What the stock does in the new extended-hours option windows
After a large move by 07:30, or an earnings release in the 16:00-16:15 window, does the stock continue or reverse by enough to beat the spread?
None of the 38 pre-registered trials passed. Every trial failed the first check: the lower bound of the mean return after costs was below zero. Continuing a large 07:30 move made 0.086% before costs but lost 0.091% after them on 1,307 trades; fading it lost more. The closest was continuing a 16:00-16:15 earnings move to the next 07:30: 0.519% before costs and 0.272% after them on 119 events, but its lower bound was -0.151%. The shortlist is empty, so validation and the holdout were not judged and the holdout stays sealed. Under the pre-registration's no-edge branch, the public product page shows the descriptive "most likely move" table only, labelled as history.
Backtested by QS Research
Pre-registration
38 trials were fixed before the data was read: 32 on a large move from the prior close to 07:30 (two size thresholds, two exits, continue or fade, and four cause groups) and 6 on an earnings release in the 16:00-16:15 window (continue or fade, three exits). Discovery is 2016-2021, validation 2022-2024 and holdout 2025-2026. Each trial is judged on a session-clustered lower bound after costs, the result at double costs, both halves, the best month removed, a per-year check and a deflated Sharpe at 38 trials. The holdout opens only for a trial that passes validation.
Registered 2026-11-02Amendments
- 2026-11-03 The data snapshot is pinned; no trial, window, grid or gate changed. approved: Recorded 2026-11-03 before any window return was computed
- 2026-11-03 Producer archive, dataset name, and the earnings, macro and corporate-action sources are recorded; no rule changed. approved: Recorded 2026-11-03 before any quote, table or window return was read
- 2026-11-03 Inputs pinned, evaluator conventions fixed, and the two pre-run checks (earnings timing and the earnings-date cross-check) reported. approved: Recorded 2026-11-03 before the counted run
- 2026-11-04 Descriptive only: an ETF version of the most-likely-move table; no counted trial, window, grid or gate changed. approved: Recorded 2026-11-04 after the counted run; descriptive only
A holdout sample is sealed: its results stay unread until the study is final.
Trials · 38 trials · 0 passed · judged at N = 38 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| H1-01 | FAIL | counted |
| H1-02 | FAIL | counted |
| H1-03 | FAIL | counted |
| H1-04 | FAIL | counted |
| H1-05 | FAIL | counted |
| H1-06 | FAIL | counted |
| H1-07 | FAIL | counted |
| H1-08 | FAIL | counted |
| H1-09 | FAIL | counted |
| H1-10 | FAIL | counted |
| H1-11 | FAIL | counted |
| H1-12 | FAIL | counted |
| H1-13 | FAIL | counted |
| H1-14 | FAIL | counted |
| H1-15 | FAIL | counted |
| H1-16 | FAIL | counted |
| H1-17 | FAIL | counted |
| H1-18 | FAIL | counted |
| H1-19 | FAIL | counted |
| H1-20 | FAIL | counted |
| H1-21 | FAIL | counted |
| H1-22 | FAIL | counted |
| H1-23 | FAIL | counted |
| H1-24 | FAIL | counted |
| H1-25 | FAIL | counted |
| H1-26 | FAIL | counted |
| H1-27 | FAIL | counted |
| H1-28 | FAIL | counted |
| H1-29 | FAIL | counted |
| H1-30 | FAIL | counted |
| H1-31 | FAIL | counted |
| H1-32 | FAIL | counted |
| H2-01 | FAIL | counted |
| H2-02 | FAIL | counted |
| H2-03 | FAIL | counted |
| H2-04 | FAIL | counted |
| H2-05 | FAIL | counted |
| H2-06 | FAIL | counted |
Limitations
- The 23-name list (22 with data in the tested years) is a survivor list chosen in 2026 and applied back to 2016; the result is about these names, not large caps in general.
- Early pre-market is thin: only 15% of the counted names' 2016 sessions were usable, rising to about 70% in 2020-21; discovery leans on 2018-2021.
- Stock, not options: the study prices the stock's own move. Option spreads in these windows are wider, so a stock rule that fails after stock costs would do worse with options. No option history exists there before late 2026.
- Bars, not fills: entries and exits are 1-minute bar closes and costs are median quoted half-spreads; size and market impact are not modelled. One spread per name and window covers the whole 2016-2021 discovery period, not each year.
- Earnings timing is a proxy: of 50 sampled 16:00-16:15 events, 39 had their first headline in the window, 6 had only a preview before 16:00 and 5 had no headline in the surrounding two hours.
- The 16:00-16:15 test has 119-124 trades, at most one per name per quarter; the smallest 07:30 cells have 50-61 trades and cannot rule out a small edge.
- The 2022-2024 validation period was summarized in the descriptive tables before the counted run, so it is not a clean out-of-sample period. No shortlist was chosen, so no validation verdict was made.
- The descriptive set is partial: the pre-registered 08:30-08:45 macro reaction and the by-name drop counts were not produced; the macro table gives only the 07:30-09:25 size ratio by release type.
- The counted run was made once, from a clean local build of the recorded release; it has not yet been re-run byte-for-byte from a signed release.
- Two SPY dividend ex-dates (2016-03-18 and 2018-06-15) are missing from the corporate-action data. SPY is not counted, but its 07:30 move defines the index-driven cause, so names on those two dates (up to 44 name-sessions) can change cause.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 0ae4b95f179e3565, published 2026-11-21