SPY straddles through macro releases: is the event-night move priced too high or too low?
On the night before a scheduled Fed decision, CPI release, jobs report or election, is SPY's at-the-money straddle priced too high or too low, and does buying or selling it pay after costs?
Neither trial passed. Buying the SPY straddle at the close before the release and selling it at the close after lost 4.5% of the debit on average (335 trades; one-sided band -10.9% to +2.3%; median -13.1%; 33% of trades profitable). Selling it lost 2.3% of the credit (band -11.0% to +4.1%; median +5.9%; 58% profitable; with the worst 5% of trades counted twice the mean is -7.9%), and its worst single night lost 3.2 times the credit, on the June 2026 jobs report. Both means are inside the smallest mean this sample could detect (about ±5.3% for the buy side and ±9.3% for the sell side), so the result is not measurable at this sample size, not evidence that the effect is zero. Neither side beat its control: buy-side signal minus control +2.9% (-5.2% to +11.7%), sell-side -3.1% (-13.0% to +5.4%). At the quote mid instead of the bid and ask, the buy side lost 1.8% and the sell side made 0.5%, so the spread is most of the sell side's loss; that mid-fill line is descriptive and never decides the verdict. On the traded nights the next-day move was below the implied move on most nights (median ratio 0.49; above implied on 17%), but the implied figure includes time value past the event. 335 of 336 event nights traded; one was refused on a wide spread.
Both means are inside the smallest detectable mean at this sample size, so the fail is against the pre-registered bar, not a finding that the effect is zero.
Backtested by QS Research
Pre-registration
Two trials were fixed before any SPY option quote was read: buy the at-the-money straddle at the close before a scheduled macro release and sell it at the close after; and sell the same straddle and buy it back at the same times. The events are Fed decisions, CPI releases, jobs reports and US federal elections from 2016 to 30 October 2026: 336 event nights, carrying 343 event types (85 Fed, 127 CPI, 126 jobs and 5 elections; 7 nights carry two events). A release before the open or at 14:00 is held from the close before to the close of the release day; an election from the election close to the next close. Fills are the quoted end-of-day bid and ask at the 16:00 close, with $0.0065 a share a leg; an entry leg needs a two-sided quote, an ask of at least $0.05 and a spread of at most 25% of its mid. The expiry is the first one listed at entry at least two calendar days after the exit; the strike is the nearest listed strike with both legs quoted. The control is the same one-night hold on the first date pair at least five sessions after the event exit that avoids event nights and the sessions before them. The pass bar needs at least 30 trades, a positive mean after costs and a one-sided bootstrap lower bound above zero at 0.05 divided by the options family's 1,868 trials; the buy side also needs positive skew, the sell side must stay above zero with its worst 5% of trades counted twice, and both sides must beat their control. The quote snapshot was pinned before the entries were built. The 2026 election is excluded because its prices were seen while the study was being written, and the five election nights in the list were already seen.
Registered 2026-11-03Amendments
- 2026-11-04 Pinned the event list and the session calendar; dropped releases on a market holiday and unscheduled Fed actions, before any option quote was fetched. approved: Recorded 2026-11-04, before any option quote was fetched
- 2026-11-07 Pinned the option-quote snapshot and recorded its fetch and byte-verification on a second machine; no quote value had been read. approved: Recorded 2026-11-07, before any option quote value was read
- 2026-11-07 Corrected the shared-control disclosure: 13 control pairs serve 27 event nights; added a first-night sensitivity line. No rule changed. approved: Recorded 2026-11-07, before any option quote value was read
Trials · 2 trials · 0 passed · judged at N = 1868 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| straddle-long | FAIL | counted |
| straddle-short | FAIL | counted |
Limitations
- Both trials fail the pre-registered bar, but both means are inside the smallest mean this sample could detect (about ±5.3% for the buy side and ±9.3% for the sell side). The result is 'not measurable at this n', not evidence that the effect is zero.
- Fills are the quoted end-of-day bid and ask at the 16:00 close, with $0.0065 a share a leg; the exit has no spread cap. A mid-fill version is reported in the write-up beside the official numbers; it is descriptive only, and the spread is most of the sell side's loss while the buy side still loses at the mid.
- The implied move is measured from a straddle that keeps at least two days of time value after the exit, so it overstates the pure event move. The realised/implied figures are descriptive, not a tested result.
- No holdout: the full 2016 to October 2026 event history was used to count the trials, so the result is in-sample.
- 335 of 336 event nights traded. One CPI night (April 2025) was refused because both candidate expiries had a leg wider than 25% of mid. The descriptive implied-versus-realised figures are over the 335 traded nights; with the refused night included, the median ratio is 0.50 and the move exceeded implied on 18% of nights.
- The control is counted once per distinct date pair, not once per event night: 13 shared pairs serve 27 nights, so the control mean is over 322 distinct pairs. Keeping only the first night of each shared pair does not change either fail.
- The elections are five nights and were already seen before the pre-registration was written; they are reported, never decisive. The 2026 election is excluded because its prices were seen while the study was being written.
- The session calendar comes from a pinned market-data snapshot whose upstream event list cannot be rebuilt from source; the study reproduces from that pin, not from raw inputs. The quote fetch ran on a different machine than the pre-registration named, and the byte-identical verification needed a market panel file outside the pinned option data (the option path does not read its values).
- The gate file the run used is a later version than the one recorded at the pre-registration merge; the thresholds this study uses (at least 30 trades, a 0.05 base alpha divided by the trial count, and 4,000 bootstrap draws) are the same, but the run's recorded gate hash is not the pre-registration's.
- The sell side has undefined risk: its worst single night lost 3.2 times the credit (June 2026 jobs report), and the mean is negative even before that tail is stressed.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 418ca364e41fa24a, published 2026-11-21