HeadroomDESK

Dealer gamma and BTC volatility: an internally pre-registered null

Pre-registered internally on 2026-11-15, sha256 2fa50e8a8a697770566be2ef32922206001ea82f9a5f00d17c315cbaa831e555 of the locked page, locked by the operator's dated line of 2026-11-15; hash recorded in the program's library on 2026-11-15; not independently timestamped; posted here on 2026-11-17, after the result was known.

By Headroom Research. Verdict: null, read on 2026-11-16 in one sealed read.

The question

When options dealers on Deribit are short gamma, their hedging buys into rallies and sells into drops, which should make BTC's next-day swings bigger and its intraday moves run further. Separately, when option takers pay up for volatility, that may itself warn of bigger swings. The study asked whether a measure of those two things, rebuilt from the options trade tape alone, makes a forecast of the size of BTC's moves (never their direction) better than one that already knows recent realized volatility and the options market's implied volatility. H2 asks only whether a four-hour move extends into the next hour; it does not forecast which way BTC will go.

The verdict

Null. H1, H2 and H3 were each not supported on the held-back year, read once on 2026-11-16. Under the locked page, a null on all three closes the candidate.

What was registered before the test

H1: more negative standardized dealer gamma (Gz, 08:00 UTC) goes with higher next-day realized variance than the baseline, θ < 0, with the floor θ ≤ −ln(1.16)/1.20305 = −0.12337, a 16% variance change per development SD (tested on n = 364 daily units, 2021-01-01 to 2021-12-31).

H3: more net taker demand for volatility over the prior 24 h (Vz) goes with higher next-day realized variance, φ > 0, with the floor φ ≥ +ln(1.16)/1.26695 = +0.11714748… (tested on n = 364 daily units, 2021-01-01 to 2021-12-31).

H2: four-hour moves extend into the next hour more when hourly Gz is more negative, c < 0, with the floor c ≤ −0.03/1.30984 = −0.0229036 (tested on n = 8,736 hours, 2021-01-01 to 2021-12-31).

The family was tested with Holm, m = 3 at α = 0.025 one-sided; H1 and H3 by Clark–West, H2 by a day-clustered t (n = 364 daily units and n = 8,736 hours, 2021-01-01 to 2021-12-31).

Power at the floors was H1 85%, H2 83% and H3 80%, so a non-rejection reads "not supported", not "underpowered" (n = 364 daily units and n = 8,736 hours, 2021-01-01 to 2021-12-31).

A hypothesis counts as supported only if Holm rejects it and its estimate on the held-back year clears its floor. The locked page is published with this release, byte-identical, so its hash can be recomputed.

Changed before the lock, all recorded on the locked page and none after it: the H1 and H3 test moved from Diebold–Mariano on QLIKE to Clark–West after the power runs showed the first could not detect an effect of useful size; H2's three-month forward row was withdrawn for over-rejecting; one planned direction check on the forward collector was withdrawn. One assumption (that the hedging channel's scale transfers) was falsified before the lock, and the construction does not rest on it.

Method, data and what was held out

Dealer positions were reconstructed from the Deribit BTC options taker tape, with the dealer taken as the maker side and option Greeks computed on the parity forward. The outcome is next-day realized variance for H1 and H3, and next-hour return continuation for H2. The development data was read a counted number of times before the lock; the held-back year was released once, for one script whose hash was filed before the run, and read once.

Sampling: daily at 08:00 UTC after settlement for H1 and H3 (n = 364 daily units, 2021-01-01 to 2021-12-31), hourly for H2 (n = 8,736 hours, 2021-01-01 to 2021-12-31).

Development: eight counted reads touched the 2016–2020 data, seven before the lock and the first step of the sealed run; 2021 was read once (development span 2016-11-29 to 2020-12-31; held-back year 2021-01-01 to 2021-12-31).

Held-back data: 10,327 instrument-year files and 2,933,039 trades, with 365 days of the BTC price series (n = 364 daily units and n = 8,736 hours, 2021-01-01 to 2021-12-31); none dropped, none imputed.

Results

Primary (screen) book. Every number below is quoted from the filed closing page, sha256 4f46cecd2f8a4a2527d84b6178f4a37ce6ace6e1168539ad4fba96b438ddc49c, at its rounding.

H1, not supported: Clark–West 1.80, one-sided p = 0.036; estimate −0.046, 95% interval −0.120 to +0.028; floor ≤ −0.123 (n = 364 daily units, 2021-01-01 to 2021-12-31).

H3, not supported: Clark–West 1.21, one-sided p = 0.113; estimate +0.035, 95% interval −0.034 to +0.105; floor ≥ +0.117 (n = 364 daily units, 2021-01-01 to 2021-12-31).

H2, not supported: t = +0.43, one-sided p = 0.667; estimate +0.0026, 95% interval −0.0092 to +0.0144; floor ≤ −0.0229 (n = 8,736 hours, 2021-01-01 to 2021-12-31).

Holm stops at step 1: H1's 0.036 faces 0.00833 and is not rejected (n = 364 daily units, 2021-01-01 to 2021-12-31).

What this does and does not show: one held-back year, one venue, one reconstruction of dealer positioning from taker flow, tested at the stated floors; it does not show that dealer gamma has no effect on BTC volatility in general; it shows that this measure did not forecast at a size worth using in 2021 (n = 364 daily units, 2021-01-01 to 2021-12-31).

In trading terms: one SD more negative dealer gamma went with about 5.7% more next-day variance, and one SD more volatility demand with about 4.6%, against the 16% that would change sizing or stops; H2's sign is wrong and its size about a tenth of the floor, so there is no continuation filter (n = 364 daily units and n = 8,736 hours, 2021-01-01 to 2021-12-31).

Each floor lies outside its estimate's 95% interval (only just for H1); this is descriptive, not a pre-registered test (n = 364 daily units, 2021-01-01 to 2021-12-31).

The direction check on the held-back year held (d = 0.343, interval 0.333 to 0.351), so the run is a verdict, not void (2021 tape: 2,933,039 trades read, 2021-01-01 to 2021-12-31).

Reported beside, with no verdict: the secondary (all-trades) book, the Diebold–Mariano test and all five descriptives have intervals spanning zero, and H2's big-move row also has the wrong sign (n = 364 daily units and n = 8,736 hours, 2021-01-01 to 2021-12-31).

2020 in-sample (step 1, descriptive, written before 2021 was opened): θ̂ = −0.059, φ̂ = +0.059, ĉ = +0.0021; the 2021 estimates have the same signs and are somewhat smaller (2020 development fit; θ̂ and φ̂ on n = 336 daily units).

H-expiry, a declared-null probe outside the family (2020 only): c = +0.82, standard error 0.65, one-sided p = 0.90, on n = 246 expiry days in 2020, of which 17 were top-decile and negative; the declared null stayed null and could not have made the study positive.

Step-2 count, explained: the classifier read 3,052,460 prints; the 2021 code tables total exactly the 2,933,039 rows read; the extra 119,421 are earlier prints of the same instruments, needed only to assign tick codes to each instrument's first 2021 prints (n = 2,933,039 trades, 2021-01-01 to 2021-12-31).

Custody, fences and disclosures

  • The held-back year is no longer unseen for this tape. No variant on it (no other lag, book, threshold or sub-period) will be run; any later idea on this options archive is a new study with its own pre-registration and its own unseen data.
  • Pre-registration custody: the locked page was hashed and filed in the program's library and signed by the operator's dated line; it was not posted to OSF or independently timestamped before the run. The research plan is registered retrospectively on OSF: osf.io/jmu36. The code is archived on Zenodo: doi.org/10.5281/zenodo.24302331.
  • Access to the held-back year was enforced by the program's data-fence register: permits for one read by the filed script, recorded as spent and removed after the run.
  • Working-seat rule events as filed: on 2026-11-10, one path lookup outside the working seat's lane; on 2026-11-15, one search command by a working seat, refused by its harness, with nothing returned and nothing spent; on 2026-11-15, a seat acted on a notification line it had written itself and the schedule checker refused it. None spent a read.
  • No stop in the run's conduct, per the reviewing ruling.
  • Data rights: no trade-level data is published. Code and derived aggregates only.

Files of record

  • Filed closing page: 4f46cecd2f8a4a2527d84b6178f4a37ce6ace6e1168539ad4fba96b438ddc49c
  • Locked page (published with this release, byte-identical): 2fa50e8a8a697770566be2ef32922206001ea82f9a5f00d17c315cbaa831e555
  • Sealed-read script sealed_O.py: 927f0ffd254bb3b85e3de94befbb5292709422cbbcedace04e23d73efdc86c90
  • Output, text: cfbe347deabafeecbf4376251d1003bd6015e5687e6e22e34394d30bbd63eba9
  • Output, JSON: 48a3f35695f63eecda9a3353fd3d5edd2ba8eabe058ed9f42bb905c9db0b0701
  • First-step output (the in-sample year): ecbc41c5988e66130df6433d1cdb4685c76ea6e64d875181b9293ac80fe6a3f9
  • Second-step output (the direction check): bc1a843a5448f650cad744b3d8830451a486390d94331f36e2476373e3181013
  • Run-started record: c5ab5c0ab04ad5290329c7d16ab48f6d6cd7b1e79aeafdba1cbc94dcf4d6f261

All Headroom Research releases, nulls included, are listed on the research page.

Updated 2026-11-17: OSF and Zenodo links added; no number changed.

Headroom Research publishes finished studies as they were filed. Information, not advice.