Do long options bought when implied vol is at or below realized vol beat buying them every month?
When a long option's implied vol is at or below the lower of its one-year and three-month realized vol, does buying it beat buying the same option on every monthly date, after costs?
Seven of the eight pre-registered tests failed. The one pass is T7, the 0.9x threshold on year-out at-the-money options: 241 trades from 33 monthly entries returned +0.844 per option against +0.241 for buying the same options every month, a difference of +0.603 with a bootstrap band of [+0.006, +1.103], and +0.907 on the 2022-2026 half. The pass is marginal: the observed difference is below the smallest difference the test was built to detect (+1.85), the median trade returned -1.0, and 55.2% of trades lost the whole premium. It is also flattered by the data end: every long-dated entry made in 2026 expires after the last close and is valued at zero, which is 19 of T7's 241 signal trades but 282 of its 1,561 baseline trades, and the long-dated entries of November and December 2025 are valued at zero for the same reason. On simple pooled means, dropping the 2026 trades lowers the difference from +0.709 to +0.617, and dropping the late-2025 ones too would lower it further; whether the bootstrap lower bound stays above zero without them is not known. The four primary tests at the tool's 1.0x threshold all failed, so the tool's 'cheap' verdict is not validated. T7 is published as the lead that the follow-up test on other names went on to test, not as an edge.
One of the eight pre-registered tests passed: T7, the 0.9x sensitivity test on year-out at-the-money options. The four primary tests at the tool's 1.0x threshold all failed, so the headline is a fail; T7 is published as the lead that the follow-up test went on to test, not as a validated edge.
Backtested by QS Research
Pre-registration
Written and amended on 2026-11-02, before any code or trade existed, and run once on a pinned snapshot of end-of-day option quotes. Four cells (30-60 and 300-450 days to expiry, at-the-money and 25-delta) were each tested at two thresholds, 1.0x and 0.9x the lower of one-year and three-month realized vol: eight counted trials, T1-T8. The signal is the tool's implied vol on the expiry's put-call-parity forward over that lower realized read, with same-day quotes only and a spread no wider than 25% of the mid. Each test buys a long call and a long put at the ask on the first trading day of every month, pays $0.0065 per share commission, holds to expiry and scores the return on premium. A test passes only if, at eight trials, the month-clustered bootstrap lower bound of signal minus the every-month baseline is above zero, every signal cell has baseline trades, the long-premium checks pass, and the difference is above zero on the 2022-2026 half alone; the 2016-2021 half is reported but not judged. One amendment before the run moved the implied-vol read from a spot Black-Scholes read to the tool's forward-based read and added the same-day quote and 25% spread filters.
Registered 2026-11-02Amendments
- 2026-11-02 Before the run: the signal's implied vol changed from a spot Black-Scholes read to the tool's Black-76 read on the expiry's put-call-parity forward, and a same-day-quote rule and a 25%-of-mid spread cap were added before the signal is read. approved: Recorded 2026-11-02, before the counted run; no separate approval is recorded.
The holdout was opened on 2026-11-02 (Read by the pre-registered pass rule in the counted run of 2026-11-02; the pre-registration set the 2022-2026 half as the judged holdout.).
Trials · 8 trials · 1 passed · judged at N = 8 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| T1 | FAIL | counted |
| T2 | FAIL | counted |
| T3 | FAIL | counted |
| T4 | FAIL | counted |
| T5 | FAIL | counted |
| T6 | FAIL | counted |
| T7 | PASS | counted |
| T8 | FAIL | counted |
Limitations
- The data ends on 2026-10-29, before the expiry of every long-dated entry made in 2026 and of the long-dated entries of November and December 2025 (their year-out choice expires 2026-12-18), so those trades are valued at zero under the pre-registration's worst-case stale-exit rule. In the passing cell the 2026 slice alone is 19 of the 241 signal trades but 282 of the 1,561 baseline trades; the baseline carries more than twice the signal's share of these zeros, so the measured difference is flattered, and the late-2025 count is not in the result. On simple, unweighted pooled means the difference falls from +0.709 to +0.617 without the 2026 slice and stays positive, which understates the drop from removing all of them, but the judged, weighted lower bound of +0.006 cannot be re-computed without them, so the pass should not be read as robust to this. The last near-dated snapshot (2026-10-01) is affected too, because its 30-60 day options expire after the data ends.
- The one pass is marginal: the observed +0.603 is below the +1.85 smallest difference the test was built to detect, the median trade returned -1.0, and 55.2% of trades lost the whole premium.
- Calls and puts are pooled and judged together. On a simple split, calls were +1.50 over the baseline on 121 trades and puts were -0.10 on 120; this split is descriptive and not weighted like the judged comparison. The follow-up test judged calls and puts separately and found calls passing and puts failing.
- Only the 0.9x threshold passed, and only for year-out at-the-money options. The four primary tests at the tool's 1.0x threshold all failed, so the tool's 'cheap' verdict is unchanged; the edge is conditional on the stricter threshold.
- Judged at the eight pre-registered trials, not against the 2,965 trials QS Research had run when the report was built.
- The counted baseline entered from February 2016, before any name had the 253 daily closes the signal needs (the first signal entries are in 2017, and 2019 in the passing cell). A later descriptive re-run under a corrected baseline reported a lower bound still above zero, but it is not part of the judged result, was not charged as a trial, and its output is not published.
- Twenty large, liquid names chosen in 2026 and still listed today: the result carries survivorship bias and does not transfer to other names without a new test.
- End-of-day bid and ask, not fills: bought at the ask with $0.0065 per share commission, monthly entries, held to expiry with no early exit, roll or hedge, in one mostly rising decade; the every-month baseline itself made +0.24 per option in the passing cell.
- Coverage is uneven: META has 104 near-dated baseline entries against 252-258 for most names (its option chain under that name appears to start in mid-2022), and AMD has 228; the cause is not established for AMD. No earnings adjustment.
- The study reports entry refusals but not the count of entered trades whose exit was valued at the worst case; the 2026 counts above are read from the yearly table. In the two long-dated 25-delta tests the yearly and other descriptive splits count one name and side the signal never trades, so they total 1,464 baseline trades against 1,404 judged.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 9f1576c808cbd61f, published 2026-11-21