Blind Bracket

The market prices these almost exactly right

Across 83 earnings events with real historical option prices, the mean straddle cost 7.68% of spot against a mean realised move of 8.05%. That is a market ratio of 0.95 — the option is priced at ninety-five cents on the dollar of what the underlying actually does.

The play's entire premise was that a disciplined filter could find the cheap tail of that distribution. Tested against real prices, it cannot.

Both winners were accidents

NVDA · 2026-08-26 → 08-27 · +$666 (+27.2%)

Pre-registered prediction A said the trade would lose, at p = 0.64. It won. The profit came from the 36% branch landing — which is evidence against the model, not for it. Of four locked predictions, the only one that scored correct was B, the claim that direction is unpredictable.

AFRM · 2026-08-27 → 08-28 · +$70 (+4.2%)

Entry at 15:45 struck the straddle at 76 while the stock closed at 77.49. That offset was worth roughly $300 against a properly centred strike — more than the entire profit. Without an unregistered directional tilt that nobody chose, the trade lost money.

Six errors, in the order they enabled each other

1. The price ceiling was an identity, not a measurement

   Expected return on premium is exactly mean(|move|) / cost − 1. The cited "+6% at 5.79%, −12% at 7.0%" reproduces from the pooled mean alone. Nothing had been measured about a special cost regime.

2. A pooled statistic was applied per symbol

   Six of eight studied symbols carry a fair straddle above the 6.5% ceiling — NVDA included. The rule was selecting low-volatility names, not cheap options. Same error that had already killed the news study's ρ of 0.432.

3. The correction was sound and sequenced wrong

   A decision-neutrality check was written against AVGO, which did not motivate the change, while omitting AFRM, which did. Checking the wrong name is how a threshold gets tuned to fit a trade without anyone feeling dishonest.

4. Fair-value denominators were drawn from dead regimes

   Every historical mean came from daily bars starting 2020–21. AFRM's 15.55% against a real 11.22% is the clearest case, and it is the one that fired a live trade.

5. The supporting study was never a backtest

   The n = 191 result reporting +30.6% assumed entry at 0.8× fair value. The market offers 0.95× on average. It was a payoff calculation under an assumed price, and the gap between it and reality is the whole play.

6. A precondition was documented but never enforced in code

   DELL's 08-28 straddle priced at 3.27% of spot against a 12.39% historical mean — apparently the cheapest setup ever seen. It expired three days before the announcement. The option contained no earnings at all.

What was actually measured

These survive the play's closure. Each is stated with the control that produced it, because in every case the control is what made the number mean anything.

• 45–53%Gap direction is unpredictable. Six pre-event price signals across 238 filing-anchored events, null 50%, SE ~3%.
• ρ −0.261Range compression forecasts move size. The only pre-event signal that survived every control — and it predicts the median, not the tail an option prices.
• 0.95×The market prices earnings straddles at 0.95 of the realised mean move. 83 events, real prices.
• t 3.58 → 1.20Post-earnings drift is market beta. Raw +2.29% became +0.64% once benchmarked against QQQ over identical sessions with point-in-time betas.
• ρ 0.077News predicts nothing. A pooled ρ of 0.432 collapsed to 0.077 within symbol; direction ran 49.6–54.0%. It was measuring which stock it was looking at.
• 34 of 68SEC item 2.02 is not earnings. Half of Tesla's filings are vehicle-delivery reports carrying identical metadata; only the filing body distinguishes them.
• 0.49%Median strike offset from a 15:45 entry. Zero expected value, non-zero variance, and mostly strike granularity rather than drift — so it cannot be timed away.
• 3.5%Observed round-trip spread cost in premium terms, from two real fills.
• ρ 0.013Per-symbol cheapness does not persist. A symbol's trailing cost/fair ratio carries no information about its next straddle, so even genuine dispersion would be untradeable.

The last thread, closed

AVGO at a ratio of 0.64 and TSLA at 0.74 were the one observation left standing when the play was shut. They were also selected because they topped a table of eight — so re-measuring them on the same events would only confirm whatever picked them. Three non-circular tests, all failed.

• p = 0.080Permutation. Shuffling returns across symbols 20,000 times while holding group sizes fixed, the best symbol averages +39% under a pure null, with a 90th percentile of +60%. AVGO's +62% sits inside that distribution. It is what the top of eight buckets looks like.
• ρ = 0.013Persistence. Cheapness is untradeable unless it is knowable in advance. Estimating each symbol's ratio from prior events only, the cheapest half returned −7.6% and the richest half −7.0% — a 0.5pp gap in the wrong direction, t = 0.09. A real effect predicts negative correlation. This is zero.
• 0.63 → 0.92Decay. TSLA's ratio drifts from 0.63 in its first half to 0.92 in its second, and its return from +62% to +3%. AVGO drifts 0.58 → 0.71. Both converge toward fair pricing within the sample — the pattern you would expect if the early cheapness was ever real and has since been arbitraged.

The persistence test is the one that settles it. Even had the dispersion been genuine, knowing a symbol has been cheap carries no information about whether its next straddle is cheap — which is the only form the knowledge could have been used in.

Buy the real prices first

Every decision before the final data pull — the ceiling, the ratio correction, the sizing rule, two live trades and a great deal of argument about entry timing and exit discipline — reasoned about an assumed entry price. The moment real option prices arrived, the play evaporated in a single afternoon.

The pull cost nothing. Alpaca's option history starts 2024-01-18, gives minute bars, and validated exactly against a trade we had already placed by hand: it read the 2026-08-26 NVDA straddle at 12.29 against the 12.25 actually paid. Four days of theory could have been four days of measurement.

The counterweight is that the free data is trades, not NBBO, so recorded costs are understated and any edge computed from them is overstated — the flattering direction. That the play still failed on optimistic pricing is what makes the verdict safe.

What outlives the play

Carried forward

• A threshold change may not authorise a trade in a name that motivated the change, and never on the same day.
• Pre-register before entry or do not run it. Later learning is appended, never merged upstream.
• A pooled statistic is never applied to an individual symbol without the within-symbol column printed beside it.
• Every precondition documented in a playbook must be enforced in code, or it will eventually be skipped.
• Exchange time comes from a timezone library, never from a fixed UTC offset.
• A profitable outcome does not retire a recorded process objection.
• A result selected as the best of N is tested against the distribution of best-of-N, never against zero.
• An effect that cannot be estimated from prior data alone is not tradeable, however real it is.