Part VIII · The whole market · Chapter 8J

From map to trade

A practical workflow, and how to test a level yourself

Everything in this part points to a modest conclusion. A GEX map describes where dealers' hedging is likely to be heavy or light, under an assumption about who holds what. The sign of its total carries some information about the day's volatility, most of which the VIX and implied volatility already hold. No study found in this book's research shows that trading at the map's levels pays. A workflow should be built on that conclusion, not on the stronger claims vendors make.

1 · What the evidence leaves you

The chapters before this one sort a map's contents into three kinds, each deserving a different weight:

The rest of the chapter works through that table. It gives a routine for reading the first two rows, a rule for turning the first row into size, and a method for testing the third.

2 · A workflow for each session

Seven steps, most before the open, keep the map in its place:

1. Check the data. Name the sign convention, the units, the expiries included and the time of the data. Open interest is one count a day, published the morning after, so anything drawn during the session is Greek drift plus an estimate of flow (Chapter 8C). If the vendor states none of this, read only where the large cells are, not their colors.
2. Find spot against the flip, and treat the flip as a band. It moves with spot, implied volatility and time; vendors compute it differently; and SpotGamma's Volatility Trigger is a separate level (Chapter 8C). Inside the band the sign of the regime is uncertain.
3. Note the large nodes, their signs, and the gaps. They describe where moves are more likely to be damped or fed, if the sign is right. They are not destinations.
4. Check vanna and the calendar. Events that move implied volatility, such as CPI releases and FOMC meetings, move every cell and the flip. So do expiries, the quarterly roll of the JPM collar, and the last hour of a same-day expiry (Chapters 8C and 8F).
5. Let the regime set a size multiplier, not a trade. The next section gives the weights.
6. Form the thesis elsewhere. The trade comes from the book's other tools (Parts X to XIII). The map can trim it or veto it, but not start it. Skylit's own "Charts First" rule says the same.
7. Log what you saw. Save the map as it stood when you acted: its time, its convention and the nodes you relied on. That log is the raw material for the test in the second half of this chapter.

3 · Regime as a size multiplier, not a trigger

Regime as a decision boundary

For practical trading the Dealer Mechanics Lab treats the regime as a decision boundary on strategy selection, and the shape of the VIX term structure joins the GEX sign in drawing it. The term structure is in contango when longer-dated implied volatility sits above shorter-dated, the usual calm shape. It is in backwardation when that order inverts under stress (Chapter 5E).

• Positive GEX, VIX in contango, no event imminent: premium-selling and mean-reversion strategies are favored. Realized volatility is likely to undershoot implied, range bets pay, and tight stops work.
• Negative GEX, VIX in backwardation, an event imminent or just past: directional and long-volatility strategies are favored. Realized volatility is likely to exceed implied, trend bets pay, and wider stops are needed.
• Transition regimes, with GEX near zero and the VIX term structure in transition: reduce size, take only the highest-conviction setups, and expect choppy intraday action.

Chapter 11G reads the regime more broadly, from eight dimensions of market state, though neither the GEX sign nor the term structure is among its inputs. The Lab's principle is that strategy selection should be conditioned on regime, and regime is most concisely characterized by the sign of net dealer gamma combined with the shape of the VIX term structure.

How the Edge Book weighs the regime

The Edge Book agrees on the direction of the effect: negative-GEX days tend to have higher realized volatility, and positive-GEX days compressed realized volatility (Chapter 10H). It gives the sign far less weight, treating GEX as a low-confidence filter: a soft input worth 10% to 15% of one filter's weight. If GEX disagrees with the VIX term structure or with the realized-volatility filters, the Edge Book ignores it. If it agrees, it adds slight confidence.

In the Edge Book's use the sign adjusts the size of other strategies, for example by trimming put-writing modestly when GEX is negative. Chapter 13F's filter card puts the trim at 10% to 15%. It never sizes from the magnitude of GEX, which it calls too model-dependent. It also rejects trading at specific GEX levels for retail accounts (Chapters 1G and 10I).

Its reasons are the model risk of Chapter 8B, the change in structure as 0DTE options grew after 2022, and a Cboe study from 2025. That study, Xu's "0DTEs Decoded" (May 2, 2025), put 0DTE market-maker hedging "at best, just 0.2% of the SPX daily liquidity". It measures 0DTE hedging only, so applying it to levels built from longer-dated open interest is an inference (Chapter 8C). The two sources agree that the sign is regime context. They differ on how much it should move a decision.

What the evidence says about the weight

One test bears directly on the weight. FlashAlpha, a vendor of dealer-exposure data, tested its own product on 1,972 SPY days from 2018 to 2026, in a study it describes as pre-registered. The rank correlation of GEX with next-day realized volatility was −0.36 raw. It fell to −0.14 after controlling for the VIX, and to −0.03 (p = 0.18) after controlling for the VIX and at-the-money implied volatility. The gap in next-day realized volatility between the top and bottom groups of GEX days shrank from 10.63 volatility points to 3.15, and then to 0.99 (p = 0.25). The vendor's own conclusion: "GEX should be used as a regime label, not a standalone forecast".

That is one vendor's test of its own product, on SPY, at a daily horizon. At shorter horizons the picture changes. Ardia and Vaudescal (2026), a working paper on five years of half-hour SPX data, find that a one-standard-deviation increase in net gamma from longer-dated options predicts next-half-hour realized variance about 36% lower. The link survives their forecasting controls. One master's thesis finds that changes in GEX help forecast S&P 500 returns, without an implied-volatility control (Jonsson and Nyberg, 2025).

Read together, the evidence supports a small weight at the daily horizon, which is the Edge Book's position. It also suggests that pairing GEX with the volatility market, as the Lab does, adds less than it seems: much of what GEX says about tomorrow the volatility market already says.

Chapter 8G's figure rebuilds this test on made-up days. Adjust for the VIX, then for at-the-money implied volatility, and the association falls from −0.36 to about −0.03.

4 · How to test a level yourself

Vendors publish hit rates. SpotGamma's own statistics for its SPX levels cover May 10, 2019 to May 28, 2024. They report that the put wall "held" in 89% of daily sessions, meaning the intraday low did not fall below it, and that SPX closed above it in 93%. After a put-wall breach, average forward returns were +14 basis points over one day, +7 over five days and +39 over ten. After a call wall breach they were −7 basis points over one day and +5 over five. Skylit says fresh nodes react "~80%" of the time. None of these numbers can be read as it stands.

A hit rate needs a base rate

A base rate is how often the outcome happens with nothing special at the level: the same hold, measured at levels that are not walls. A put wall that sits well below spot is "held" on most days whatever its gamma, simply because most days do not fall that far. SpotGamma's report gives no base rate, no comparison with arbitrary strikes at the same distance, and no standard errors.

How much does distance alone give? For a price that wanders at random with no drift, the chance that the day's low stays above a level $d$ daily standard deviations below the open is $2N(d) - 1$. The chance that the close ends above it is $N(d)$, where $N$ is the standard normal distribution function.

A level 1.5 to 1.6 daily standard deviations below the open "holds" about as often as SpotGamma reports for its put wall, whether or not dealers hold anything there. With a VIX of 16 the rule of 16 puts a day's standard deviation near 1% (Chapter 5E), so 1.5 standard deviations is about 90 points at SPX 6,000. The report does not say how far below spot its walls sat, so its 89% cannot be read either way. Real returns have drift and fat tails, so this table, built on a random walk, is a teaching baseline. The honest base rate comes from the data, as the controls below describe.

Where 2N(d) − 1 comes from: the reflection principle

Step 1: the question. A price starts at 0 and wanders at random with no drift over one day, ending with standard deviation 1 in these units. What is the chance that its path touches $-d$ at some point during the day?

Step 2: the mirror. Take any path that touches $-d$ and ends above it. Reflect the part after the first touch in the line $-d$. The reflected path is just as likely, because up and down steps are equally likely, and it ends below $-d$. Every path that touches and ends above has exactly one partner that ends below.

Step 3: count. Every path that ends below $-d$ must have touched it. So the paths that touch are those ending below, plus their mirror partners ending above: twice the paths ending below. The chance of a touch is $2\Pr(\text{end} < -d) = 2N(-d)$.

Step 4: the hold. The low stays above $-d$ with probability $1 - 2N(-d) = 2N(d) - 1$, since $N(-d) = 1 - N(d)$. At $d = 1.5$: $2 \times 0.9332 - 1 = 0.866$, the 87% of the table.

The forward returns need the same care. A +14 basis point average after a breach means little without its standard error. The report gives no count; suppose the breaches are the 11% of about 1,270 trading days in the sample, about 140, and a day's standard deviation is 1%. The standard error of a one-day average is then about $100/\sqrt{140} = 8.5$ basis points, and +14 is about 1.7 standard errors from zero: not distinguishable from nothing at the usual 5% level.

How often a level holds by chance

The figure below measures holds in made-up sessions where the answer is known. Levels sit every 25 points. A touch is price reaching a level from at least 3 points away. The touch holds if price then moves the bounce target back from the touch before it moves the break point through it.

Two sliders set the bounce target and the break point, each from 2 to 20 points; they start at 5 and 20. A switch chooses the market: a pure random walk, where nothing happens at the levels, or one with dealers long gamma at every level. In the second, hedging damps moves within 6 points of a level and adds a small pull toward it, Chapter 8E's pinning in miniature. New sessions draws another 200. The upper panel shows one session with its holds (circles) and breaks (crosses). The lower panel plots the hold rate over 200 sessions against the ratio of break point to bounce target, with the curve a random walk gives.

Read the first state. Then set the bounce target to 2 with the break point at 20. Set both to 10. Then switch on dealers long gamma and try 5 and 20, 10 and 10, and 20 and 5, watching both the hold rate and the minutes near a level.

bounce target

5 pts

break point

20 pts

market random walk dealers long gamma at the levels

new sessions

holds: 80% of 1,069 touches

chance alone: 80%

minutes near a level: 24% (walk 24%)

In a random walk: the hold rate differs from chance by 0 points, and the time spent near a level by 0 points.

In the random walk the default reads "holds: 80% of 1,069 touches" against "chance alone: 80%": nothing happens at the levels, and four touches in five hold. The chance of rising $X$ points before falling $Y$ is $Y/(X+Y)$, here $20/25$. A bounce target of 2 with a break point of 20 gives 89% of 1,183 touches against 91% by the formula, the same 89% SpotGamma reports for its put wall. Equal distances of 10 give 49% of 1,063 touches.

Put dealers long gamma at the levels and the answer depends on the definition. At 5 and 20 the hold rate rises to 87% of 444 touches, 7 points above chance. At 10 and 10 it falls to 46%, 4 points below, and at 20 and 5 to 12%, 8 below. Touches fall to 444 because price that sticks near a level rarely backs 3 points away to approach it again. What changes for certain is the time near a level: 69% of minutes within 3 points of one, against 24% in the random walk. Long-gamma hedging makes price stick near a level, not bounce off it. A test of "bounces" can miss a real effect or invent one, depending on distances you choose.

The $Y/(X+Y)$ formula follows from one idea. A fair game's expected position never changes, so starting at 0, the chance $p$ of reaching $+X$ first satisfies $pX - (1-p)Y = 0$, which gives $p = Y/(X+Y)$. Any definition of "hold" has such a base rate built in. Unless your test compares with it, the number you report is mostly your choice of $X$ and $Y$.

Write the test down first

A level test should be registered before you look, as a trial registers its primary endpoint. Pre-registration means writing down the hypothesis, the definitions, the data and the analysis before seeing any result, so the test cannot be tuned to the data. For a level, write down seven things:

1. The level: which map, which vendor, which convention, and the time at which you read it, for example the map as it stood at 10:00 ET.
2. The touch: how far away price must be first, and whether a gap through the level counts.
3. The outcome: the bounce target, the break point and the time limit, for example the close.
4. The comparison: the base rate at matched control levels, described below.
5. The sample: the dates, starting after the day you freeze the rule, and the number of touches you need.
6. The analysis: one primary outcome, the difference in hold rate from the controls, with standard errors that allow for touches in one session being alike. Name the controls for the VIX and implied volatility too.
7. The decision: the result that would make you use the level, and the result that would make you drop it.

Controls: matched levels, the VIX and implied volatility

The control arm is a set of levels that differ from the tested level only in not being nodes. Match three things:

• Distance from spot, because distance alone sets most of a hold rate, as the table above showed.
• Roundness. Listed strikes sit on round numbers, and round numbers may attract orders for reasons that have nothing to do with gamma. A control at 5,900 is fairer than one at 5,912.
• The day. Compare the node with controls on the same session, or on sessions of the same volatility.

Volatility is the confounder to fear most. Nodes and holds can both be more common on calm days, and the FlashAlpha test above showed how much of GEX's information is the VIX's.

No look-ahead from intraday open interest

Open interest is counted once a day, overnight, and published the next morning. A map drawn at 2 PM is an estimate, and a map drawn tomorrow morning knows what today's traders did. A backtest that defines today's levels from the day's final map, or from the next morning's open interest, uses information nobody had when the trade was made. That is lookahead bias (Chapter 3A), the market's version of immortal time bias.

LuxAlgo, a charting-indicator vendor, puts the rule plainly: "Do not backtest today's gamma levels across earlier dates as though they were known then. A valid historical gamma study needs timestamped historical levels, their update times, and a supported way to align that data with the price series". Fixed levels help here. SpotGamma publishes its levels each morning, and they "hold for the trading session", so a log of them can be tested. Skylit's nodes move all day, so the level a trader acted on may no longer be on the map. Unless you saved it, you cannot test it.

Out of sample, and how many tries

Every choice you try is a trial: another bounce target, another break point, another expiry, another kind of node. Try twenty definitions and keep the best, and the best beats chance by luck, as the deflated Sharpe ratio of Chapter 1D corrects for strategies. Freeze one definition, then test it on data that comes after the freeze. If you must use history, split it by time and use walk-forward testing, never a random split (Chapter 3A).

Then count how many touches you need. The arithmetic is a trial's sample-size calculation.

A test that needs nine years is a reason for patience, not for skipping the test. It also explains why no vendor statistic in this book's research comes with the counts, controls and errors that would make it readable.

The figure below turns the worked example into a planner. Sliders set the base rate, starting at 80%, the hoped-for edge above it, starting at +5 points, how often the level is touched, starting at one session in five, the touches per session, starting at 1, and their correlation, starting at 0.2. Significance and power stay at the example's 5% and 80%. The curve gives the years of sessions a test takes against the edge, on a logarithmic scale. A dashed twin gives the count when the base rate is estimated from as many control touches.

base rate

80%

hoped-for edge

+5 (85%)

touched on

1 session in 5

touches per session

1

correlation ρ

0.20

touches needed = 470

× design effect 1.0 = 470

sessions = 2,350 ≈ 9.3 years

from controls too: 905 of each

About 9.3 years. The level is touched about 50 times a year. Halve the edge to 2.5 points and the test takes about 39 years.

At the start the test needs 470 touches, 2,350 sessions or about 9.3 years, and 905 touches of each kind if the controls are estimated too. Halve the edge to 2.5 points and it takes about 39 years. Now set 3 touches a session: the design effect of 1.4 raises the count to 658, yet the test takes 4.4 years, because each touching session now gives three touches, worth 2.1 independent ones. A small edge at a rarely touched level takes decades to test, which is one more reason to treat such a level as context.

The last figure shows what trying many definitions does. It scores up to 100 definitions of a hold on the same 50 made-up sessions of a random walk, about 500 touches of levels where nothing happens. Each definition has its own bounce target and break point, from 2 to 20 points, and so its own base rate, break ÷ (bounce + break). The slider sets how many definitions you try, K, from 1 to 100 and starting at 20, and the solid line keeps the best edge over its base rate among the first K. The dashed line runs that best rule, frozen, on 50 fresh sessions. New sessions draws others.

definitions tried, K

20

new sessions

best of 20: +3.4 points

bounce 11, break 17

holds 64.1%, chance 60.7%

fresh sessions: −0.4 points

An edge from nothing. Nothing happens at these levels, yet the best of 20 definitions beats its own base rate by 3.4 points over 501 touches. Frozen and run on 50 fresh sessions, the same rule falls short of its base rate by 0.4 points.

At the start, the best of 20 definitions, a bounce of 11 and a break of 17, holds 64.1% of 501 touches against a chance rate of 60.7%: an edge of 3.4 points from nothing. On fresh sessions the same rule falls 0.4 points short of its base rate. Set K to 1 and only the level test's default is tried: +1.0 points, 81.0% against 80.0%. Set K to 100 and the best edge reaches +4.5 points, while on fresh sessions that rule gives −1.0. About 500 touches is the sample the worked example asks for to tell 85% from 80%. With twenty tries on that sample, luck alone supplies most of a 5-point edge, which is why the definition must be frozen before the test.

5 · Skylit's rules as hypotheses you can test

Chapter 8C found Skylit's trading rules untested. Each can be restated so that a trader with a log of the map can test it. The last column says what the book's mechanics predict, so a test can also check the theory.

Each row needs the discipline of the previous section: definitions frozen in advance, controls matched on distance, roundness and volatility, maps as they stood at the time, and enough touches.

6 · A checklist

Before any trade the map touches, run down the list:

• Data: convention, units, expiries and data time named; intraday parts understood as drift plus estimated flow.
• Regime: spot against the flip, with the flip read as a band; the VIX term structure and realized against implied volatility read first.
• Weight: the GEX sign used as a size multiplier within the range your framework allows, never from its magnitude.
• Calendar: events, expiries and the last hour of the same-day expiry checked.
• Thesis: the trade formed from something other than the map.
• Levels: any level in the plan has passed a registered test with a base rate, or is treated as context only.
• Log: the map saved as it stood, for the next test.

7 · When to step aside

Some days the map is least reliable when it looks most useful. Step aside from trades that lean on the map when:

• The inputs are unknown. The vendor states no convention, or intraday nodes rest on a classification it does not describe.
• Implied volatility is about to jump. CPI and FOMC days move every cell and the flip at once (Chapter 8C).
• The same-day expiry is in its last hour. At-the-money gamma grows most there, and the largest nodes jump from strike to strike.
• Spot sits inside the flip's band, where the sign of the regime is a coin toss.
• Vendors disagree on the sign. Two conventions, two opposite readings: the disagreement is the information.
• The three preconditions of Chapter 8H coexist: short dealer gamma, mechanical one-sided flow and degraded depth. Then size goes down and stops widen, whether or not a catalyst arrives.
• Your level has not passed a registered test, or has failed one.
• The tool costs more than it can add. At $699 a month, a map is a 16.8% yearly hurdle on a $50,000 account (Chapter 8C).

8 · Check yourself