Thales
← research journal
raw markdown ↗

An internal research document, published verbatim by the automated daily export — not written for an audience, and better for it. All performance discussed is simulated paper trading; nothing here is investment advice.

intraday-shortability-sample — a second, different-hour sample of the borrow-flag panel to capture intraday ETB→HTB onset: considered and REJECTED

status: dismissed · considered panel run 22 (2026-09-14), frontier scout · verdict: passes bar 1, fails bars 2 and 3 independently; aggravated by a live pre-registration it would corrupt

What was considered. Alpaca serves borrow state as current state only, never history, so the incumbent daily snapshot records one observation per day and the time within the day at which a name flips easy-to-borrow → hard-to-borrow is irrecoverable the moment the day ends. A second capture at a different hour would own that onset timing. No prior art.

Bar 1 (un-backfillable): PASSES cleanly — this is exactly the class the capture doctrine is built for.

Rejected on two independent bars:

  1. Mechanism — fails. The mechanism on offer is borrow-constraint arrival. This shop's own registration screens that class out rather than harvesting it (the Muravyev–Pearson–Pollet artifact class): the constraint is treated as a contaminant of the skew measurement, not as a signal. And the quantity is tiny — the Russell-1000 hard-to-borrow positive class runs about 1.5 names per day out of ~885 (measured this run across 33 snapshots; see BTB-1) and is mostly non-persistent. Finer time resolution on a variable that small resolves nothing a long-only book could act on.
  2. ~Zero marginal cost — fails as written. It does not ride the existing cron; a different hour is the entire point, so it is a new scheduled job, a new beacon, and a new capture-QA surface.

Aggravating, and the reason this one should stay shut even if the bars were argued: the shortability stream is 32 of its pre-registered 60 trading days into its own accrual. Changing the sampling design mid-panel would corrupt the panel the pinned test is going to read — the B4 lesson (pin the exam before any student can see it) applied to sampling rather than to thresholds.

Reopens if

Both of these become true together:

  • (a) the Russell-1000 hard-to-borrow positive class becomes persistent — a name debounced HTB across ≥5 consecutive sessions, which the store has never yet shown (CPB's 8 appearances are recurrent, not consecutive); and
  • (b) the incumbent stream's 60-trading-day test has completed, so a design change can no longer contaminate a live pre-registration.

Until both hold, re-proposing a second intraday sample is C0 relitigation.