The Squeeze, Under Oath

GameStop and AMC are the poster children of an era — the two tickers that taught a generation of retail traders a new folklore: the gamma squeeze, the whale alert, the earnings lottery ticket, the conviction that options flow is a machine you can point at a stock and fire. I traded the autumn-2021 echo of that folklore, and the first article in this series counted what it cost me. This one goes back to the source and puts the legends under oath.

The evidence: the complete GME and AMC option chains, January through June 2021, at one-minute resolution — every quote, every OPRA print, every greek. For cross-examination, two other panics off the same tape: Volmageddon — February 5, 2018, the day the short-volatility trade died — including the original VXX note, and the yen-carry unwind of August 5, 2024. An hourly panel around 535 earnings reports, 2016 through 2024, in eighteen liquid names. And 9.18 billion rows of 2025–26 minute-level option chains as the calm-market control. Eleven billion rows in all, interrogated the way the first article interrogated my broker confirmations — the same method, pointed at the stories every retail options trader tells: it was a gamma squeeze; the smart money telegraphs its moves; earnings straddles print; the spread doesn’t matter if you’re right.

To be clear up front: no positions were taken. Nothing here is a backtest of a strategy I claim I would have run. This is measurement — what the quotes and prints actually did, in the places where the folklore says the bodies are buried.

PART I — THE FOUNDING MYTHWas GameStop actually a gamma squeeze?

The story everyone tells: retail bought calls, dealers who sold the calls had to buy stock to hedge, the buying pushed the price up, which forced more hedge-buying — a feedback loop, a machine, a gamma squeeze. It’s the founding myth of modern retail options trading, and in 2021 I believed it enough to trade like it was a law of physics. The tape lets you check.

The standard yardstick is dealer gamma exposure — GEX — the dollars of stock dealers must buy or sell per 1% move to stay hedged, computed from open interest and each contract’s gamma. I computed it daily for GME across the whole first half of 2021, under the standard convention. Two things jump out of the chart below. First: through the famous week, GEX was tiny — between +$1.1M and +$7.4M per 1% move in the run-up, briefly negative (−$0.9M) on January 28. Second: the biggest GEX day of the entire six months wasn’t in January at all. It was May 27, at +$33.4M per 1% — four months after the squeeze, on a day nobody wrote folk songs about.

FIG. 1 — The engine that wasn’t. Through the squeeze week (shaded), naive GEX ran $1–7M per 1% — against days when $7.5B of options premium changed hands. The mechanism’s biggest day came in May, to no effect anyone remembers.

Scale is the argument. On January 27, the biggest single day of call premium in the whole six-month window — $7.47 billion gross traded, against $1.13B in puts — the GEX print says a full 100% move would have obligated dealers to trade on the order of a billion dollars of stock even under the most aggressive assumption, that they were short every open call. A billion dollars of forced hedging, on a day when seven and a half billion of options premium alone changed hands, and vastly more in the stock itself. The feedback loop existed; it was also a rounding error.

And the loop was actively dismantling itself. Gamma shrinks as implied volatility explodes — at the IVs GME options reached, a contract’s gamma is a fraction of its calm-market self — so the more the squeeze ran, the less gamma there was to squeeze. By January 27 the front-weekly at-the-money call carried a median IV of 937%. The options market wasn’t an accelerant by then. It was a very expensive scoreboard.

Open interest tells the same story from another angle, and it’s the part almost nobody quotes:

Call open interest fell into the parabola — positions were being closed over the January 22 weekend, not opened — while put OI tripled in five sessions. Billions in “call buying” on falling call OI means the premium was churn: opens, closes, day-trades, paper passing hands. And the minute data delivers the verdict on causality. Over January 25–27, pooled at 15-minute horizons across 1,128 minutes: call premium flow predicting the next fifteen minutes of price, correlation +0.06. Price over the past fifteen minutes predicting call flow: +0.19. On January 26 the “lead” correlation was actually negative (−0.11). The honest summary of +0.06 is no detectable lead at all.

Here is the fair version of the myth — the narrower one the tape does support. In mid-January the squeeze mechanics were real: if dealers had hedged every open call on January 27, they’d have held 40.7 million shares — roughly 58% of GME’s ~70M shares outstanding, ~80% of the float — and that figure had jumped 12.5M shares in a single day on January 13, the +57% day. An options market obligating most of a stock’s float is a real machine. So: gamma-squeeze mechanics plausibly powered the January 13–22 run-up. But the vertical part everyone means when they say “gamma squeeze” — the 61-to-469 parabola that ran from January 25 to the morning of the 28th — was a short squeeze and a buying stampede in which the options market was a symptom. For contrast, AMC in June is what a genuine call-accumulation wave looks like: call OI up 71% in three weeks, $4.52B of calls traded on June 2 alone, 171M delta-weighted shares, GEX peaking at +$38.1M per 1% — bigger than anything GME printed in January. And even that produced a rally, not a singularity.

One caveat carries the whole section: GEX’s sign convention (dealers long the calls, short the puts) is an assumption, and in January 2021 — when retail was the one buying calls — likely a wrong one. The load-bearing finding survives either sign: single-digit millions per 1%, in a stock trading billions. The magnitude, not the direction, is the confession.

PART II — THE REPLAYThe squeeze week, one minute at a time

Numbers summarize; the tape performs. Below is the squeeze week itself — January 22, then the five sessions of January 25–29 — reconstructed minute by minute: the price, the front-weekly at-the-money call’s implied volatility, its bid-ask spread, and the premium flowing through calls and puts, resetting each morning. Drag the scrubber, or press play and watch a market decompose in real time.

Watch for three things. The IV: a median 574% on January 22, rising to 1,199% by the 29th, with single minutes printing near 1,997% — an options market pricing the possibility that anything, including nothing, could happen. The spread: 3.2% of mid on January 19, 11.2% by the 29th, with minutes where the market was quoting a spread wider than most stocks move in a year. And the gaps — the flat-lined IV stretches on January 28, when the tape simply stops: trading halts, 97 minutes’ worth, the exchange’s own circuit breakers tripping over and over while the price gapped nearly $300 in a morning.

FIG. 2 — The scrubber. Price above, ATM implied vol below; the readout shows the quoted spread and the day’s running call and put premium at the selected minute. IV goes dark where the market did.

The thing the replay teaches that no daily chart can: the violence was administrative. January 28 isn’t a crash so much as a market being switched off and on fifteen times — price teleporting between halts, quotes vanishing, the ATM option unpriceable for minutes at a stretch. If your mental model of that week involves executing trades at the prices on the chart, the replay is the correction.

PART III — FIVE MINUTES TO FOURVolmageddon, minute by minute

Three years before GameStop there was a purer catastrophe. Through 2017, selling volatility — shorting VIX futures through products like the XIV note and SVXY fund — had become retail’s favorite perpetual-motion machine. On Monday, February 5, 2018, the machine ate its owners in a single afternoon. I pulled the minute tape for that week: SPY, VXX (the original 2018 exchange-traded note, long dead), and SVXY, February 1 through 9.

FIG. 3 — The three-tier panic. SPY’s ATM vol nearly doubles (14.3% → 27.9%); VXX’s reaches 2.46× its calm low; SVXY’s prints 417% — 7.2× — though by then its “at-the-money” strike was the lowest one listed, 2.5× the share price. Gaps = no valid quote.

The famous part of Volmageddon happened after the close — the VIX-futures melt-up between 4:00 and 4:15, the 5:15 p.m. indicative-NAV print that vaporized XIV. None of that is on the options tape, and I won’t pretend to show it to you: the last option print on February 5 lands at 3:59:59.836 p.m., and then the record simply ends, like a security camera cutting out before the vault door opens. What the tape does show is the before and after, and the before is the part worth studying, because the market knew. Watch the afternoon in minute detail:

FIG. 4 — Liquidity leaves first. Spreads are flat all morning, go vertical after ~2:30, and peak at 3:13–3:22 p.m. — 56.7% of mid on SPY’s ATM, 84.6% on VXX’s — while the price is still finding its low (3:11) in an orderly slide.

The compression of it still astonishes me. SPY’s 30-day ATM vol sat at its day low, 14.26%, at 10:50 a.m. — a boring Monday. By 3:14 p.m. it printed 27.91%: +96% in four and a half hours, with +43% of that in the thirty minutes ending 3:11 and the completion of the double arriving at Tuesday’s open (33.3%). SPY’s ATM spread, 1.01% on a calm day, averaged 8.62% in the last hour — 8.6× — and VXX’s ran 11.5× its calm self. A third of the entire day’s SPY options premium — of $2.195 billion — traded in that final hour. Skew told the same story: the gap between 25-delta put vol and call vol went from 3.5 points on February 1 to 14.4 on the 5th, then halved to 7.0 on the 6th — not because puts calmed down, but because upside call vol doubled to catch up. After a crash, the market prices violence in both directions.

And the positioning post-mortem, from the open-interest snapshots: SVXY — a fund about to lose ~83% of its value between sessions, $71.12 at Monday’s close to around $12 at the next observable print — carried 358,800 put contracts at an OI-weighted strike of $58. After the gap, 96.1% of that put book was deep in the money; its chain didn’t even list a strike below $30 until February 7. Then the flip: within four days SVXY’s put/call ratio went from 1.97 to 0.60 as call OI doubled — the hedges cashed out, and the bottom-fishers moved in. VXX’s put OI — 1.67M contracts of listed short-vol conviction at an average strike 46% below where the note closed — grew to 2.34M by February 9. The trade that had just detonated was being re-boarded before the smoke cleared.

PART IV — THE MORNING THE CARRY BROKEAugust 5, 2024, in 390 minutes

The third panic is the modern one: the unwind of the yen carry trade, Monday, August 5, 2024 — the day the VIX printed ~65 before the open and financial television briefly rediscovered the word “contagion.” The minute tape tells a drier story, and the drier story is more useful.

FIG. 5 — The panic that was over by 9:32. Monday’s 0DTE ATM vol peaked at 103.7% in the session’s second minute — the same minute region where SPY printed its low of the day — and had given back half the spike by 10:40.

Everything happened at the open, and then the market spent the day walking it back. SPY’s first quoted minute, 9:31: 0DTE ATM vol 102.4% (Friday baseline: 47.9%), 30-day vol 36.6% (Friday: 18.7%), and the price — 510.89 — which stood as the low of the entire cash session. The panic peaked at 9:32, at 103.7%. Half the front-vol spike was gone by 10:40 — a half-life of about 1.1 hours. The 30-day vol decayed slower, half-life ~2.4 hours, and never fully normalized that day, closing 9.5 points above Friday. Two weeks later it was 13.0% — below its pre-crash level, the whole episode priced out and then some.

About that VIX 65: at-the-money 30-day SPY vol never traded above 36.6% all session. The headline number lived in the deep out-of-the-money wings and in quote widths that made the wings barely priceable — the first minute’s 30-day ATM spread was 9.0% of mid, one minute printed 27.1%, and the open-half-hour averaged 4.6–4.9× Friday’s. QQQ ran about a third hotter than SPY throughout — 121.8% vs 91.0% on the 0DTE at 9:35 — which is what you’d expect when the unwinding trade was long tech. The cleanest visual of the day, though, is the term structure — the curve of vol against days-to-expiry, snapshotted five times:

FIG. 6 — Fear has a maturity date. At Monday 9:35 the curve was inverted by 56.6 points from 0DTE to a month out (+24.7 even excluding the 0DTE); by noon it had collapsed halfway; two weeks later it sat below the pre-crash Friday, upward-sloping, as if nothing happened.

PART V — ANATOMY OF A PANICThree crises on one yardstick

I pulled three panics expecting to find one animal in three cages. What the tape shows is three different animals:

2018 was a slow-then-sudden afternoon in which liquidity led: spreads went vertical forty minutes before the price low, and the actual kill happened off the listed tape entirely. 2021 was a rolling multi-day boil where the crisis was the market structure — halts, unpriceable quotes, four-digit vol — and the causal arrow ran from price to options, not the reverse. 2024 was over in a minute: the low, the vol peak, and the worst spreads all printed by 9:32, and the rest of the day was mean reversion. If there is one common thread, it’s this: in every case the option market’s prices stayed honest — wide, ugly, but honest — while its liquidity was what actually broke. The spread, not the vol, is where panic lives. Which raises the practical question: what does that cost?

PART VI — THE TOLLWhat crossing the spread actually costs, calm vs panic

Of everything the 2021 account paid to trade, commissions were the small half of the toll. The big half is the spread — the gap you cross going in and coming out. With prints matched to the same contract’s quotes, you can measure it exactly, and the folklore that “fills happen at mid” dies quickly: only 1.8% (SPY) to 7.9% (AAPL) of calm-market prints land within a basis point of the mid.

FIG. 7 — The toll booth. Calm markets: in the single names you pay roughly the quoted half-spread, 66–97 bps of the option’s price each way; SPY’s effective runs ~2× its tight quote, because its 0DTE-heavy mids move even inside 5 seconds. Panics: 1.9–4.3× the calm quotes. In the GME week, fills actually landed inside the blown-out quotes (42% of prints) — effective 149 vs 239 bps quoted at the moment of the print (256 across all quotes) — and still cost ~2× calm.

The gradient across the chain is the part I wish I’d known in 2021. Deep out-of-the-money options — the lottery tickets, |delta| under 0.1 — carry a median quoted half-spread of 285.7 bps, against 50.6 bps at the money: 5.6× more expensive per dollar traded, before they even start decaying. Short-dated is the same story: 0–1 DTE options cost 169.5 bps effective, drift-controlled — double their own quote, triple a 90-day option’s. Cheap options are not cheap; they are expensive options with small price tags. The 2021 account — median three days to expiry, strikes a rung too far out — was paying the maximum possible toll on every single round trip, and per the first article’s ledger it made 378 of them.

PART VII — 535 EARNINGS NIGHTSThe straddle, the crush, and whether either is predictable

Every earnings season the same debate: the options are “pricing in” an X% move — is it too much or too little? Buy the straddle if you think the market’s underpricing the gap; sell it to harvest the overnight collapse in implied vol, the famous crush. I ran the experiment on 535 earnings reports across 18 liquid names, 2016–2024: buy the at-the-money straddle at the last close before the report, sell it at the first live quote after. No selection, no timing, every event.

FIG. 8 — The house edge, mostly. Mean implied move 6.59% vs mean realized 5.43% (medians 5.59% vs 3.88%); the straddle overpriced the move in 66.4% of events. The dots far above the line are why anyone plays: NVDA May 2023 realized 25.6% against 6.8% implied — +279% on the straddle.

The at-the-money IV crush is real and brutally reliable: entry IV averaged 106.5%, exit 83.5% — 23 points gone overnight, a median relative drop of 24.3%, and that’s measured an hour into the reaction session, after part of the crush already happened. IV rose overnight in only 22.1% of events. So selling straddles prints, right? At the mid, mostly: the short straddle won 56.8% of nights with a median take of +6.28%. But the mean is −1.57% — the occasional NVDA-shaped monster eats months of harvest — and the mirror-image long straddle shows the same skew from the other side: mean +1.57%, median −6.28%. The typical night loses; the tails pay for everything. Where have I read that before.

Then the spread walks in and takes both sides’ lunch. Crossing it — buying at the ask, selling at the bid — drags the long straddle’s mean from +1.57% to −5.72%: a 7.28-point toll per round trip, roughly the size of whatever edge either side ever had. SMCI is the cartoon version: +18.77% mean at mid, −16.06% crossing the spread. The most honest sentence I can write about earnings straddles is that the market prices the move about right, the crush is real, and the bid-ask spread is wider than both facts combined.

FIG. 9 — Six years of house edge, then three years of the opposite. The long straddle lost at mid every year 2017–2021, then made +8.59%, +9.35%, +14.28% in 2022–24 — mega-cap earnings gaps outran what the straddles charged, even paying the spread.

Which brings the obvious next question: if the crush is this reliable, is it predictable — can you rank tonight’s IV against the same stock’s recent earnings and know when the straddle is rich? I ran the baseline: each event’s entry IV ranked against that name’s trailing eight events, 391 qualifying events. Correlation between the rank and the straddle’s return: +0.017. Nothing. Rank against the size of the crush: +0.107 — statistically present, explaining about 1% of the variance. The decile calibration isn’t even monotone: +11.5% in the lowest-IV decile, −12.2% in the second, +24.1% in the seventh, −3.7% in the top. “IV is high versus its own history” — the entire premise of every earnings-vol screener I have ever seen — carried no information the straddle price hadn’t already digested.

PART VIII — THE WHALE ALERTSDoes “unusual options activity” telegraph anything?

The last piece of folklore is the youngest: the “unusual options activity” alert, the $2M call sweep breathlessly screenshotted as evidence that somebody knows something. I took every options print of $200,000 or more across 23 liquid names in 2025 — 798,652 prints, $885.6 billion in premium — classified each as buyer- or seller-initiated from its position in the quoted spread, scored every symbol-day by net aggressive bullish premium, and asked what the top decile of “whale days” predicted.

FIG. 10 — The smart money, followed for ten days. Big bullish-flow days: +0.42%. Doing nothing: +1.11%. The days when whales bought puts: +1.61%. In 2025 the alert ran mildly contrarian.

The entire “edge” is a next-day pop of +0.175% against a +0.114% baseline — t-statistic 0.48, statistical silence. By ten days out the whale-alert days had lagged the ordinary drift of the same names (t = −1.97), and the heavy put-buying days had beaten it. Follow the alerts against fading them and the gap is t = −2.37 — the wrong sign for the story. The mundane explanations write themselves: 69.9% of these giant prints executed inside the quoted spread — negotiated blocks, not desperate sweeps; the single biggest print of the year, $943.6M of COIN puts, was deep in the money, almost certainly two institutions passing a position between them; and without open-interest confirmation you can’t tell an opening bet from a closing one, or either from one leg of a spread. One year, one regime, honestly caveated — but the folklore needed the effect to be large and positive, and it was small and backwards.

PART IX — THE PULLEleven billion rows through a residential internet connection

A word on what “the tape” means here, because the month of engineering is its own story — the chart above is its equity curve. The vendor’s terminal is a local Java process that speaks HTTP; everything else I wrote: a puller that decomposes a campaign into units — one symbol, one expiration, one month — fetches them through a worker pool, writes each to a parquet file, and logs every attempt to an append-only manifest. The manifest is the whole design. Crash, laptop sleep, auth failure, rate limit: on restart, the puller reads the log, skips what’s done, and re-queues what isn’t. Nothing is ever lost twice.

Final tally: 11,242,791,833 rows across 29,544 parquet files and 195.7 GB, in four campaigns over 32.6 days — of which only 225 hours had any write activity. Peak throughput, 135 million rows in a single hour; best ten-minute burst, 3.3 million rows a minute. Of 29,947 units attempted, 98.65% were eventually written; 350 were legitimately empty; 53 failed permanently, all in the glacial open-interest campaign, none touching the windows this article is built on. The manifests logged 10,796 errors along the way — and 99.5% of the units behind them were retried to success. The two worst nights deserve their own exhibit:

The first incident is the one I’d frame: I started a second session on another machine, the terminal enforced its one-login rule by killing the first, and the worker pool — doing exactly what I’d told it to — burned through the entire remaining queue in nine minutes, logging a failure roughly every 87 milliseconds. The design absorbed it: five hours later the same manifest that recorded the disaster replayed it as a to-do list. If the first three articles were about how I behaved under drawdown, this is the version of me that finally learned to write the discipline down where a crash can’t forget it.

PART X — THE CALM-TAPE PROBLEMWhat this means for training a model on quiet markets

There’s a reason beyond curiosity that I pulled the panics specifically. The other thing on that disk is 9.18 billion rows of 2025–26 minute-level option chains — the corpus I’ve been using to train small sequence models on market microstructure. Everything in this article is, from that project’s point of view, a single warning repeated five ways: the tape a model trains on and the tape that decides its worth are not the same distribution.

Concretely. A model trained on the 2025–26 chains has learned a world where SPY’s ATM half-spread is 59 bps; the GME window ran 4.3× that, and fills migrated inside the quotes — the calm-tape relationship between quote and fill inverted. It has learned IV surfaces that move points per day, not 48% in 44 minutes. It has never seen a session with 97 minutes of dead quotes, or a term structure inverted 68 points, or an “at-the-money” strike 2.5× the share price because the chain hadn’t caught up with reality. And the three panics it would need to generalize across don’t resemble each other — an afternoon avalanche, a five-day boil, an opening gap. Even the “calm” data has regimes hiding in it: the same mechanical straddle lost money for five straight years, then made high-single-to-double digits for three. Whether anything a model learns from quiet tape survives contact with a panic is not a question this data can answer — three events is an anecdote, not a distribution. I flag it because it is the open question of the project, and because the standard failure mode in this field is discovering distribution shift the expensive way. This article was, among other things, my attempt to make that discovery on paper instead of in a position.

CODAWhat the tape confessed

This is what eleven billion rows said under questioning. The gamma squeeze that founded the folklore was mostly a short squeeze wearing an options costume — the mechanism was real in the run-up and a rounding error in the parabola, and the flow chased the price. The whale alerts, in the one year I could measure, pointed mildly the wrong way. The earnings straddle is priced about right on both sides, and the spread eats the residue. The crush is real but yesterday’s IV rank tells you nothing tradeable about tonight’s. And panic itself — the thing I actually traded through in 2021 without ever seeing its plumbing — turns out to live less in prices than in liquidity: the spread blowing out forty minutes before the low, the quotes going dark mid-squeeze, the toll at the door doubling exactly when everyone needs the door.

None of this made me money. All of it would have saved me some. The 2021 account paid the panic toll nearly four hundred round trips over without knowing the price of anything but the premium; the version of me that writes these articles finally got the itemized bill.