Earnings straddles held through the release: is the implied move mispriced?
If you buy or sell the at-the-money straddle at the last close before an earnings release and close it at the next close, does it make money after costs, and is the option-implied move too high or too low?
Both trials fail. Buying the at-the-money straddle at the last close before the release and selling it at the close after lost 17.3% of the debit on average (9,410 trades on 1,003 entry dates; one-sided band [-19.5%, -15.2%]; median -26.9%; 23.1% of trades made money). Selling the same straddle lost 35.0% of the credit (9,409 trades; band [-41.1%, -29.4%]; median -13.4%; 36.8% made money; with the worst 5% of trades counted twice the mean is -48.1%). Both beat the pre-registered control about a month later on the mean (+4.2% [+0.5%, +7.7%] for the long; +9.9% [+0.9%, +20.2%] for the short), but both means and bounds are below zero, so neither passes (bounds judged at 1,659 options-family trials for the long and 1,660 for the short). On the pre-registered comparison, which includes at least two days of time value in the implied figure, the realised move was smaller than the implied in 74.0% of the 9,402 paired events (median realised/implied 0.57). That did not pay after costs: the median quoted spread was 13.3% at entry and 15.9% at exit (mean 21.3% and 23.0%), and at mid fills the short still lost 2.9% on average even though its median was +9.0%; the long's at-mid mean was +1.8% but its median was -10.1%. The short mean of -35.0% is inside the +/-37.0% smallest detectable mean at this sample, so the data cannot rule out a smaller true mean, but the pass bar is a positive mean and its upper bound is -29.4%. The long mean is lifted by a few very large winners; its median trade lost 26.9%, worse than the short's 13.4%. The result is a fail for both trials, and the implied move is a magnitude prior only, not a trade signal.
Backtested by QS Research
Pre-registration
The pre-registration was binding before any post-release option quote for this family was read. Two trials: buy the at-the-money straddle at the last close before the release and sell at the close after it, or sell the same straddle and buy it back at the same times. The reaction session is the session of the earnings filing if it was accepted before 16:00 ET, otherwise the next session; a release before the open that was filed after the close is placed one session late, and those rows stay in the counted run. The contract is the nearest listed expiry at least two days after the exit, and the listed strike nearest the entry close with both legs quoted, ties to the lower strike. Entry is at the ask for the long and the bid for the short, exit at the bid for the long and the ask for the short, with $0.0065 a share per leg; a refused entry is counted, never dropped. The control is the same one-night hold about a month later. The pass bar is at least 30 trades, a mean after costs above zero, a one-sided lower bound above zero at 0.05 divided by the options family's counted trials (1,659 for the long trial and 1,660 for the short), positive skew for the long trial with entry-date resampling, month resampling for the short with its worst 5% counted twice, and beating the control. The data snapshot was pinned and the entry lists were committed before the run; there is no holdout. The long side's sign had been seen in an earlier exploratory check, so the long result can only confirm it.
Registered 2026-11-02Amendments
- 2026-11-02 Froze the data-fetch specification before any option quote was read: one pinned snapshot holding the signal and control chains, the sessions to fetch, the expiry and strike rules, the per-event status list and the coverage counts to check before the pin. No trial, rule or pass bar changed. approved: Recorded before any option quote was read
- 2026-11-03 Pinned the data snapshot and recorded the coverage counts by status and year, before the entry lists were generated and before any quote was read. No trial, rule or pass bar changed. approved: Recorded before the entry lists and the counted run
- 2026-11-03 Recorded how the engine applies the contract rule: each entry names its own two candidate expiries with a fallback, the nearest quoted strike is used, and the quote-age, spread, minimum-ask, fee and participation settings are listed. No trial, rule or pass bar changed. approved: Recorded before any entry file was written and before any quote was read
Trials · 2 trials · 0 passed · judged at N = 1659 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| straddle-long | FAIL | counted |
| straddle-short | FAIL | counted |
Limitations
- The universe is survivors only: 547 names chosen from today's listings, applied back to 2021. The 10,004-event list has 21 events with no reaction session and 129 whose control window falls after the data end; those 150 are left out of both sides, so each side's denominator is 9,854 events. The 21 are not counted as refused entries. Every other missing side is a counted refusal.
- Fills are end-of-day quoted ask and bid, with $0.0065 a share per leg. A limit order could do better; the mid-fill version is reported but not judged. Returns are on premium: the long on its debit, the short on its credit, and the short's worst trade lost 37.6 times its credit.
- Event dates are the realised filing dates; a date that moved is a small lookahead. The long side's sign was seen in an earlier exploratory check before the pre-registration, so the long result can only confirm it; the short side, the control and the slices were unseen.
- The 10,004-event list cannot be rebuilt from its source: a fresh collection gives 10,222 events with 9,159 in common. The pinned snapshot is fixed, but the selection rule was not recorded, so reproducibility from source is not established.
- The pre-registered size-tier and before-open/after-close cuts were not produced: no size field exists, and the timing split that is shown is by the 8-K rule's same-session/next-session placement, not by before-open/after-close. The timing field that would identify pre-market rows is empty, so the pre-market sensitivity slice could not be run.
- The spread figures are medians: 13.3% at entry and 15.9% at exit. The mean spread was 21.3% and 23.0%, so the average cost is higher than the median suggests.
- The long mean is lifted by a few very large winners: the median long trade lost 26.9%, worse than the short's 13.4%. At mid fills the long mean is +1.8% but its median is -10.1%; the short mean is -2.9% but its median is +9.0%.
- The control is one month after the release, not a matched quiet date. The pre-release premium is higher, so the same dollar spread is a smaller percentage of the signal's premium; the +4.2% long and +9.9% short beats are the pre-registered comparison, not a trade signal, and can be flattered by that denominator.
- Seven events traded different expiries between the long and short trials and one also a different strike, because the expiry fallback is applied per side. They are left out of the paired slices; only one event traded in one trial only. This is a small departure from selling the same straddle.
- The short mean of -35.0% is inside the +/-37.0% smallest detectable mean at this sample, so the data cannot rule out a smaller true mean. The pass bar is a positive mean and its lower bound; the short's upper bound is -29.4%, so it fails.
- The significance correction is not a single fixed number: the long bound used 1,659 options-family trials and the short 1,660; the published family N is 1,659. The pre-registration fixed two trials under this name, but the bar divides by the wider options family count.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current b10afe88d344f298, published 2026-11-21