10:00-11:30 ET, disjoint frozen sub-windows
Repeated later-session prints were episodes, not independent signals
Thousands of rows collapsed to a handful of daily leaders, and the best-looking cell still lost to controls.
In simple words
What happened
The big dataset was mostly repeated observations of the same moments. Once those duplicates stopped counting as fresh evidence, there was no robust directional edge.
Who would pay?
A truly informed repeat buyer should keep paying because the underlying move has not yet incorporated their information. The signature must predict the named direction after daily episode collapse.
How it was tested
Large-lot repeats, bursts, and sequences were preregistered as L1-L4 families. Correlated rows collapsed to one episode leader before inference; the named and opposite options were both graded.
What would prove it wrong?
The strongest apparent subgroup had only four leaders, 2024 was -50%, raw-event returns were negative, and the opposite option was stronger. No feature cleared minimum sample or control bars.
Leader-collapsed evidence
What each setup actually did
Discovery strip A; the best 10:30 cell had only four daily leaders
| Setup | Leaders | Mean | Median | Hit rate | Controls | Decision |
|---|---|---|---|---|---|---|
| L2Best 10:30 repeat/burst cell | 4 daily leaders | Negative raw events | Unstable | Insufficient n | Opposite right stronger; 2024 -50% | REFUSED |
What we keep
Learnings
- 01
Ten prints in one burst are not ten independent bets. Independence lives at the episode or session level.
- 02
When both the call and put can win around the same burst, the event may predict volatility—not direction.
- 03
A feature with only four daily leaders cannot support a production rule, however dramatic the raw tape looks.