What does a level hold rate measure?
Call wall, put wall and flip in three books, against a random line at the same distance
Key finding. Gamma-level products publish hold rates in the high eighties and nineties. A line one expected move from price holds about that often with no options data behind it. Scored against a random line at the same distance, the call wall, the put wall and the zero-gamma flip of the volume, open-interest and measured books add nothing: the edge is within a standard error of zero in every one of nine rows.
Gamma-level products increasingly publish a backtest beside their levels: a hold rate, a break-during-the-day rate, a break-at-the-close rate, often in the high eighties or the nineties. One public example is MenthorQ's Backtesting Results: 1D Move on the SPX Index. Over November 2019 to August 2023 it reports the next day's SPX close below its 1D Max 85.02% of the time, above its 1D Min 87.62%, and inside both 72.63%. The page gives no baseline.
A hold rate without a baseline cannot tell a level from its distance. A line far enough from price holds most days because price rarely travels that far, whatever put the line there. The question worth asking of any level is not how often it held, but how often it held more than a line drawn at the same distance for no reason.
What one expected move does on its own
If the day's move were normal and the level sat one standard deviation away, the close would stay below the upper line 84.1% of the time, above the lower one 84.1%, and between them 68.3%. The SPX market does a little better than normal at the centre: over 1,088 sessions the close landed inside one straddle-implied expected move 72.4% of the time, because the distribution of the close is peaked and fat-tailed rather than normal. Those are the numbers a line one expected move away earns by being there. 85%, 88% and 73% are what that page reports for levels built from the expected move over a different window and a close-to-close horizon; the resemblance is the point. Nothing in a result of that size says the options market added anything beyond the width of the band.
The test: each level against a line at its own distance
Each level is read once, at 10:00 ET, and judged against the rest of the session. For every session the distance from spot to the level is expressed in that session's expected move, and the baseline is how often any line that far away, on that side, was crossed across all 1,099 sessions. "Edge" is the level's hold rate minus that baseline, with a day-bootstrap standard error.
| level (at 10:00 ET) | sessions | held at close | random line | edge | held all day | random line | edge |
|---|---|---|---|---|---|---|---|
| call wall, volume book | 1,098 | 76.2% | 76.6% | −0.4 pp (±1.3) | 58.4% | 58.9% | −0.5 pp (±1.4) |
| call wall, open interest | 1,077 | 76.2% | 78.1% | −1.9 pp (±1.2) | 61.6% | 60.8% | +0.7 pp (±1.3) |
| call wall, measured book | 1,077 | 75.2% | 76.4% | −1.2 pp (±1.3) | 56.6% | 56.9% | −0.3 pp (±1.3) |
| put wall, volume book | 1,092 | 79.5% | 79.4% | +0.1 pp (±1.2) | 60.6% | 59.8% | +0.9 pp (±1.4) |
| put wall, open interest | 1,080 | 81.8% | 81.9% | −0.2 pp (±1.1) | 66.0% | 65.2% | +0.8 pp (±1.2) |
| put wall, measured book | 1,077 | 81.2% | 79.6% | +1.6 pp (±1.1) | 64.7% | 62.3% | +2.4 pp (±1.2) |
| flip, volume book | 1,044 | 63.7% | 64.0% | −0.3 pp (±1.5) | 30.1% | 31.8% | −1.7 pp (±1.3) |
| flip, open interest | 1,077 | 79.3% | 79.4% | −0.1 pp (±1.1) | 62.1% | 61.6% | +0.5 pp (±1.2) |
| flip, measured book | 1,062 | 93.6% | 94.1% | −0.5 pp (±0.7) | 89.0% | 89.0% | 0.0 pp (±0.9) |
| spot + 1 expected move | 1,099 | 92.0% | 92.0% | 0 by construction | 87.6% | 87.6% | 0 |
| spot − 1 expected move | 1,099 | 91.1% | 91.1% | 0 by construction | 83.8% | 83.8% | 0 |
Read the last two rows first. A line one expected move above or below price at 10:00, drawn with no book at all, held at the close on 92% and 91% of sessions. That is the size of number a hold-rate panel prints, and it is earned entirely by distance.
Then the nine book levels. Every wall and every flip holds within about one standard error of a random line at its own distance. The measured book's flip holds 93.6% of the time, which sounds like a strong level until the baseline says a line that far away holds 94.1%. The one cell that leans positive, the measured book's put wall at +2.4 ± 1.2 points during the day, is two standard errors from zero in a table of eighteen tests, which is what chance produces about once in a table this size.
A holdout that looks like an edge and is not
Scored on the last 250 sessions alone, the call walls show +4 to +6 points over the baseline. So does the plain line one expected move above spot: +4.0 ± 1.2 over the same sessions. A level with no book behind it cannot have an edge, so the lift belongs to the period, not to the walls: recent sessions travelled less far than the full sample's baseline expects, and every line held a little more often. A holdout needs its own control, or it measures the calendar.
What to ask of any hold rate
Four questions settle whether a published hold rate says anything about the level:
- The baseline. How often did a line at the same distance, on the same side, hold over the same sessions? Without it the number is a statement about distance.
- The distance. How far was the level from price when it was read, in expected moves, not points? A level two expected moves away that holds 95% of the time is doing nothing.
- The read time. Was the level fixed before the outcome it is scored on? A level that moves through the day and is scored against its own later values already knew the answer.
- The error. Is the interval clustered by day? Minutes inside one session are not independent trials, and a count of "positive outcomes" without an interval cannot be told from luck.
Our own levels answer all four on this page and on the earlier hold-rate study of the flip, the hold band and the prior day's range, where the same baseline left one borderline survivor. We draw the walls and the flip because they describe where the dealer book sits, not because they hold better than distance, and we do not print a hold rate beside them.
Method
Source: the published per-minute feature store, SPX, 2022-04-14 to 2026-09-23, 1,099 sessions
with a 10:00 ET row and at least 300 minutes after it. The level is spot minus the stored signed
distance for each book (call wall, put wall, zero-gamma flip), read at 10:00. A level on the
wrong side of spot, or within 0.02 expected moves of it, is dropped for that session. "Held at
close" means the last print is on the level's side; "held all day" means no minute close after
10:00 traded through it. The expected move is the stored expected range at 10:00. The baseline
for a level at distance z above spot is the share of all sessions whose high after 10:00 (for the
all-day test) or whose close (for the close test) rose at least z expected moves, and the mirror
for levels below. Standard errors are a bootstrap over sessions. The holdout is the last 250
sessions. The walls and flip are the terminal's own, the same values drawn on the chart and
available in LAB as cr0_dist, ps0_dist, flip_dist and their
volume and open-interest twins.
Part of FirmTape research. Measured on the free session archive; every session is free to replay.