2 June 2026 · YK Research
Contents
The Mispricing
Bloomberg reported that QVR Advisors, Benn Eifert's volatility firm, was closing its hedge fund after a roughly 30% loss from January through April, with assets having reached about $1.6B at their peak earlier in the year. I do not know QVR's exact book. This is not a post about mocking another trader. Eifert's public writing has shaped how many people, me included, think about volatility markets.
The useful lesson is broader: in volatility, the trade can be intelligent and still be sized wrong. Volatility is not broken when it hurts. That is what it is supposed to do. If the left tail never arrived, nobody would pay investors to absorb it. And right now the tape makes the sizing mistake easy, because the headline number says calm while the names underneath do not.
Read the Tape: The Index Is Calm, AI Vol Is Not
Start with the number everyone watches. VIX closed at 16.0 on 1 June, the 24th percentile of its own trailing year and below the 18.1 one-year mean. Twice in the last twelve months it spiked, to 26 in late November and toward 31 intraday in late March, and both times it round-tripped to the mid-teens within weeks. By the only gauge most people check, this is a calm market. That is exactly the setup that makes a vol seller feel underpaid for the risk they are actually holding.
Hold that picture. The rest of this note is about why a book calibrated to that calm index can be badly mis-sized, and what a survivable sizing rule looks like instead.
Market-Neutral Is Not Risk-Neutral
A volatility book can look clean on a risk report. Delta-hedged. Beta-neutral. Diversified across underlyings. Low historical correlation to equities. Sensible model. Reasonable signal. Long record of attractive carry.
Then the wrong variable moves. The risk is not only that stocks fall. The risk is the full stress path: implied volatility reprices, skew steepens, correlations rise toward one, hedges become expensive, liquidity thins, margin pressure increases, and rebalancing has to happen at the worst possible time.
What The Risk Report Can Miss
- A book can be delta-neutral and still be short gaps.
- It can be index-neutral and still be short correlation.
- It can be diversified and still be short liquidity.
- It can be statistically hedged and still be short the assumptions inside the hedge.
The model may not be wrong in a normal sense. The signal may still contain information. The trade may still have positive expected value. But if the distribution widens, the old position size quietly becomes leverage.
The Trade Can Be Right. The Size Can Still Be Fatal.
Most carry strategies fail for the same reason: they pay often enough to feel safe. A strategy that earns small, steady returns creates psychological comfort. It produces clean monthly statements. It makes the Sharpe ratio look scientific. It encourages larger allocations because the bad state has not arrived recently.
But Sharpe is not survival. Sharpe measures the smoothness of the road already traveled. It does not prove the road still exists.
Same drawdown limit. Same signal. Different tail assumption. The correct size changes before the backtest proves it.
Old-Regime Sizing Is How Carry Trades Die
The most dangerous moment for a systematic volatility book is not when the model obviously stops working. It is when the model still appears to work, but the payoff distribution has changed.
The screen still says the belly of the curve is rich. The model still says implied volatility is high versus realized. The dispersion signal still says index vol and single-name vol are misaligned. The optimizer still says the position is within risk limits. But the market has changed the cost of being early.
Cem Karsan's regime argument matters here. If baseline volatility is repricing higher because of fiscal stress, inflation uncertainty, geopolitics, and weaker central-bank suppression, old signals do not have to become useless. The subtler risk is that the signal remains useful while the size attached to it becomes wrong.
The AI-Volatility Example, In Data
This is especially relevant in AI-linked equities, and it is where the abstract sizing argument meets a measurable market. AI is the defining equity theme of this cycle. It drives earnings revisions, capex expectations, retail attention, index concentration, option activity, and narrative momentum. So look underneath the calm index. The AI complex is not pricing calm at all. NVDA carries 42% ATM implied vol, AVGO 62%, AMD 71%, ARM 94%, MU 104%. The average across the AI leaders is around 68%, roughly three times QQQ at 21% and nearly five times SPY at 14%. That spread is implied dispersion: the market pays up for single-name moves while pricing the basket as quiet.
The index looks tame only because the components are assumed to move in different directions on different days. A book spread across multiple AI winners may still be short one common factor: the market's willingness to keep underwriting the AI narrative at current valuation, liquidity, and volatility levels. Every basis point of low implied correlation between those red bars and the blue ones is something the book is quietly short, and correlation is the first thing a regime break sends toward one.
And the repricing the VIX print is waiting for has already happened in the names that drive the book. Implied vol in the AI hardware names is not just high, it is pinned at the top of its own one-year range. AVGO, MU, AMD, and ARM all sit at an IV rank of 100, and every one of them is in backwardation, near-dated vol bid above longer-dated, the classic shape of a market pricing an imminent event rather than calm carry. Meanwhile SPY and QQQ sit in contango at IV ranks of 16 and 50. The stress is concentrated, not absent.
Worse, the carry is thinnest exactly where the tail is fattest. Plot the volatility risk premium, implied minus realized vol, against beta and the relationship runs the wrong way. AVGO and MU still pay a real premium, 23 and 15 vol points, on the lowest betas in the group. But ARM and AMD, the two highest-beta names at 3.4 and 2.4, trade at a negative premium: realized vol is already running above implied, so you pay to hold them. NVDA and PLTR sit near zero. The names most likely to gap in a beta-driven selloff pay you least, or nothing, to warehouse their risk right now.
A sizing rule that scales by historical Sharpe or trailing carry would load up on exactly these names, because the spreadsheet still reads them as cheap insurance to sell against a calmer window. The live data says the opposite. The edge is not "AI stocks move a lot." Everyone knows that. The possible edge is more specific: the options market may misprice how single-name event risk, index concentration, and crowding interact. But if that is the edge, sizing is the product. Not the scanner. Not the signal. Not the clever vol surface chart. Sizing decides whether the trader is still solvent when the best opportunities finally appear.
A Better Volatility-Sizing Rule
Stress before expected return
Model liquidity as pro-cyclical
Treat confidence as a warning
What Would Change My Mind
A regime-break argument needs disconfirming evidence. Otherwise it becomes a story that can explain anything. The good news is that this version is measurable.
The old sizing framework deserves more confidence if the data flips: semis IV ranks falling from 100 toward the bottom half of their range, term structure rolling from backwardation back into contango across NVDA, AMD, AVGO, MU, and ARM, the implied-realized premium turning positive again in the high-beta names, and the AI-versus-index dispersion ratio compressing from ~3x back toward 2x, all while liquidity depth improves and correlation spikes fade quickly.
If that evidence appears, recent volatility-fund drawdowns may prove to be a crowded-positioning event rather than a structural break. Until then, the safer assumption is that volatility is being paid for a reason, and that the index print is lagging the names.