Year-out ATM options bought when implied vol is at most 0.9x realized
On 87 S&P 100 names the first test never saw, do year-out ATM calls or puts bought when IV <= 0.9x the lower realized read beat the same options bought every month?
Calls pass and puts fail. Cheap year-out ATM calls returned +0.62 per option over the every-month baseline, with band [+0.21, +1.04] and +0.45 on the 2022+ holdout, across 602 trades in 2016-2026. Cheap puts did worse than the baseline (-0.18, band entirely below zero). The calls pass is judged at the pre-registered family of 10 trials; it would not survive deflation against every trial the desk has run (lower bound about -0.015).
Calls and puts are judged as separate trials: calls pass and puts fail (1 of 2). The stated pass is the calls' verdict; the summary states the puts' fail.
Backtested by QS Desk
Pre-registration
One pre-registered cell re-tested unchanged from the first IV/RV test: 300-450 DTE ATM, IV / min(RV 252, RV 63) <= 0.9, monthly entries held to expiry, bought at the ask plus commission. Calls (R1) and puts (R2) are judged separately at trials = 10, each against the every-month baseline with a month-clustered bootstrap band and the 2022+ holdout. The data snapshot was pinned before the run.
Registered 2026-11-03The holdout was opened on 2026-11-05 (Henry, QS admin session, 2026-11-05 about 12:20 PT: "you decide : [this study]'s holdout: does it count as opened?" Decided opened by the QS admin on Henry's delegation: the holdout was read by the pre-registered pass rule Henry approved.).
Trials · 2 trials · 1 passed
| Trial | Verdict | |
|---|---|---|
| Long-dated calls (cheap vs realized vol) | PASS | counted |
| Long-dated puts (cheap vs realized vol) | FAIL | counted |
| Long-dated calls (cheap vs realized vol) | PASS | not counted |
| Long-dated puts (cheap vs realized vol) | FAIL | not counted |
Limitations
- End-of-day bid/ask quotes, not fills; bought at the ask.
- Monthly entries held to expiry; one decade (2016-2026), one mostly rising regime. The baseline itself made +0.26.
- The edge is uneven by year: 2019-2020 are the largest, and 2017, 2018 and 2023 went the other way.
- Passes at the pre-registered family N = 10, not at the desk's global trial count.
- No earnings adjustment.
- 326 data-check failures (missing parity forward, or basis outliers) are listed, not patched; names that were acquired or spun off stop at their last quoted chain.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 86b77b51eeb514ac, published 2026-11-06