Every 0DTE dashboard quotes an "expected move" - the one-sigma range to the close implied by ATM volatility. Folklore says the straddle overstates the move most days. As far as we can find, nobody has published the number: when the screen says +/-X points to the close, how often does the close land inside?
We measured it on 82,378 five-minute readings over 1,088 SPX sessions (2022-04-14 to 2026-08-26). At every frame: ATM IV interpolated at spot, EM = S * sigma * sqrt(tau to 16:00), then the fact - did the close land within EM of that reading's spot? Shares are day-clustered; readings with under 15 minutes left are excluded from the headline.
Conservative by four points, everywhere
| threshold | theory | measured |
|---|---|---|
| inside 1.0 x EM | 68.3% | 72.4% +/- 0.9 |
| inside 0.8 x EM (straddle breakeven) | 57.5% | 62.0% +/- 0.9 |
| inside 0.5 x EM | 38.3% | 42.4% +/- 0.8 |
The folklore is right in sign and modest in size: the quote overstates the delivered close-move by ~4 percentage points of coverage. The four points sit at every threshold - a level of implied vol, not a fat-tail premium. And it is remarkably uniform: coverage runs 71-74% at every hour of the session and in every year 2022-2026. Nothing has drifted as 0DTE volume grew.
One observation the naive story does not predict
Cut by the dealer-gamma state at that minute and the gradient runs backwards from intuition: coverage is HIGHER in the short-gamma tail (74.5% at gamma percentile <= 20 vs 71.0% at >= 80). Implied vol prices short-gamma days even more conservatively than the extra movement they deliver to the close. Two to three standard errors - an observation to check, not a result to lean on.
Caveats named in full (minute closes, close-calibration not range-calibration, overnight excluded) plus a complete reproduce recipe from the public per-session files - per-strike IVs and spot are all you need, no paid data: https://firmtape.com/research/how-accurate-is-the-expected-move
All finished sessions replay free at https://firmtape.com/sessions
Update: the variance test (2026-08-28). Coverage asks how often the close lands inside; it does not ask whether the quote is the right SIZE. We added the standard checks from the vol-forecasting literature. Result: E[z^2] = 1.034 +/- 0.062 - the expected move is an essentially unbiased VARIANCE forecast, roughly uniform across the day - and under the QLIKE loss it decisively beats a trailing-realized-vol benchmark (1.50 vs 2.05, t -4.2). So "conservative by four points of coverage" is not overpricing: the variance is priced right, and more closes land inside +/-1 EM than the lognormal predicts anyway. Both can be true only if the move distribution is leptokurtic versus the lognormal - more mass in the middle AND fatter far tails at the same variance. The sharper one-liner: right variance, wrong shape. Sizing against the expected move as a one-sigma is honest; treating the 28% outside as lognormal tails is not. Full numbers in the addendum on the study page above.
