Event-night reversal on SPY: does buying a gap-down open after CPI or jobs news pay by the close?
When CPI or jobs-report news lands while the market is closed and SPY opens below the previous close, does buying at the open and selling at the close pay after costs, and does it beat the same rule on ordinary days?
No. Buying SPY at the open on a CPI or jobs-report morning and selling at the close lost 2.9 basis points a trade after costs over 147 discovery trades from April 2016 to June 2022, against a 2.2 basis point loss for the same rule on 1,083 ordinary sessions. The month-clustered lower bound was -12.5 basis points, only 3 of 7 years were positive, double costs also lost money, and the deflated Sharpe was 0.000025 against a 0.80 bar. The measured loss is inside the smallest mean this sample could detect, about 16 basis points, so the fail is against the pre-registered bar, not proof that the edge is zero or negative. The gap-down version had only 26 trades, below the 30-trade floor, so it is reported as not measurable at this sample size, not as evidence of no effect; its discovery mean was -16.3 basis points against +7.8 for ordinary gap-down sessions. Validation had nothing to test and the holdout stays sealed.
Trial 1 (every event session) failed after costs, but its mean is inside the smallest mean this sample could detect, so the fail is against the pre-registered bar, not a finding that the effect is zero. Trial 2 (a gap down of at least half a typical day's move) had only 26 trades, below the 30-trade floor, and is reported as not measurable at this sample size, not as evidence of no effect.
Backtested by QS Research
Pre-registration
Two trials were fixed before any CPI or jobs-report price was read. On each event session, buy SPY at the 09:30 open and sell at the 16:00 close; trial 2 takes the same trade only when the open is at least half a typical day's move (0.5 sigma) below the previous close. The event list is the pinned CPI and jobs-report nights from January 2016 to 30 October 2026: 253 sessions, of which 246 were judged after dropping 6 with too little prior history and 1 early close. The split is 60/20/20 by date: 147 discovery, 49 validation and 50 holdout. The control is the same rule on 1,834 ordinary sessions that are not event or election sessions, not adjacent to one, not on a Fed statement day, and not an early close or one of the first 61 sessions with too little history. Costs are 2.5 basis points a side, with a doubled-cost stress. A trial passes only if, in discovery, its month-clustered lower bound is above zero, it stays positive at double costs, its per-year record holds, it beats the same rule on ordinary sessions, and its deflated Sharpe clears 0.80; validation uses the shortlist rule. The last 20% is a sealed holdout that only the study's owner can open. Amendment 1 pinned the price file and the session counts before any price row was read; Amendment 2 pinned the session table and the filled dividend records and fixed the control-stage and deflated-Sharpe readings before the run, and was approved after the result was known. Election nights are reported apart and never judged.
Registered 2026-11-04Amendments
- 2026-11-04 Pinned the daily price file and the session counts (246 judged, 1,834 control, 6 without enough history, 1 early close) before any price row was read; every ex-date's prior close is dividend-adjusted and two missing dividend records were filled. approved: Recorded 2026-11-04 before any price row was read; reviewed the same day.
- 2026-11-04 Pinned the session table and the filled dividend records; fixed two readings before the run: a control session takes the stage of its date, and the deflated Sharpe uses the whole shared trial ledger. No trial, rule or bar changed. approved: Recorded 2026-11-04 before the run; approved by the study's owner the same day, after the result was known.
A holdout sample is sealed: its results stay unread until the study is final.
Trials · 2 trials · 0 passed · judged at N = 2 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| Every event session: buy SPY at the open, sell at the close | FAIL | counted |
| Gap down of at least 0.5 sigma: buy SPY at the open, sell at the close | FAIL | counted |
Limitations
- The open is not the overnight low: most of a dark night's recovery can happen before 09:30, and a pre-market entry is a different, untested trade with thinner quotes.
- Only SPY, daily official open and close prices, a flat 2.5 basis points a side and one share; size and the opening auction's depth are not modelled.
- The family is judged only on CPI and jobs-report sessions. Election nights are five seen sessions, reported apart and never counted.
- The gap-down trial has only 26 discovery trades, below the 30-trade floor. It is not measurable at this sample size; its numbers are shown for what they are.
- The measured loss on trial 1 is inside the smallest mean the sample could detect (about 16 basis points), so the fail is against the pre-registered bar, not evidence that the edge is zero or negative.
- The pre-registered minute-by-minute overnight-low table was not built. It is descriptive and would not change the verdict.
- The counted run was a clean local build of the release commit, and no byte-identical re-run on a signed release has been recorded. The pinned inputs and the full result file are committed.
- The deflated-Sharpe bar used 1,739 trials on the ledger at the run. That ledger is live, so a fresh run would use a larger count; the fail is unchanged because both trial Sharpes are negative.
- Mega-cap earnings nights and any news not on the pinned release list are out of scope; a list chosen today would be a survivor list.
- The holdout, the last 20% of event sessions from 2 August 2024, was never read and stays sealed.
- Amendment 2 was recorded before the run but approved after the result was known; neither reading can change the verdict.
- The smallest detectable mean was promised by the pre-registration but not printed by the run; it was computed afterwards from the pinned files (about 16.3 basis points for every event session and 46.4 for the gap-down trial) and changes no gate outcome.
- Before the run, a descriptive table pooling five ETFs including SPY, with the median open-to-close return after overnight moves on macro-release and other days, was added to another study's records. It has no SPY-only figure and no gap-down condition; whether it was looked at before the run is not recorded.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 21dd6f142f591277, published 2026-11-21