Buying an at-the-money call or put five sessions before an earnings release and selling the session before it
Does an at-the-money call or put bought five sessions before a company's earnings release and sold at the close of the session before it make money after costs, and does it beat the same hold in a quiet week?
Both pre-registered trials fail. An at-the-money call bought at the close five sessions before the filing and sold at the close of the session before it lost 13.0% of its premium on average after costs, across 9,363 trades on 957 entry dates (band -17.8% to -8.2%). The same put lost 19.0% (-23.6% to -13.9%). Calls lost 8.4 percentage points less than the same hold in a quiet week (band +1.6 to +17.0 points), so the call side shows a real pre-filing edge over quiet weeks, but it does not make the trade profitable; puts showed no difference (-0.4 points, band -7.6 to +7.9). The quiet-week control itself lost 21.4% (calls) and 18.7% (puts) over four sessions, so the spread and four days of decay dominate. The call side is a lead for a separate pre-registered test, not a result on its own.
Backtested by QS Research
Pre-registration
Two trials were fixed before any pre-filing quote was read: buy an at-the-money call (trial 1) or an at-the-money put (trial 2) at the close five sessions before the filing day, and sell at the close of the session before it. Entries pay the ask plus $0.0065 a share; exits receive the bid less $0.0065 a share. The contract is the strike nearest 0.50 delta with 9 to 40 days to expiry, nearest 14. Each trade is compared with the same contract rule in a quiet window 21 to 25 sessions after the filing. A trial passes only with at least 30 trades, a positive mean after costs, a positive one-sided bootstrap lower bound of the mean, positive skew, and a positive lower bound on the signal-minus-control difference. The entries were committed with their sha256, the combined data snapshot was pinned, and the release is recorded.
Registered 2026-11-01Amendments
- 2026-11-01 Before any result: a quiet-window control (the same contract rule in a window 21 to 25 sessions after the filing), one combined pinned snapshot for signal and control quotes, and entries generated mechanically from the event lists. approved: Recorded before any result and reviewed
Trials · 2 trials · 0 passed · judged at N = 1646 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| runup-call | FAIL | counted |
| runup-put | FAIL | counted |
Limitations
- End-of-day quotes only. Buys pay the ask and sales receive the bid, with $0.0065 a share charged on each side; real fills and size are not tested.
- The names are today's listings, so companies that later delisted are missing. That flatters the call side, which is the side with the positive gap to its quiet-week control.
- The pass rule's own power check (whether it can detect a planted edge) qualifies a pass; this study has none, so it does not change the reading.
- Filing dates are the realised dates, not the dates the companies announced in advance. A moved date is a small look-ahead.
- The event list behind the data cannot be rebuilt from recorded inputs. Results stand on the pinned list: a rebuild gives 10,222 events against 10,004, with 9,159 in common, so the selection rule is not reproducible.
- The significance bound is corrected for the whole trial ledger, 1,645 at the call run and 1,646 at the put run, not for the two trials this study pre-registered. The pre-registration names a family of two but its pass bar divides by the options family's counted trials; the backtest read the whole trial ledger. The two legs are therefore not judged at the same N, the bands are much stricter than a two-trial correction. The count the backtest read was a live ledger number, not a pre-registered one; the judged family is recorded as 1,646 for both legs, while the call's bound remains at 1,645.
- Coverage is incomplete and control skips are not broken out. 323 of the 10,004 events (3.2%) are dropped before entries because a fetch was not ok (which includes control windows past the data end; the split is not recorded). The signal has 9,363 trades and 318 skips (267 chain mismatch, 40 no expiry in window, 8 no quote, 3 split in life); the control has 9,270 trades and 411 skips, 93 more than the signal, with no reasons recorded. The comparison is not event-for-event.
- One event (PSA, Public Storage, entered 2022-07-28) had no usable quote for the pre-filing exit in either leg, and the engine fell through to expiry on 2022-08-19, through the release. That is not the pre-registered exit; it stays in the counted trades, though it moves neither mean by more than 0.01 points and does not change the verdict. The result reports no unfillable trades.
- The entries cover 542 underlyings and 537 produced at least one trade; the pre-registration's 550 and the event-list check's 547 are counts of the list, not of this study's entries (pre-registration erratum, 2026-11-21). The 2021-2026 window is one regime, with no holdout: every year, including 2026, was read.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 394025e30f01fab7, published 2026-11-21