Who took liquidity is the wrong question

00 · The scorecard

Seventeen assertions, one by one

Each row quotes the article in its own words (an ellipsis marks a trim; nothing is paraphrased), then the finding and the reason. The sources follow in the sections below.

01 · The hinge

Role and inventory are different questions

Every argument in the article that touches signed participant data swaps one question for another. "What does the hedging cohort hold" becomes "who crossed the spread". They have different answers, and a positioning product measures only the first.

As written, the assumption is true. The article refutes it by changing what "providing liquidity" means. A market maker is a liquidity provider in the classes it is appointed to because it registered as one and took on the obligations that come with registration: a continuous two-sided market during the day, and dealing for its own account when supply and demand come apart. Firms and broker-dealers carry the other side of customer imbalances in practice, and both desks that publish treat them as hedgers. The role belongs to the participant in the product. It stays in place on a fill where the market maker happened to cross the spread.

The tag also has regulatory consequences. The options market-maker exemption from the SEC's net capital rule, which Cboe reminded its permit holders of in Regulatory Circular RG13-166, is available only to a firm whose securities business is "The (b)(1) exemption is only available to an options market maker (i) whose securities business, except for an occasional non-specialist related securities transaction for its own account, is limited to that of acting as an options market maker on a national securities exchange". A market maker that trades outside that business more than occasionally loses the exemption.

The article's definition is passivity on individual fills, and by that definition market makers fall short about half the time in SPX. Barclays, working from the same participant categories, uses the ordinary meaning:

Under either definition, the thing a positioning product measures stays the same. A market maker who lifts a customer's resting offer ends the trade short exactly the contracts a market maker who was lifted ends short, with the same delta, gamma, vega and the same hedge to follow. The aggressor flag records how the spread was crossed on one fill. The tag records who is in the business of carrying the other side. The inventory records what is held. Figure 1 draws the same trade both ways.

Figure 1. One customer buy of 1,000 calls against a market maker, drawn with each side as the aggressor. The exchange record, the resulting inventory, and the liquidity provider are identical in both panels.

The article then reports that in SPX during April 2025 market makers were aggressive on about half of their fills, and concludes that a positioning model built from participant tags is unreliable. A market maker that crosses the spread on half its fills still holds, at the end of the day, the net of everything it bought and sold. The aggressor share describes execution style. It does not determine the sign or the size of the position.

Crossing the spread is also part of the job. A market maker lifts an offer to leg the spread it was just hit on, hits a bid to flatten a strike it is too long, and crosses the market to hedge the vega of a fill that just arrived. Each of those reads "aggressive" on a trade report, and each is a market maker managing inventory. A liquidity provider that was passive on most fills would be running a different business.

The article's own evidence has a consequence the article does not draw. Its first section says the aggressor flag does not tell you how the hedging cohort is positioned. Its accuracy figure, later on, is a model's ability to reproduce that flag. If the first section is right, the accuracy figure certifies a model on the wrong target. The flag itself is exact: it records which side rested. A signed inventory is calculated without it, because who rested is not the quantity a positioning product measures.

02 · The evidence

Two tables that contradict the text around them

The only data in the article is a pair of April 2025 aggressor-versus-passive tables for SPX and SPY. Read as published, they contradict two of the sentences written about them.

By count the market maker line is near even at 51.76%, and so is the customer line at 48.57%. Firms are 33.21% aggressive by count and professional customers 42.48%. By size nothing is even: firms 75.57% aggressive, broker-dealers 69.25%, market makers 56.43%, professional customers 53.82%, customers 45.49%. The table does not say "equal no matter what the type".

The SPY table shows market makers passive on 92% of their fills by count and 81% by size, with customers aggressive on 95% by count. That is the pattern the article says tags would produce if they meant anything, and it sits in the article's own second table, directly under the sentence saying it does not appear.

Figure 2. Aggressive share of size by capacity, April 2025, re-drawn from the two tables published in the article. The market maker line is highlighted in each panel.

The SPX and SPY panels differ for reasons of order handling, not of role. SPY option volume is dominated by retail marketable orders routed through internalization and price-improvement mechanisms, where the market maker is the responder and is recorded as the passive side by construction, and the one-cent increment leaves little room to rest inside the spread. SPX is an institutionally traded, single-listed product outside the penny program, quoted in nickels below three dollars and dimes above under Rule 5.4, and usually wider than one tick, with orders worked and rested on both sides all day, and a substantial share of size printing through complex orders, auctions and floor crosses. In that market a desk takes a resting customer order when the price suits its inventory, and that fill is recorded as aggressive.

The sample has two further problems. April 2025 contains the tariff shock, so a one-month table blends a stress regime with a recovery and cannot support the separate "in times of stress" claim made later. And on a package execution, an auction response or a floor cross there is no resting side in the ordinary sense; the matching engine assigns the "aggressor" label, and nothing in the order book shows who wanted the trade.

03 · The record

What the exchange knows about each trade

The article treats OPRA as the source of all option data and reasons from what OPRA lacks. Participant-signed positioning comes from the exchange's own trade and clearing records, which contain the fields OPRA omits.

The premise is right and the inference is wrong. OPRA carries price, size, exchange, timestamp and sale condition codes. It carries no participant origin, no buy or sell side and no position effect, so a vendor that needs a side infers one by comparing the print to the quote, which is the midpoint heuristic the article itself rejects. The exchange sells a separate product because its own record contains fields that are not in the OPRA feed. Cboe Rule 6.1(e) lists what a trading permit holder must file for every transaction so the exchange can match and clear it:

Side, capacity, account, and the opening or closing status of every non-market-maker transaction are in the record by rule. The capacity is also disclosed by the exchange itself: Rule 6.2(a) sends it to the counterparty on every trade.

The Open-Close data set is the aggregation of those fields. Cboe's own product description:

Figure 3. Fields on a public OPRA last sale beside fields in the exchange record. The three fields the article says nobody has are all in the exchange record, by rule.

Cboe has published the ten-minute intraday file on its original exchange since January 2011 and the one-minute file since October 2019. VS3D positions are built on the ten-minute file, so the position you see at 10:30 is the position through 10:20.

That describes an open-interest product. A product built from signed participant volume receives the buy or sell, the origin and the position effect as fields, so there is no structural assumption to make about who bought the puts when the record says which origin bought them. The article has grouped four vendors by the shape of their output and then assigned every one of them the weakest input.

04 · The marks

What the rulebook says about "open" and "close"

The article asserts an enforcement asymmetry and a market-maker exemption. The rulebook does contain routes to a wrong mark. They are worth stating in full, because none of them changes a position calculated by netting buys against sells.

Two provisions help the article's side. Rule 6.1(h) lets a trading permit holder that does not know an order's opening or closing status report either value at its discretion, and default an unknown capacity to the broker-dealer code:

Rule 6.6(b) lets the position-effect field be revised through the Clearing Editor on the trade date, and only a change away from the customer capacity code needs a reason code:

Two consequences follow. The position-effect field on the non-market-maker side is a reported field with a defined default, and the broker-dealer category is the one origin that can absorb unknowns, which is a reason to read it with more care than the others. Against that, the accuracy of every reported field is the trading permit holder's responsibility, and a customer's written confirmation must state whether the trade opened or closed a position:

The market-maker exemption is real: Rule 6.1(e)(9) excepts market-maker transactions from the opening or closing designation. It exists because a market-maker account is carried net, so the position in a series is purchases less sales and no designation is needed. The rulebook states that convention in the exercise rule for American-style index options:

The same paragraph is the closest thing in the rulebook to the asymmetry the article asserts: an opening designation may not be re-marked closing except to remedy a good-faith error, with no symmetric clause. It is narrower than the article's version. It governs post-trade re-designation in one class of options, not the standard applied to the original mark, which Rules 6.1(i) and 8.14 place on the trading permit holder in both directions. The penalty exposure that does exist applies to the opening mark: Rule 8.30 position limits and Rule 4.4 closing-only restrictions both test the opening designation, and the minor rule violation plan carries a fine for trading in a restricted class. A permit holder has reason to be careful before marking open.

None of this changes the market-maker position, because the position calculation does not use the designation at all. A market maker's position in a series is its buys less its sells, and that is the quantity we accumulate for each origin every ten minutes. A sale reduces what the seller holds whether the ticket said open or close. We use the designation later, as a check: the customer side is reconciled against listed open interest, and a wrong mark shows up there.

As a blanket claim this is false. The field exists, it is required, it is confirmed to the customer in writing, and it has one stated exception and one stated default. The article could have argued a measured mislabel rate, how often, in what class, and by how much, and it does not.

Mis-marks happen. They do not change a net position: buys less sells is the same number whatever designation each ticket carried. A mis-mark does break any method that sums only the trades marked opening, which is why we check against open interest instead. The worked example is on this site. On September 2 firm accounts bought about 18,000 of the September 9 7825 calls under a closing code, taking the firm net in that series from a small short to 17,765 long by the close, against a market-maker net of 17,914 short, with the difference in the other capacities; another 1,225 followed the next morning. A closing buy of that size requires a pre-existing short of that size, and our record for the series, back to its first trade, shows none. The next morning the clearing house published 18,522 contracts of open interest in a series whose open interest had been a few hundred, 757 above the firm line in our record. Open interest is the total across every participant, so it bounds the position without confirming which column holds the long. It does refute the closing code on the ticket. The direction is the reverse of the one the article describes: here an opening buy was marked closing. Designations are wrong in both directions, and the position calculation does not depend on them.

05 · The institutions

How bank research desks build dealer positioning

The article's closing argument appeals to unnamed institutions. Two desks have published a method for this in the last year. Both name the volatility-surface route among the alternatives, and neither builds on it.

BofA Global Research publishes a Systematic Flows Monitor with an SPX dealer gamma series and a methodology appendix. The appendix lists three ways one could sign an option trade, of which a local volatility surface is the third, and then says where its own estimates come from:

The same desk's year-end volatility review footnotes its dealer gamma exhibit the same way:

Barclays Derivatives Research introduced its SPX option positioning estimate in September 2025. Its method section names the same three candidates, calls them more sophisticated than the naive model, and says which one it builds on:

Neither desk publishes a comparison of the two routes, and neither says the surface approach is worse. What they establish is narrower than the article's claim and enough to answer it: with all three options in front of them, both desks built on exchange participant records, and BofA states the reason in writing. Both also disclose what their series does not do. Barclays' estimate is end-of-day and, in its own words, "the analysis uses end-of-day data and does not capture intraday inventory adjustments - an increasingly important factor given the proliferation of 0DTE options." That is a real limit of that implementation, and the one part of the article's cadence complaint that applies to a bank series. A position refreshed on the ten-minute file does not have that limit.

Our position on this predates the article. In March 2024 we published the three ways a dealer gamma series can be built: naive open interest, aggressor-tagged flow, and participant-tagged exchange volume, and said which one holds up day to day. The same month we wrote that inferring SPX market-maker positioning from trade-level aggressor inference is a fool's errand, for the SPX reasons in section 02. In February 2024 we said in public that at least one bank desk built its dealer view from Cboe-tagged open/close data. Both desks quoted above have since put the method in writing.

VS3D is built the way the fourth row describes: signed participant volume from the exchange, accumulated by strike and expiry, and the market-maker position is the net of that volume, with no inference in between.

06 · Coverage

"Only SPX and VIX"

This packs two claims together. The first, that open/close data exists only for SPX and VIX, is false. Cboe publishes the Open-Close Volume Summary across its four US options exchanges, C1, C2, BZX and EDGX, end of day back to January 2005 on C1 and January 2018 on the other three, with ten-minute and one-minute intraday files on all four. The article's own SPY table breaks April 2025 volume down by capacity: broker-dealer, customer, firm, market maker, professional customer. Only the exchange records those tags, and the article prints them for an ETF option on the page after saying participant-tagged exchange data exists only for SPX and VIX. Research desks use the clearing house's open/close record for single stocks, though for a different purpose:

Those are retail sentiment gauges rather than dealer positioning, and they are cited here to show that the signed record exists and is used well beyond SPX. The second claim, that single-stock dealer positioning cannot be built this way, is one we partly agree with, for a reason the article does not give. SPX and SPXW are Cboe proprietary classes listed on a single exchange, so the exchange's record is the complete listed market in them, FLEX included. A multi-listed stock option is spread across many exchanges, and one exchange's file is a sample of that name's flow, so we publish the one complete position and leave the partial ones alone.

07 · The accuracy claims

Eighty-five percent of what, measured how

The article's method deserves a fair statement in its strongest form, since the article does not give it one. The accuracy figures can then be read for what they measure.

After a fill, the side that was resting behaves differently from the side that crossed. It replenishes or pulls, changes the size it shows, and moves its quote in the direction of the risk it just took. Read across the whole order book at high resolution, that is a real microstructure signal and a better instrument than any midpoint heuristic. Identifying the resting side on a fill is a question with a right answer, and a model can get most of it right. The article leaves three things unaddressed.

The first is the input. The article says its engine ingests every modified quote, every cancelled order and every depth-of-book quote from OPRA. OPRA disseminates last sale and best bid and offer messages per exchange; it carries no orders, no cancels and no depth of book. Those come from each exchange's own feed, and no exchange feed attributes a resting quote to a named firm, so the resting firm has to be inferred from a feed that was itself inferred.

The second is the position effect. Knowing who rested does not tell you whether the fill opened or closed a position, and building an inventory from fills requires that field. A quote-revision model has no such field. It must either assume every dealer fill opens a position or model volume against open interest, which is the method the article's own second section calls a structural assumption.

The third is SPX itself. A substantial share of SPX size trades as complex orders, auction responses and floor crosses, where the matching engine assigns the aggressor and there is no post-fill quote revision to read.

The article gives four figures without a sample, a period, or a definition of a hit.

The answer key measures the wrong thing, and the article says so itself. The 85% is scored against the trade-level aggressor flag, which side was resting when the print happened. That is a well-defined fact and a fair label for a model that sets out to identify the resting side. It is also the field the article's first section says does not tell you how anyone is positioned. Both can be true at once, and if they are, the headline accuracy figure certifies the one thing the article has already said not to draw a position from. The article contains no measured claim about the quantity a positioning product measures, which is the sign and the size of what the hedging cohort holds.

The figure is also ordinary. The published literature on classifying option trades by the quote rule reports accuracy in the high seventies to low eighties of percent on Cboe data (Savickas and Wilson, Journal of Financial and Quantitative Analysis, 2003). Against that benchmark, 85% identification of the resting side is in line with a thirty-year-old heuristic, and "midpoint averages 50%" is below what that literature finds for quote-based rules. The figure may well be right.

The 52% measures a decision rule rather than open/close data. Open/close data outputs no aggressor label, so the only way to score it on one is to add a rule on top, and the obvious rule is "when a market maker is on the trade, treat the market maker as the resting side". In SPX that rule is close to a coin flip, for the reasons in section 02, and it is also close to the market-maker aggressive count share of 51.76% in the article's own April 2025 SPX table. If the 52% is that number under a different name, it describes the aggressor mix, and the article should say which sample and which rule produced it. No one building a signed inventory uses that rule. The inventory adds a market maker's buys and subtracts its sells, without asking who was resting.

"Half of the fills crossed the spread" and "half of the positions are wrong" are different statements, and the second does not follow from the first, for the reason in section 01. Two numbers would test a signed inventory: the mislabel rate on the position-effect field, and any position, on any date, that the inventory got wrong. The article gives neither.

08 · Hedging

Quotes first, futures for the residual, and the desks that publish size the residual

Quote adjustment and futures hedging happen in sequence. A desk that gets hit on a strike first tries to trade out of the risk in options: it reprices, changes the size it shows, and waits for offsetting flow. What it cannot internalize is a residual delta, and the residual is hedged in the underlying. The desks that publish size that residual in futures terms. Barclays reports SPX dealer gamma against S&P futures volume in the last hour of the session, "the S&P gamma positioning has historically shown a long bias, generally ranging between $0-10bn, ~0-20% of last 1hr SPX futures volume (0-4% of daily futures volume)", and is explicit that the window is a working assumption:

BofA, which calls the cohort delta-hedgers and flags "dealer gamma" as the colloquial name, quantifies the effect on S&P e-mini realized volatility by estimating the market impact of trading the required delta in the futures, with hedging confined to the last hour, thirty minutes or fifteen minutes of the session. The published institutional view is internalization first, then a residual that lands in futures and clusters into the close. Both desks size the hedge in the underlying.

The article draws two further conclusions that do not follow. The first is "100% control over prices". That does not describe a class where dozens of registered market makers compete on one order book, where the minimum increment is a nickel or a dime, and where, on the article's own SPX table, customers were the passive side of 54.51% of April 2025 size. Customers set the price on more than half the volume, and a desk that could set the price would not need the hedge. The second is that a filled limit order closes a position. Deemed-closing is position accounting. A market maker quoting a series it does not hold opens risk on the first fill, and everything in this section is about what it does with that risk.

The article is right about gamma and wrong about what follows from it. A long-gamma hedger never reverses direction without spot reversing: it sells the underlying as spot rises and buys as spot falls, on both sides of the strike. The hedge always opposes the move, and its size is proportional to gamma, which is largest at the strike and largest of all on the day the option expires, so at a strike where dealers are long gamma the opposing flow is heaviest, in both directions, and that is the pinning mechanism. Barclays, on its own participant-built series, finds the same thing: "Further analysis confirms a statistically significant negative relationship between dealer gamma and SPX volatility, consistent with the view that high (positive) gamma exposure dampens volatility through mean-reverting hedging flows, whereas low (negative) gamma amplifies moves via momentum-driven dealer hedging activity".

Vanna is a different claim. Vanna hedging requires implied volatility to change; if implied volatility is flat while spot oscillates around a strike, a vanna model produces no flow and no pin. Explaining a pin with vanna also needs a systematic relationship between spot and that series' implied volatility at that strike, which is a stronger empirical claim than the gamma account, and the article does not test it. The article's intuition about "flow without a spot move" does describe charm: into expiry, delta decay moves the required hedge with spot standing still. And into an expiration the expiring line's vega goes to zero, so its vanna goes to zero while its gamma reaches its maximum, which makes the expiring line the one case where vanna cannot explain a pin. The published work on expiration pinning is built on delta-hedge rebalancing of option positions (Avellaneda and Lipkin, 2003; Ni, Pearson and Poteshman, Journal of Financial Economics, 2005; Golez and Jackwerth, Journal of Financial Economics, 2012, for S&P 500 futures).

Figure 4. Futures to trade for a long-gamma position on a $1 move in spot, at each level relative to the strike. The hedge opposes the move on both sides, and it is largest at the strike.

09 · Naive GEX

The assumption the article attacks is the one signed data removes

Agreed on the assumption, with the qualification both desks make. BofA publishes the naive series beside its own and finds it "often appreciably larger in magnitude". Barclays finds that the long-call, short-put assumption "holds directionally" at the index level while it "significantly overstates the magnitude of positioning". Across the whole dealer exposure the naive sign is usually right and the size is wrong. It fails where it is used, line by line at a strike, and the levels a reader trades off are built strike by strike. The article then files VolSignals under it:

VS3D carries the market-maker position by strike and expiry as the net of signed participant volume, with the customer and firm sides shown beside it, and no open-interest parsing anywhere in it. On the /positions page published August 31, the expiring series' sized lines were tested against the naive convention one by one: 20 of the 36 lines over 300 contracts carried the opposite sign to what naive GEX would assign, including calls where customers were the buyers and puts where customers were the sellers. That is a direct test of the assumption the article says vendors rely on, run on the data the article says cannot be built, and published a week before the article appeared. A level printed at a strike is a claim about that strike, and being right on aggregate direction does not defend it.

On the 7790 line at the August 12 close, listed open interest across all expiries was under a thousand contracts and the measured market-maker net was close to flat. Any screen showing a large standing dealer position there was showing contracts that did not exist. That comparison, published August 14, is on our site as well.

Signed open/close data is the exchange's record of which origin bought and which sold, and whether the customer opened or closed. "Highly error-prone" needs a measured error, and the article does not measure one.

10 · What the article does not contain

The test that was never run

The first half is right. The record reports what a trade did to an account, not why it was done. The data set is licensed to anyone who pays for it, which is how bank desks, independent researchers and VolSignals have it. The article gives no source for the exchange's motive, and a motive would not change what the fields contain.

11 · For the record

What VolSignals positions are, and what they are not

Every SPX option print, with the participant on each side, netted by strike and expiry, on the last two years of proprietary data, refreshed every ten minutes through the session.

What this build cannot do

What would change our mind

We would accept and publish any of three results: a measured position-effect mislabel rate on the non-market-maker side of SPX, by class and size band, from a stated sample; any series and date where our market-maker net exceeds the published open interest for that series; or a side-by-side of two market-maker gamma series, one built from participant records and one from a surface, against next-day SPX realized volatility on a stated sample.

Sources and disclosures

The article under review. "Debunking A Widely-Held Belief (And Lie) About Option Flow And Visualization Methodologies", OptionTeller (Substack), September 8, 2026, optionteller.substack.com. All passages attributed to "the article" are quoted verbatim from the published text; the two tables in section 02 are re-typed from the images as published.

Exchange documents. Rules of Cboe Exchange, Inc., current C1 rulebook (Rules 4.4, 5.4, 5.51, 5.52(d), 6.1(e), 6.1(h), 6.1(i), 6.2, 6.6, 6.20(e), 8.14, 8.30, 9.5) and Regulatory Circular RG13-166, cdn.cboe.com. Cboe DataShop, Open-Close Volume Summary product description, datashop.cboe.com. OPRA message content per the OPRA binary data recipient specification published at opraplan.com.

Institutional research. BofA Global Research, Systematic Flows Monitor, September 4, 2026 (methodology appendix, "SPX option gamma positioning"); BofA Global Research, Global Equity Volatility Insights, December 2025 (exhibit source line on dealer gamma); Barclays Derivatives Research, "Reading the Dealer's Hand: Unpacking SPX Option Flows", September 23, 2025; Barclays, "The Long & Short of It", August 18, 2026; UBS US Equity Derivatives Strategy, December 12, 2025. These notes came across our desk as licensed research; short passages are quoted with attribution for the purpose of comparing stated methodology, and no chart, table or extended passage from any of them is reproduced here. Where a note states a caveat about its own method, it is quoted as well.

Literature. Savickas and Wilson, "On Inferring the Direction of Option Trades", Journal of Financial and Quantitative Analysis, 2003. Avellaneda and Lipkin, "A market-induced mechanism for stock pinning", Quantitative Finance, 2003. Ni, Pearson and Poteshman, "Stock price clustering on option expiration dates", Journal of Financial Economics, 2005. Golez and Jackwerth, "Pinning in the S&P 500 futures", Journal of Financial Economics, 2012.

VolSignals pages referenced. volsignals.com/positions (August 31, 2026: naive-convention test on the expiring series), the measured-versus-inferred note on the 7790 line (data at the August 12 close, published August 14, 2026), volsignals.com/gamma (September 1, 2026), volsignals.com/sep9-7825c.

Scope. This page addresses the article's methodological assertions and nothing else. It makes no claim about any individual, and no claim about the accuracy of any competing product beyond what its own published text asserts. Informational only, not investment advice.

Index vol & positioning intelligence · volsignals.com · @VolSignals · Informational only, not investment advice