SPY in Thanksgiving week 2026: historical base rates and dealer gamma after the November expiry (descriptive)
What does SPY do in Thanksgiving week 2026 (23-27 Nov), given the setup at the 20 Nov close: a 2.3% down day on November monthly opex, a two-day RSI near 0.4, and a second session below the 200-day average?
Descriptive, not a test. This is a snapshot written on 20 Nov 2026, before Thanksgiving week, of what SPY did after similar setups in history and where option dealers' hedging sat. The condition matching the 20 Nov state (a drop of at least 2%, a two-day RSI at or below 2, and a close below the 200-day average) occurred on 28 sessions in 23 separate episodes since 1993: 79% were up four sessions later, with a mean of +2.25% and a median of +1.95% (10th to 90th percentile -1.08% to +5.57%). Since 2016 the same state is only 6 sessions (83% up, +2.90%). The first five sessions below the 200-day average after at least 60 closes above, when the two-day RSI was under 5, were 30 sessions in 18 episodes: 57% up and +0.18% on average; the first session of an episode was 17 sessions, 47% up and -0.38%. Thanksgiving week itself was up in 73% of 33 years (+0.63% average); when the prior week was down, 85% of 13 years; when the prior Friday was down, 62% of 13. The week after November opex was up in 70% of 33 years (+0.69%); when opex day was down, 62% of 13. Net dealer gamma within 20% of spot was negative: -$11.9 billion per 1% move across all expiries and -$8.6 billion with the 20 Nov expiry removed; the cumulative gamma ladder never crossed zero (no flip level), the largest put-gamma level was 700, and the largest call-gamma level was 700 across all expiries and 775 with the 20 Nov expiry removed. The study's judgement is about a 60% chance SPY is up for the week, a range of about 670-727, and no trade. That is a judgement, not a count, and the week's outcome was not known when it was written. The study was not pre-registered: the record is retrospective, written on 2026-11-21 after the results, and 26 conditions are a count of what was looked at, not a family fixed in advance.
The study is a descriptive snapshot, not a counted test: it sets no pass bar and counts no trials. The 26 conditions are a count of what was looked at, not a pre-registered family, and nothing was judged against them.
Backtested by QS Research
Pre-registration
No pre-registration existed. A retrospective record was written on 2026-11-21, after every result was seen, to describe the question, data, rules and judgement exactly as the committed code and write-up show they were run. It changed no rule and set no pass bar. The question was what SPY would do in Thanksgiving week 2026 given the 20 Nov setup. The data were a pinned daily SPY price history from 1993 to 13 Nov 2026, four closes from 16-19 Nov from a consolidated feed, one provisional intraday 20 Nov bar (open 713.56, high 718.29, low 698.03, last 698.106), and per-contract SPY open interest and implied volatility from a licensed options feed kept outside the repository; only by-strike gamma outputs are committed. The record states the rules as run: Wilder RSI(2), simple 200- and 50-day averages, close-to-close returns, forward returns over 1, 3, 4 and 5 sessions, episodes that merge signal days within five sessions, and a 95% normal interval on the mean. It lists 26 distinct conditions looked at: 11 oversold and drop conditions on the full history and from 2016, 3 Thanksgiving-week conditions on two windows, 3 post-opex conditions, 4 follow-up conditions (three around a first close below the 200-day average, plus a drop-and-close-in-range condition), and 5 post-hoc RSI(2) cuts chosen after seeing the 20 Nov reading. The oversold and drop, follow-up and post-hoc conditions are read at 1, 3, 4 and 5 sessions; the first 11 also on two spans, the Thanksgiving-week conditions on two windows, and the follow-up and post-hoc conditions on all days and first of episode. The follow-up run used a slightly later tick whose exact last and low were not logged, so a few of its rows differ slightly from the printed command. The gamma map is Black-Scholes gamma from mid implied volatility at a 4% rate, net of call minus put gamma, with walls at the largest call and put gamma, a cumulative-ladder flip within 10% of spot, and strikes within 20% of spot, computed with and without the 20 Nov expiry; the expiring 20 Nov contracts are valued at 6.5 hours to expiry in the all-expiries column. The record notes that the scripts say the first 21 conditions were fixed before their returns and the 5 post-hoc cuts were not, but no committed record before the results shows the first claim; the count of 26 is a count of conditions looked at, not a pre-registered family.
Registered 2026-11-21Trials · 0 trials · 0 passed · judged at N = 26 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| all days | RESULT | not counted |
| RSI2 < 5 | RESULT | not counted |
| RSI2 ≤ 2 & below MA200 | RESULT | not counted |
| drop ≤ −2% | RESULT | not counted |
| RSI2 ≤ 2 & below MA200 & drop ≤ −2% (today's state) | RESULT | not counted |
| same, since 2016 | RESULT | not counted |
| fresh MA200 break, sessions 1–5, RSI2 < 5 (today) — all days | RESULT | not counted |
| same — first day of each episode | RESULT | not counted |
| drop ≤ −2% & close in bottom 10% of range — first of episode | RESULT | not counted |
| Thanksgiving week, prior Fri close → half-day close, 1993–2025 | RESULT | not counted |
| … prior week down | RESULT | not counted |
| … prior Friday down | RESULT | not counted |
| week after November opex (3rd Fri → next Fri) | RESULT | not counted |
| … opex day down | RESULT | not counted |
| RSI2 ≤ 1.0, all days | RESULT | not counted |
| RSI2 ≤ 1.0, first of episode | RESULT | not counted |
| RSI2 ≤ 1.0 & below MA200, all days | RESULT | not counted |
| RSI2 ≤ 1.0 & below MA200, first of episode | RESULT | not counted |
| RSI2 ≤ 0.42, first of episode | RESULT | not counted |
| net total | RESULT | not counted |
| net within ±2% | RESULT | not counted |
| gamma flip | RESULT | not counted |
| put wall | RESULT | not counted |
| call wall | RESULT | not counted |
| 690 | RESULT | not counted |
| 695 | RESULT | not counted |
| 703 | RESULT | not counted |
| 705 | RESULT | not counted |
| 710 | RESULT | not counted |
| 715 | RESULT | not counted |
Limitations
- No pre-registration: the record is retrospective, written on 2026-11-21 after every result was seen; no rule was pre-registered and no pass bar was set. The 60% call and the 670-727 range are the study's judgement, not a tested result, and the week's outcome (the 27 Nov close) was not known when it was written.
- The base rates use a provisional intraday 20 Nov bar (last 698.106 at about 11:13 PT), not the 699.33 close. The follow-up run used a later tick whose exact last and low were not logged; a few of its rows differ slightly from the printed command. The setup table's 4% range position is from that later tick; the 698.11 bar is at 0.4% of its range.
- Samples are small and overlap: the today's-state row is 28 sessions in 23 episodes (6 since 2016), the fresh-break row is 30 sessions in 18 episodes (17 first-of-episode), and the calendar subsets are 13 years. Four- and five-session windows overlap; the intervals are normal 95% intervals on the mean, not bootstrap or episode-clustered.
- The 26 is a count of distinct conditions looked at, not a pre-registered family. The post-hoc RSI(2) cuts (at or below 1.5, 1.0 and 0.42, and the 1.0 split above and below the 200-day average) were chosen after seeing the 20 Nov value and are colour, not evidence. The main table mixes full-history and since-2016 rows and all-days and first-of-episode bases.
- No holdout was reserved and no trial was counted; nothing here is tested against a trial bar. The study is a descriptive snapshot, not a forecast with a measured probability.
- The gamma map uses licensed per-contract open interest and implied volatility that are not committed; only the by-strike outputs are in the repository. Open interest is as of the 19 Nov close (file stamped 20 Nov), and the live page and the recompute disagree on size at several strikes; that is not reconciled. The all-expiries column values the 20 Nov contracts at 6.5 hours to expiry, not zero.
- The report format used here has no data-pin field for a transcribed table, so the report carries no pinned data snapshot or release commit: the daily-price snapshot is named in the write-up but not recorded as a pin, and the options inputs are not committed. The published record cannot be re-run from a pin. The 16-19 Nov closes come from a consolidated feed without dividend adjustment and were appended to an adjusted series.
- No positive-gamma level exists between spot and 721 (strike 673, below spot, has a small positive net of +0.7M). The VIX figures in the setup table are not in the committed live snapshots and cannot be checked from the published record.
- The contract count and the by-expiry put open-interest breakdown in the write-up are not in the committed result files; only the total 498k puts at 700 is in the summary output.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 1a65309366d59b98, published 2026-11-21