Options premium bursts drift in the burst direction over the next 15 minutes

The opening-window family fixed the clock and asked whether the tape predicts the day. This asks the question practitioners actually pose when they watch a feed: when an unusually large slug of options premium hits at some arbitrary moment, does price follow it over the next quarter hour? A yes would have meant that large aggressive prints leak information as they are absorbed, and that the leak is visible and exploitable in real time rather than only at a fixed hour of the morning.

Tests whether EXTREME aggregate premium bursts (options-net-premium, all trade types pooled, dollars) predict same-direction drift over the following 15 minutes, at any point in the 10:00-15:29 session body; a variable-event-timing claim, unlike the fixed opening-window family (HYP-001/002/003, all NULL). A null does not falsify trade-level burst definitions (sweeps/blocks), aggressor- partitioned bursts, or other horizons. Signal-side census disclosed above; no outcome data was examined before the freeze.

Nine liquid names (two index proxies and seven mega-caps) over seven months of sessions. A burst is a one-minute bucket whose absolute net options premium reaches eight times the running mean absolute premium of that same session up to the previous bucket, so the baseline is causal and adapts to the day rather than being imported from history. Bursts are eligible between 10:00 and 15:29, they need at least thirty prior buckets of baseline, and an accepted event blocks further events for the length of the outcome window, so the sample carries no overlapping observations. The threshold was chosen from a signal-only census (five, eight and twelve times the baseline yield roughly 6,000, 3,300 and 2,000 events) and eight was preregistered as the extreme level before any outcome was computed. The outcome is the fifteen-minute log return beginning one full bucket after the event bucket closes, signed by the burst direction, so a positive mean means price continued the way the premium leaned. The permutation null flips the sign of every event in a session together, which respects the fact that events cluster within days. 999 draws, preregistered positive direction.

Nothing follows the burst. The mean signed post-event return was 0.32 basis points with t = 0.69 and a permutation p of 0.51 over 3,341 events in 140 sessions. The point estimate is correctly signed and economically negligible, and 3,341 events is a sample large enough that a drift worth trading would have been visible. This is the strongest of the directional nulls in one specific sense: it frees the test from the fixed opening window, so the family's earlier failures cannot be blamed on having looked at the wrong time of day. The lab also asked the companion magnitude question of the same events, whether bursts predict how much price moves rather than which way, and that came back null too, which retired the burst representation in both outcome spaces at once.

Formation windows close roughly a minute before outcome windows open, and measured feed publication latency consumes most of that slack. Live availability of each signal at the formation boundary is therefore tight rather than generous.

A burst here is a burst in aggregated premium, with all trade types pooled and no view of the individual prints beneath it. The registered claim is therefore about extreme aggregate premium, and a null does not falsify trade-level burst definitions such as sweeps or blocks, aggressor-partitioned bursts, or horizons other than fifteen minutes. Twenty candidate events dropped during construction where an outcome window ran past the data available for it, and those drops are reflected in the reported count rather than filled in.

Like every result the lab has published so far, it is a discovery-substrate result: preregistered, but not yet confirmed on data the research process has never seen.