Event studies of dealer hedge flow usually suffer from a convenient disease: the narrative gets written after the outcome is known. Point-in-time replay removes the excuse. Before pulling a single stamp we fixed three questions: (Q1) did signed hedge-flow impulse lead 15- and 30-minute forward SPY returns on August 2 and August 5? (Q2) did pin risk release before the August 5 opening gap? (Q3) did flow-implied GEX flip negative earlier in the session than settled-OI GEX? Then we collected.
553
5-minute stamps, 7 sessions, 6 endpoints each
−3.90%
The Friday-close to Monday-open gap
37
Pin-risk score at Friday's close - no alarm
7/7
Sessions short-gamma from the 09:30 open
Method
Sessions 2024-08-01 through 2024-08-09; a 5-minute grid from 09:30 to 16:00 ET (79 stamps per session); per stamp we replayed /v1/flow/options/SPY/summary, /v1/flow/levels/SPY, /v1/flow/pin-risk/SPY, /v1/flow/gex/SPY, /v1/flow/dealer-risk/SPY, and /v1/exposure/summary/SPY via the Historical API. Every scalar field lands in minute_grid.csv (553 rows x 68 columns); derived analyses in hedge_flow_vs_returns.csv and pin_risk_track.csv; the collection script ships in artifacts.zip. The frozen at=2024-08-05T10:30:00 fixtures in our OpenAPI spec come from the same tape, so any single number is independently checkable.
Q1: Did hedge-flow impulse lead returns? Weakly, and only when it mattered least to know
Signal: the 15-minute change in the flow-implied dealer delta adjustment (flow_dex_adjustment from the dealer-risk endpoint), and the same construction on net tape volume. Outcome: forward SPY return 15 and 30 minutes ahead, within-session only.
| Scope | Signal (15-min change) | vs fwd 15-min ret | vs fwd 30-min ret |
|---|
| Aug 2 (first leg down) | flow DEX adjustment | -0.14 | -0.18 |
| Aug 2 | net tape volume | -0.12 | -0.20 |
| Aug 5 (the unwind) | flow DEX adjustment | -0.19 | -0.07 |
| Aug 5 | net tape volume | +0.13 | +0.03 |
| All 7 sessions pooled | flow DEX adjustment | -0.06 | -0.04 |
| All 7 sessions pooled | net tape volume | +0.02 | -0.10 |
The pre-registered answer is mostly no. On the two event days there is a weak contrarian lead - rising flow-implied dealer delta preceded slightly negative 15-30 minute returns, consistent with hedging demand peaking near local extremes - but at |r| of 0.1-0.2 on ~70 observations per day, this is a lean, not a signal. Pooled across the week it is indistinguishable from zero. If you have seen event-day charts implying dealer flow "called" the carry unwind minute by minute, this is what the same data says when the questions are fixed in advance.
Q2: Did pin risk release before the gap? No - the score never flashed
The cleanest reading of "pin release predicts the break" would be a high pin score on Friday collapsing into the close. The replay says otherwise: at Friday August 2's final stamp the pin-risk score stood at 37 of 100 (session range 22-62), spot 533.01. Monday reopened at 512.23 - the -3.90% gap - and the first Monday stamp scored 55 with the magnet re-anchored at 520. The full 90-minute tracks around the gap are in pin_risk_track.csv. Whatever information the weekend carried, the Friday pin state did not encode it - which is the expected result for a weekend macro shock originating in the yen, and worth having on the record precisely because it is unexciting.
Q3: Did flow GEX flip before settled GEX? The question dissolved on contact
We expected to time two intraday sign flips. Instead: settled net GEX opened negative on all seven sessions (ranging -8.0B to -14.6B across the week) and never crossed zero intraday; the flow-adjusted version agreed on sign at every single stamp. The regime work was done overnight, in the OI print, before any 09:30 tape - the intraday story of the carry unwind is amplitude inside an already short-gamma regime, not a regime change. The flow adjustment moved the magnitude (the flow_gex_pct_shift column), and that column is where any future version of Q3 should look.
What we found that we did not ask
In 343 of 553 stamps (62%), the pin-risk magnet strike sat more than 10% from spot - e.g. a 450 magnet with SPY at 545, or 350 at the Friday close. This is the full-chain wall convention doing what it is defined to do: the magnet tracks the largest OI concentration chain-wide, and in this window that sat in far-dated round-strike LEAPS piles rather than near the money. It is a convention to be aware of, not a bug in the score - the score's proximity and gamma components behaved sensibly throughout - but if you consume the magnet strike directly in stress windows, filter it by distance-to-spot. We publish that observation here for the same reason the coverage page publishes gap counts: you would find it anyway, and finding it here first is the point of replayable data.
Limitations
- One event week, one symbol, a 5-minute grid. Q1's within-day correlations have ~70 observations per session; treat the signs as descriptive.
- Pre-registration constrains us, not the data: with different signal constructions (score-weighted, 0DTE-excluded, OI-confirmed) the lead-lag answer could differ - each is one more loop over the same replay, which is rather the point.
- Standard replay semantics apply: start-of-day OI (the no-lookahead choice), and the year-boundary
official_oi defect does not touch this window.
Replaying the most dramatic options week of 2024 at 5-minute resolution produced two nulls, one structural fact, and one convention warning - which is exactly what pre-registered event studies are supposed to produce when the narrative does not survive contact with the tape. The grid, the tracks, and the collection script are downloadable above and rerun in about twenty minutes of API time against Alpha-tier replay. Related: the dealer-positioning narrative replay of the same day, the flow replay reference, the March 2020 replay, and live vs settled GEX explained.
