Part VIII · The whole market · Chapter 8B

Who holds what

The sign problem, and how positions are estimated

1 · Why the sign is an assumption

Gamma is identical, and positive, for a call and a put at the same strike and expiry (Chapter 5C). So an option's type, strike and expiry say nothing about which way its line counts in GEX. Only the side the dealer holds does, and that side is not in the public data. A working paper by Chilingarian (2026) puts it in one sentence:

Open interest is the public count, and it is built to be sideless. The OCC computes it once a day, overnight, from end-of-day records that its clearing members submit and that it then reconciles. It shows the count on its own site the next morning and distributes it to the industry the next business day. Open interest rises by one when a trade opens a new contract and falls by one when a trade closes one. It does not change when a contract passes from one holder to another. Nothing in it says whether a dealer, a customer or neither holds either side.

So every GEX model fills in the sign $s_i$ of (8A.1) with a sign convention. That is a rule, or a model, that decides for each line whether dealers are long or short. The rest of this chapter is about how good those rules are.

2 · The vendor formula and its two conventions

The public formula for GEX came from a vendor. SqueezeMetrics first published its "Gamma Exposure (GEX)" white paper in March 2016 and revised it in December 2017. Its per-line formula is $\Gamma \cdot \text{OI} \cdot 100$ for calls and $\Gamma \cdot \text{OI} \cdot (-100)$ for puts, in shares per $1 of index move. Multiplying by $S^2 \times 0.01$ gives the per-1% form most vendors quote. SpotGamma calls the 0.01 scaling "the industry convention":

Here $\text{sign}_i = +1$ for calls and −1 for puts. It is (8A.2) written out, with $\text{OI}_i = N_i$, $M = 100$, and one sign convention built in. Each term is the dollar gamma per 1% move of the open interest on one line. The sign is the whole game, because it encodes the assumption about which side of each option type the dealer holds. The white paper's own claim was that high GEX goes with low volatility the next day. Chapter 8G sets that claim against later tests, including a vendor's test of its own product that finds most of the signal already in the VIX and at-the-money implied volatility.

The Dealer Mechanics Lab says that every later vendor adopted this formula in some form, and names SpotGamma, Tier1Alpha and Cboe's own derivative dashboards. Of the three, only SpotGamma documents a formula of this form, and it applies its own sign rules. Tier1 Alpha's method is unpublished. Cboe's published research does not assume a sign at all: it measures positions from its own trade records, as the sections below describe.

Two conventions are in wide use, and both are assumptions about who holds what, not observations:

The 0DTE-era row needs customers to be net sellers of puts, as in cash-secured puts and put credit spreads. At the same time it needs them to be net buyers of calls. A customer who sells a call, in a call credit spread or an iron condor, leaves the dealer long that call, which pushes the call sign back toward +1. Which of those flows dominates is exactly what the convention assumes.

The figure computes (8B.1) for a made-up book around a spot of $100. The book holds puts at strikes 88, 92, 95, 98 and 100 and calls at 100, 102, 105, 108 and 112, all with 30 days to expiry at 18% volatility. The upper chart draws total GEX at every spot from 85 to 115, with the spot where the total crosses zero dashed and the current spot solid. The lower chart splits the total at the current spot by strike. The spot shift slider moves spot from −12 to +12 around $100, and two buttons switch between the conventions. Slide spot down to 94 and read the total, then switch the convention.

spot shift from $100

+0.0

sign convention JPM-era 0DTE-era

spot = $100.0

total GEX = +$0.014B

gamma flip = $99.5

It opens at spot $100.0 under the JPM-era convention, with total GEX at +$0.014B and the zero crossing, which the readout calls the gamma flip, at $99.5. Under this convention the profile is negative below the flip, lowest near 92 to 94 at about −$0.07B. Above the flip it is positive, highest near 106 at about +$0.08B. At spot 94 the readout shows −$0.070B: dealers are assumed short the put gamma around them, so their hedging amplifies moves.

Switch to the 0DTE-era convention and the whole profile turns upside down. Spot 94 now reads +$0.070B and spot 106 reads −$0.080B, and the side below the flip becomes the side where hedging damps moves. The flip itself stays at $99.5, because reversing every sign moves no zero. Chapter 8C reads the gamma flip as a level, and notes that it is not the same thing as SpotGamma's Volatility Trigger. The caption gives the 0DTE-era convention's usual rationale, that customers sell premium. The section on the 0DTE era below shows that the rationale does not survive the data.

This is not a minor calibration issue but the dominant source of model risk in every public GEX product. Model risk is the risk of being wrong because a model's assumption is wrong. Receive a claim that "GEX flipped negative today" with skepticism. Ask first which convention the publisher uses, and whether it matches the actual flow in the product you trade.

3 · The JPM era: where the first sign came from

SqueezeMetrics built its method on the assumption that dealers are net long calls and net short puts. The reasoning came from the dominant flow at the time. Institutional asset managers were the main sellers of calls, through covered calls and the famous "JPM collar" rolled by JPMorgan's Hedged Equity Fund. Institutional hedgers such as pension funds and endowments were the main buyers of puts. Dealers stood on the other side of both: long the calls they bought from yield-hunting asset managers, and short the puts they sold to hedgers. The white paper gives the reasoning in its own words. Calls are "sold by investors and bought by market makers (reflecting covered call strategies)". Puts "are mostly purchased by investors for downside protection and sold by market makers."

Under that assumption, a rise in SPX makes the dealers' long calls gain delta, so dealers sell stock to stay flat, which damps the rally. A fall makes the puts' deltas more negative. That leaves the dealers who are short them longer in delta, so dealers sell stock again, which amplifies the fall. The result is the familiar pattern of slow grinds up and fast crashes down. The Dealer Mechanics Lab attributes the asymmetric character of equity volatility cycles directly to this asymmetric dealer positioning.

What position data say about that era

The put half of the assumption has the strongest support of anything in this chapter, because several studies could see who held the puts:

• Garleanu, Pedersen and Poteshman (2009) used CBOE data that identify the aggregate long and short open interest of public customers and of firms' own traders, from before weekly options existed. End users were net long index options, most of all out-of-the-money puts, and net short single-stock options. So dealers were net short index puts.
• Chen, Joslin and Ni (2019) measured public investors' net buying of deep out-of-the-money SPX puts, January 1991 to December 2012. On average the public bought this crash insurance from intermediaries, so dealers were net short it. When the public's net buying fell, which the authors read as intermediaries' constraints tightening, options grew more expensive. Risk premia rose across many assets, funding liquidity deteriorated and broker-dealers deleveraged.
• Constantinides and Lian (2021) model and test the same market: market makers are net sellers and customers net buyers of index puts. As disaster risk rises, customers want more puts while market makers grow more credit-constrained in writing them, so the net amount bought falls just as prices rise.

The last two carry a warning for every convention. The dealers' short put position is not fixed. It is a quantity they choose, and it shrinks exactly when their constraints bind, which is when stress arrives.

The call half has weaker support. Covered-call writing and collars do leave dealers long calls, when dealers take the other side. By Cboe's count, assets in derivative-income funds, which sell options, grew about sixfold from 2019, from about $20 billion to over $120 billion. Yet Cboe's own note on those funds found that buying of upside calls for leverage, not call selling, dominated the flows it examined. The position study in the next section finds dealers' net call position near zero.

The one position whose holder is known

One large index-option position is public by construction. The JPM collar of JPMorgan's Hedged Equity Fund is reset every quarter, with about 35,000 SPX contracts per leg by vendors' counts. The fund holds stocks and, as vendor commentary describes it, "buys a put spread below the market for downside protection and sells calls above the market to finance those puts". The fund is long the upper put of the spread, so whoever sold that put, typically dealers, is short it. At that one strike the short side is short gamma by construction, not by assumption. Its hedging joins a move through the strike instead of opposing it. Chapter 8C returns to what that strike looks like on a map.

4 · The 0DTE era: a story and the data

SqueezeMetrics' assumption dates from 2016 and 2017. After Cboe added Tuesday and Thursday expirations to SPX weekly options in 2022, every weekday had an expiry, and same-day options took over the volume. Cboe's figures put 0DTE options at 5% of SPX volume in 2016, over 40% after the new expirations, about 50% in August 2023, and a record 62% in August 2025. Cboe also estimates that retail makes up about 50% to 60% of SPX 0DTE trading.

The Dealer Mechanics Lab's account of this era runs as follows. Retail and small institutional accounts sell premium, through credit spreads, iron condors and cash-secured puts, with dealers buying from them. So the new net positioning is plausibly the opposite of the JPM-era assumption: dealers net long puts, bought from customers who write them, and net short calls, sold to customers who buy them. That is the 0DTE-era convention.

The flow data do not support its premise that customers overwhelmingly sell premium:

• Cboe's own data describe customer 0DTE activity as "very balanced" in 2023, spread across hedging, yield harvesting, tactical leverage and systematic trading. The 2025 update calls it "extremely balanced", and adds that "over 95% of all 0DTE trades are done in a limited risk format (either long options outright or short via spreads)".
• A tabulation of uncertain attribution, for January 2020 to June 2023, puts customers' 0DTE purchases from market makers at 14.95% of all SPX and SPXW volume, and their sales to market makers at 14.06%.
• Beckmeyer, Branger and Gayda (2023), an SSRN working paper, identified retail SPX trades through Cboe's price-improvement auctions, February 2021 to September 2023. More than 75% of those retail trades were in 0DTE contracts. That retail slice bought more 0DTE options than it sold, leaving market makers net short. The paper also reports the market makers' aggregate net short 0DTE position growing. It is a narrow slice, about 6% of 0DTE trading by the paper's identification, against Cboe's estimate of 50% to 60% retail.

Studies that measure market makers' positions rather than assume them find something more specific:

• Dim, Eraker and Vilkov (2023), "0DTEs: Trading, Gamma Risk and Volatility Propagation", is an SSRN working paper (SSRN 4692190), first posted in November 2023 and circulated in 2024. It rebuilt market makers' inventory from Cboe's records of end users' buying and selling. They found that "Market Makers' inventory, as measured by the net gamma of their positions, is on average positive and negatively related to future intraday volatility". The paper infers positions from flow; it does not attribute them to retail premium selling, which is the Lab's gloss.
• Adams, Fontaine and Ornthanalai (2024), an SSRN working paper, found that "Market makers match most 0DTE order flows, but absorb end-users' trades in long-dated options that eventually become 0DTE positions."
• Adams, Dim, Eraker, Fontaine, Ornthanalai and Vilkov (2025), "Do S&P500 Options Increase Market Volatility? Evidence from 0DTEs" (SSRN 5641974), answer the question in their title with "On the contrary", and continue:

• Amaya, Garcia-Ares, Pearson and Vasquez (2025), a working paper, used proprietary records of every SPX and SPXW trade from January 2020 to June 2023. They found that the gamma of the aggregate market-maker position, "while typically positive, is often negative".
• Cboe's own measurement (Xu, 2023) put market makers' net 0DTE gamma at about $170 million to $670 million through the day. At 3:30 PM its median was +$173 million, with a typical range reaching down to −$1.1 billion.

Put together, the measured sign of dealer gamma in SPX same-day options is positive on average, and not always. It comes mostly from positions in longer-dated options that age into expiry, not from same-day premium selling, which market makers largely pass through. The 0DTE-era convention can land on the right average sign for a put-heavy book, but for a reason the data contradict. Chapter 8G gives the effect sizes these studies report.

5 · What position data show

The most direct evidence on the sign for all SPX options, 2011 to 2023, is a working paper by Terstegge (2025). It builds intermediaries' aggregate positions from Cboe's daily open and close volume files:

• Calls: intermediaries' buy and sell volumes are roughly balanced, and their net call position hovers near zero.
• Puts: sell volumes consistently exceed buy volumes, leaving a persistent net short put position of about 19 million contracts, concentrated in deep out-of-the-money puts.
• Risk: without hedge adjustments, a 10% fall in the S&P 500 would cost roughly 200% of the dealers' option inventory value. Even a dealer perfectly hedged at the close cannot re-hedge a large overnight or weekend move, because equity liquidity is thin then.

By contracts, then, dealers look like the put half of the JPM-era convention with flat calls, and nothing like the 0DTE-era convention's "dealers long puts". Two cautions apply. Cboe's files miss hedges done off the exchange or in futures options. And contracts are not gamma: a deep out-of-the-money put carries little gamma until a crash brings spot toward it.

The second caution is how "dealers short puts" and "dealers long gamma on average" can both be true. The worked example shows one way, with a made-up book.

That pattern, damping on normal days and amplifying in a sell-off, is what Chapters 8D and 8H describe. It also means that a single sign per option type is wrong for SPX. The JPM-era convention gets the puts' direction right by count but misses the long gamma near the money. The 0DTE-era convention can match the average gamma sign but signs the puts the wrong way, and it would miss the turn in Step 4.

The figure below draws the worked example's book against spot. One slider moves spot from 5,000 to 6,200, starting at 6,000. Two set the book: the block at 6,000 with 7 days left, from short 200,000 to long 200,000 contracts and starting long 100,000, and the short puts at 4,800, from none to 5 million and starting at 1 million. The curves give each block's gamma exposure and their total, in billions of dollars per 1% move, and a ring marks spot.

spot

6,000

at 6,000, 7 days

long 100,000

puts at 4,800

short 1,000,000

by count: net short 900,000 contracts

GEX at 6,000: +$7.76B

sign changes at about 5,838

Long gamma: damping. At 6,000 the near block gives +$11.51B and the far puts −$3.75B. Short by count, the dealers are long by gamma, because the gamma sits near the money.

It opens on the example: net short 900,000 contracts, yet +$7.76 billion of GEX at 6,000, with the sign changing at about 5,838. Drag spot to 5,800 and the readout gives Step 4's −$2.22 billion. At 5,400 it is −$9.51 billion, all of it from the far puts. Return spot to 6,000 and raise the short puts to 5 million: the total is −$7.26 billion even at 6,000. The count of contracts never changes as spot moves. Only the gamma does, and the gamma decides the hedge.

Single stocks follow different rules

Positions in single-stock options look different again. Lakonishok and co-authors (2004, published in 2007) used CBOE open interest split by investor type, 1990 through 2001. Investors other than market makers held about four times more long call than long put open interest, and more short than long open interest in both calls and puts. Market makers were therefore net long both. Christoffersen and co-authors (2018) find illiquidity premia in equity options consistent with evidence that market makers there "hold large and risky net long positions", which the premia compensate. Kanne, Korn and Uhrig-Homburg (2023) find market makers net long options on some underlyings and net short on others.

Vendors disagree with those data and with each other. SpotGamma states that its "working assumption" for equities "is that dealers are short options for both puts and calls". It also models market makers "to be short puts and long calls on index products, due to the prevalence of option strategies like collars and covered calls". Barchart's free pages sign calls positive and puts negative for every ticker. Episodes of heavy retail call buying can leave dealers short gamma in a single name, the setting of Chapter 8I. Lipson, Tomio and Zhang (2023), a working paper, find that retail option trading after commissions fell to zero raised the volatility of the underlying stocks. The channel was dealers hedging higher-gamma contracts in the direction of price moves.

So no single convention fits every underlying. The Dealer Mechanics Lab says that for SPY and individual stocks the appropriate convention may be regime-dependent, and the single-stock evidence supports that.

6 · Every guess at once: the sign map

Neither convention is observed, so the honest check is to try every positioning in between. The figure colors each one by the GEX it gives, for the made-up book of the profile builder above. Across is the share of call open interest that dealers are short, and up the share of put open interest they are long. Dealers hold the rest the other way. The JPM-era convention is the bottom-left corner and the 0DTE-era convention the top-right one. One slider moves spot from 88 to 112, and another scales all the put open interest from a quarter to four times its size. Move spot and watch the zero line turn, while the two conventions stay on opposite sides of it.

spot

$100.0

put OI scale

×1.00

JPM-era corner: +$0.014B

0DTE-era corner: −$0.014B

GEX > 0 on 50% of the square

puts carry 47% of the gamma

The whole square: from −$0.224B, dealers short every option, to +$0.224B, dealers long every option.

At spot $100.0 the corners read +$0.014B (JPM-era) and −$0.014B (0DTE-era), the profile builder's totals at 100, and the puts carry 47% of the book's gamma. The whole square runs from −$0.224B, with dealers short every option, to +$0.224B, with dealers long every option. So both conventions sit close to the zero line: a modest error in the assumed shares flips the sign.

The readout says that GEX is positive on 50% of the square, and it says so at every spot and every put scale. That is a property of the model, not of the book: swapping every share for its opposite reverses GEX, so the zero line always passes through the middle. What moves is the line's tilt. At spot 94 the puts carry 73% of the gamma, the line is shallow, and the sign depends mostly on who holds the puts. At 106 they carry 24%, the line is steep, and the calls decide. Scale the puts by four at spot 100 and they carry 78%. The corners then read −$0.301B and +$0.301B.

The position data put SPX near the bottom edge, midway across: calls flat and puts short. On this book that point is negative at every spot. But the square gives every strike the same shares, and real books do not, as the worked example above shows. The square answers "how sensitive is the sign?", not "what is the sign?".

7 · Demand-based option pricing

A foundational paper sits under all of this: Garleanu, Pedersen and Poteshman (2009), "Demand-Based Option Pricing", Review of Financial Studies, Vol. 22, No. 10, pages 4259 to 4299. In their model, dealers who cannot hedge perfectly charge for the risk left in the positions they absorb. Demand-based option pricing follows: demand pressure raises an option's price in proportion to the variance of its unhedgeable part, and raises other options' prices in proportion to covariances. The Dealer Mechanics Lab summarizes the paper as showing that net end-user demand for index options is the primary driver of the deviation of option prices from no-arbitrage benchmarks. The deeper the dealers' positioning imbalance, the larger the deviation, and the larger the hedging activity needed to manage it.

The Garleanu-Pedersen-Poteshman framework establishes that option prices are not pure forecasts of future variance plus a risk-free rate. They also depend on who is buying and who is selling, beyond the model fundamentals. The Dealer Mechanics Lab takes this as the structural justification for the variance risk premium (Chapter 10C) and as the academic foundation for the GEX framework as a whole.

The Edge Book reads the same paper more narrowly (Chapter 10I). In its reading, demand pressure affects option prices at the level of option pricing, not the price of the underlying, and its effect per unit of demand is small. Nor does the paper validate specific level-trading rules. Both readings are defensible, because the paper tests prices, not what dealer gamma does to the index.

Later work fills in the mechanism. Bollen and Whaley (2004) found that net buying pressure for index puts moves the implied volatility of S&P 500 options most strongly. For stock options, call demand dominates. This is the causation Chapter 8A mentioned: heavy demand for index puts steepens the skew. Muravyev (2016) split the price impact of option trades into an inventory-risk part and an information part, and found the inventory part larger. Fournier and Jacobs (2020) model an index-option market maker who demands a larger variance risk premium after it incurs losses. The inventory GEX tries to measure is the same inventory that sets option prices, and dealers are paid to carry it.

8 · How positions are estimated

Between two of OCC's daily counts, every intraday GEX number is imputed from the trade tape rather than observed. Vendors do this with varying sophistication, and the imputation errors can be substantial intraday, especially in high-volume sessions. Four sources feed the estimates:

Cboe's end-of-day Open-Close subscription is listed at $500 a month, the 10-minute intraday product at $1,000 and a one-minute product at $5,000, with history from January 2018. So the side of an SPX trade is observable within minutes for a few thousand dollars a month. The side of an SPY trade is split across exchanges and, for retail orders that wholesalers internalize, partly off the exchanges altogether (Chapter 2B). That is why the one vendor method that claims to make no assumptions covers SPX only, and why most retail maps assume a sign.

Volume is not position

Trading volume says little about who is left holding what. Cboe's 2023 note gives an example from August 15 of that year. Over 100,000 contracts of a 4440-strike put traded in the day. Yet market makers "were left short only about 3,000 contracts by the end of the day, or just 3% of the gross volume". Its 2025 update states the principle: "what matters for determining the potential impact of market maker gamma hedging activity is the balance of the volume between buys vs. sells, not the notional size".

Trade classification and its errors

Where the data do not record sides, vendors infer them. Three rules are common. The quote rule calls a trade above the mid-quote a buy and one below it a sell, and cannot classify one at the midpoint. The tick rule compares the price with the previous trade's. The Lee-Ready algorithm combines the two. This is trade classification, and its accuracy has been measured. Savickas and Wilson (2003) tested the rules against a proprietary CBOE data set that records the true direction of each option trade. The quote rule, Lee-Ready, the Ellis-Michaely-O'Hara rule and the tick rule signed correctly 83%, 80%, 77% and 59% of the trades each could classify. That is an error rate of 17% to 41%.

The errors are not spread evenly. Trades outside the quotes and trades against reversed quotes were highly misclassified by all four rules. Components of index-option complex trades, executed away from the retail automatic system, were misclassified almost 50% of the time by any method. Dropping them, 15% of the sample, raised the quote rule's success to over 87%. A second problem is who the "customer" is. A study led by Muravyev (2020) found that most option trades originate from investors who time their executions, such as proprietary traders and institutions with execution algorithms. So the liquidity taker is often not a customer facing a hedging dealer. And the midpoint can be a poor guide to value at the moment of a trade.

The figure shows what wrong sides do to a map. A made-up day has 240 same-day trades at nine strikes around 6,000, each with a true side. The outlines are the dealers' true net position at each strike, positive where they are long. The filled bars are the position estimated when each trade's side is misclassified with probability e, which the slider sets from 0% to 45%, starting at 20%. Rings mark the largest node in each, and "new draw" misclassifies the same trades afresh. Set e inside the published range and press new draw a few times.

misclassified, e

20%

new draw

signs flipped: 2 of 9 strikes

largest node: 6,000, unchanged

total net error: −820 contracts

This draw: 43 of 240 trades misclassified, and a total of −665 contracts against a true +155. On average the errors leave each strike at 1 − 2e = 0.60 of its true net.

At 20%, the first draw misclassifies 43 of the 240 trades. Two of the nine strikes, 5,990 and 6,010, flip sign, and the largest node stays at 6,000 but shrinks from a true +1,058 contracts to an estimated +504. The total is the most fragile number: an estimated −665 contracts against a true +155, an error of −820. On average the errors leave each strike at $1 - 2e$ of its true net, 0.66 at 17%, 0.60 at 20% and 0.18 at 41%. Raise e toward 41% and press new draw: the largest node starts to wander from one strike to another.

Two findings bound how well any of this can work. Ardia and Vaudescal (2026), an SSRN working paper, built net gamma from public open interest over five years of half-hour SPX data. It tracked a dealer-inventory benchmark at a correlation of 0.99 for options other than 0DTE, but only 0.53 for 0DTE. The benchmark shares its open-interest starting point with the proxy, so 0.99 is an upper bound on agreement. And market makers match most 0DTE flow, as Adams, Fontaine and Ornthanalai found, so their same-day gamma comes mostly from older positions that sit in yesterday's open interest. For longer expiries, yesterday's open interest is most of the story. For same-day options, the flow intraday maps chase is mostly offset, and the part that is not is the part public data measure worst.

The figure below makes Cboe's point with a made-up session in one same-day put: 2,000 trades, 100,000 contracts in all. Customers buy a share of the volume you set, starting at 51.5%, and market makers take the other side of every trade. A flow-based map signs each trade with a rule that errs with probability e, starting at 13%, about the quote rule's error once complex trades are set aside. It signs the legs of spreads, 15% of trades to start, by a coin toss, as with the complex-order legs misclassified almost half the time. New day draws another session.

customers buying

51.5%

misclassified, e

13%

spread legs

15%

new day

held: short 3,024 (3.0% of volume)

inferred: short 1,742

volume signed wrongly: 17%

Close. The map is off by 1,282 contracts, against a true position of 3,024. Its errors scale with the 100,000 contracts traded, the true position with the balance between buys and sells.

At the start, market makers end short 3,024 contracts, 3.0% of the volume, close to Cboe's example. The map signs 17% of the volume wrongly and infers short 1,742: the right sign, 1,282 contracts off. Set the buying share to 50% and the true position shrinks to long 112, while the map infers long 1,366. Its errors stay as large, because they scale with the 100,000 contracts traded. Press new day a few times at the start's settings: on some days the map gets the sign wrong. Balance decides the hedge, and balance is what a classified tape measures worst.

9 · Vendors' methods side by side

Every vendor answers the sign question somehow, and the answers range from a fixed rule to measured positions. As their own documentation describes them:

A few statements in that material deserve their labels. The "no assumptions" of the Unusual Whales method holds only as far as Cboe's own flags are complete: a dealer trading as a broker-dealer, or hedging in SPY or futures options, falls outside them. Volland claims "over 90% accuracy over all expirations, with 99% accuracy in 0DTE options" for its classification, without a disclosed benchmark. Its creator, a competitor of the others, says the open-interest convention is "right maybe ±75% of the time", a figure with no support offered. Skylit says only that its positioning is "derived by Skylit from that data, not reported by the exchanges". Its calculations run on "custom inference models and proprietary intelligence built from years of research into dealer microstructure" (Chapter 8C). A working paper by Sen (2026) sums up the market: commercial services "cost roughly US$99 to US$1,999 per month, cover mainly U.S. equity indices, and disclose little of their methodology".

SpotGamma, to its credit, published a sensitivity test of its own model. On data from February 7, 2020, it assumed the model "is only 50% correct and the remaining 50% open interest is randomized to be long or short", and repeated the run at higher shares. The total moved with the share it got right, which is the sign map's lesson in a vendor's hands.

10 · What no public data show

Every public GEX model sees mainly the standard listed chains: contracts that trade on Cboe and the other exchanges, cleared through the OCC, with strikes on a fixed grid. That dataset is comprehensive, but it captures only one side of the market, the listed side. The Dealer Mechanics Lab calls the rest the invisible book, and of all the limits it names this as the one that bites hardest in practice. "Absent from the chains most vendors use" is the more accurate phrase, because some of it does reach OCC.

FLEX options sit alongside the listed market, trading on Cboe but with a customizable strike, expiration, exercise style and settlement (Chapters 2B and 4A). OCC clears them and publishes FLEX open interest reports, but they sit outside the standard chains vendors ingest. The Dealer Mechanics Lab cites Cboe's figures: FLEX volume averaging 1.3 to 1.4 million contracts a day in 2025, more than ten times the 2019 level. Total FLEX open interest is 35 million contracts, up from 2 million in 2019, a 17.5× increase. Buffer ETFs, the defined-outcome ETF products built from FLEX, held more than $70 billion by mid-2025. Issuer literature says that their market makers "hedge through the broader listed options complex". So part of the FLEX book's gamma reappears in listed open interest, with a sign no convention assigns correctly.

Over-the-counter (OTC) books are harder still. OCC has cleared some over-the-counter S&P 500 index options since 2014, but bilateral books leave no public record. Foreign exchange shows what records make possible. Anderegg, Ulmann and Sornette (2022) rebuilt aggregate dealer gamma in EURUSD and USDJPY options from trade-repository data, October 2017 to June 2018, and found it negative. No equivalent exists for U.S. equity-index options traded off the exchange.

Structured products are the largest unseen layer: autocallables, principal-protected notes and range accruals, which retail and private-wealth channels sell to clients around the world. An autocallable pays a coupon and can be redeemed early when the index is above a set level. It exposes its holder to losses if the index falls through a lower barrier. The issuing bank's hedge replicates the structure's option exposure dynamically. The Dealer Mechanics Lab says the autocallable book alone is measured globally in hundreds of billions of notional, dominated by retail flow in Europe and Asia. The evidence on these hedges comes from other markets. In Korea, issuers of step-down autocallables are long vega, and their hedging went with lower implied volatility (Yoon 2014; Ji and co-authors 2015). In January 2024 China's CSI 500 and CSI 1000 indexes fell through the knock-in levels of "snowball" autocallables, set at 70% to 80% of initial levels, and a case study (Yin, 2026) lists institutional hedging pressure among the effects. For the S&P 500, a BIS box by Todorov and Vilkov (2024) proposed that option dealers "effectively dampen volatility when they hedge structured products", without measuring it.

The Lab's central claim is that the invisible book matters most precisely when GEX matters most. On a quiet day, FLEX and over-the-counter hedging is small and self-cancelling. On a stress day it can dominate everything the listed GEX number shows. That happens when a barrier is approached, when an autocallable book nears a knock-in level, or when an institutional buyer suddenly rolls a multi-billion-dollar collar. The claim is plausible, and nothing public quantifies it. The Lab adds that Volmageddon's VIX call wing was in FLEX, and that the March 2020 vanna cascade ran through structured-product hedges as much as through listed puts. Those are its interpretations. The documented accounts of February 2018 attribute the amplification to the rebalancing of leveraged and inverse VIX products (Chapter 8H), and no published estimate tests the March 2020 claim. The Lab expects the next stress event to feature a flow that surprises every published GEX model, and calls that almost certain.

Theory supports the worry, if not its size. Gennotte and Leland (1990) built a model in which hedging demand is not fully observed, and found that crashes "can occur even with relatively little hedging". The usual reading is that unseen hedging sales are mistaken for informed selling. Grossman (1988) argued that a traded put reveals the hedging demand behind it and a replicated one does not. Read together, they give listed open interest its value, since it is visible, and they make the unseen hedges the dangerous part.

11 · Units

Four units are in use, and vendors often do not say which. For one line with gamma $\Gamma$, open interest OI and spot $S$:

Skylit and GEXBot state no unit for their values, so their numbers cannot be converted at all.

An SPY or XSP contract carries about one tenth of an SPX contract's dollar gamma at the same moneyness, because its underlying is about one tenth the size (Chapter 6D).

12 · Six limits of the framework

The model has serious limits, and you should know them before you read any GEX chart with conviction. The first four are the subject of the sections above.

1. The counterparty assumption: as the profile builder's toggle made plain, the sign is fragile. The standard SqueezeMetrics convention assumes dealers are long calls and short puts, and the position data fit neither that nor its opposite for SPX. The right convention depends on which counterparty actually holds each option in aggregate, which is hard to observe.
2. Open interest publishes with a lag: the OCC publishes open interest T+1, that is, on the next business day. Intraday GEX is imputed by inference from the trade tape, with the classification errors described above.
3. FLEX, over-the-counter and structured-product books are outside the chains: their hedges are real, and public models cannot size them.
4. Black-Scholes gamma is wrong under skew: GEX uses Black-Scholes gamma at one volatility, as if the volatility surface were flat. The Dealer Mechanics Lab says that real put-wing gammas under steep skew can be several times the Black-Scholes number. It adds that the error does not reverse: it consistently understates put-side gamma in equity markets. Measured convexity departs from it in another way. Willeboordse (2026), across five underlyings from 2020 to 2025, found the correlation between empirically measured and Black-Scholes gamma significantly negative at a daily horizon, −0.25 to −0.47. The author attributes it to implied volatility moving against spot. Hull and White (2017) showed that the delta minimizing a hedge's variance folds in the expected change in volatility. A dealer using it trades differently from the Black-Scholes hedge GEX assumes.
5. GEX is positioning, not outcome: it tells you what dealers would do given a move, not whether a move happens. The effect sizes in the academic literature are real but modest. Ni, Pearson and Poteshman (2005) find that on expiration dates the returns of optionable single stocks are altered by an average of at least 16.5 basis points, 1996 to 2002. The paper does not study the index. At the index level, Golez and Jackwerth (2012) find S&P 500 futures pulled toward the at-the-money strike when options on the futures expire. Just before SPX options expire, the futures are pushed away from it. The notional shift is at least $115 million per expiration day. Barbon and Buraschi (2021), first posted in 2020, measure dealers' gamma imbalance. In one version of the paper, one standard deviation of it goes with a fall of about 12 basis points in a stock's absolute daily return, against an average of about 109. Another version gives 5 to 25 basis points for stocks, against an average of 180. For the indexes tested, one standard deviation goes with a fall of more than 20 basis points in the absolute daily return. Chapters 8E and 8G give these results in full.
6. Dealers may not hedge the way the formula assumes: Hu, Kirilova, Muravyev and Ryu (2023), an SSRN working paper, studied account-level data for Korean index options. Market makers reduced risk mainly by rebalancing their inventory within minutes, trading the options back out, rather than by delta hedging. Only a minority consistently delta hedged. In S&P 500 options, daily trading volume exceeds net position changes by a factor of 32. The authors conclude that "delta hedging is less universal than commonly assumed". Dong (2025) found that an option trade moves the underlying stock by 0.56 basis points over five minutes, about one sixth of a stock trade's effect. Multi-leg trades had negligible effects. Open-interest gamma can therefore overstate the hedge flow of (8A.3).

13 · An auditable recipe

Two working papers show what a disclosed method looks like. Chilingarian (2026) sets out "a three-model taxonomy" and ratifies one model, the SqueezeMetrics open-interest proxy, "as a written contract". It fixes the aggregation formula and "its unit scale", gives "a correctness rule" for the gamma flip, and shows on "a real 10,223-contract SPX snapshot" how switching conventions changes GEX. Sen (2026) publishes a complete engine for S&P 500 and gold futures options. It prices with Black-76 at each strike's own implied volatility and uses the standard dealer-sign convention. It solves for the zero-gamma level numerically and adds walls, pinning strikes, and vanna and charm overlays, and it runs for US$42.80 a month.

Neither paper makes the sign right. What they make is the sign visible, so that a reader can see the assumption, change it and watch the answer move. That is the standard to hold any GEX number to.

The two traditions this book draws on agree on the method and differ on the weight. The Dealer Mechanics Lab reads the 0DTE-era evidence as putting SPX dealers net long gamma on most days. It makes the sign of GEX a central input to the regime (Chapter 8D). The Edge Book gives three reasons for caution: vendor models disagree, the 0DTE era changed who holds what, and a Cboe study of 2025 put 0DTE market-maker hedging at about 0.2% of SPX daily liquidity. It keeps only the sign, as low-confidence regime context worth 10% to 15% of one filter's weight, and never sizes from the magnitude (Chapter 10H). This chapter's evidence cuts both ways. The measured average sign for SPX same-day options is positive, as the Lab says. Yet the conventions behind most public numbers do not describe the positions, which supports the Edge Book's low confidence in any vendor's reading.

14 · Check yourself