Short Interest / Days-to-Cover Deciles (probe): a killed candidate

Verdict: KILLED Stage: screen prototype
Identity: equity_short_interest_dtc

The first genuinely NEW INPUT in months rather than another transformation of price and volume: bi-monthly FINRA short interest for the whole US tape, 2017-12-29..2026-07-15 (3,804,024 rows, 23,314 tickers), unlocked by the Polygon Starter upgrade and ingested in full. ONE pre-registered config, no sweep: long the bottom days-to-cover decile, short the top, dollar-neutral, PIT top-2000 by ADV, price >= $5, positions formed on settlement + 10 BUSINESS DAYS (FINRA disseminates ~8; the extra margin makes lookahead un-arguable) and held to the next availability date. *** CORRECTION 2026-08-05 — THE FIRST PUBLISHED NUMBERS FOR THIS ENTRY WERE WRONG. *** The probe computed returns as log(close).diff() on data/lake, which stores RAW AS-TRADED closes: not split- or dividend-adjusted. Every forward split therefore booked a catastrophic fake loss — AAPL 2020-08-31 as -135%, NVDA 2024-06-10 as -229%, TSLA 2022-08-25 as -110%. The bias was DIRECTIONAL, not noise: splits happen in high-priced mega-caps, mega-caps have enormous ADV, and days_to_cover = SI/ADV is therefore tiny for them, so every fake collapse sorted into the LONG (bottom-DTC) leg. The production feature engine refuses raw prices outright; this standalone probe bypassed that guard. Fixed in scripts/lib/px_adjust.py, which adjusts the RETURN series only and deliberately leaves price floors and ADV ranks on raw as-traded values (back-adjusting levels would introduce a look-ahead the raw panel does not have). Both affected probes were repaired in the same pass. WHAT CHANGED: gross Sharpe before any cost was published as -0.220 (full) and +0.002 (easy-to-borrow); corrected it is -0.062 (NW t -0.19) and +0.204 (NW t +0.66). The published claim that there was 'no edge for frictions to eat' was therefore FALSE — there is a small gross edge, it is simply not statistically established and it is entirely consumed by costs. The KILL STANDS but on different reasoning. Corrected net of 6bp one-way at 20x annual turnover, on the easy-to-borrow universe that is actually shortable: -0.026 / -0.228 / -0.792 at 50 / 300 / 1000 bp per year of borrow. The top days-to-cover decile IS the hard-to-borrow bucket, so the 50bp row describes a trade nobody can put on; at a realistic 300bp it is -0.228. Against the pre-registered rule it fails gate (b) Newey-West t >= 2 and gate (d) survives 300bp on easy-to-borrow, and passes (a) and (c). Correlation to the live book is -0.017, so the diversification arithmetic was favourable and it still did not matter — the same lesson as mechflow_tom: a near-zero bar lowers what a candidate must BEAT, never what it must PROVE. The sign is NOT flipped and re-tested, and the gates were NOT relaxed to admit it once the corrected numbers looked better; that is precisely the search a pre-registration exists to forbid. KILLED. Reproduce: scripts/ingest_short_interest.py then scripts/probe_short_interest.py.

Why it was worth testing

This died at the screen stage, before a full walk-forward was ever run. Screening exists so that ideas which cannot clear a coarse, cost-aware bar do not consume the far more expensive machinery behind it. A screen kill is a cheap kill, and it is published for the same reason as an expensive one: the trial was still spent, and it still raises the evidence bar for everything already in the book.

The result

Measure	Value
Screen net Sharpe	0.2040

What this does and does not say

It says this configuration, on this data, net of the costs we charge, did not clear the bar it pre-registered. It does not say the underlying economic effect does not exist, that no implementation of it works, or that someone with different data or different execution would reach the same conclusion. A null is evidence about a test, not a proof about a market.

It also does not say the trial was free. Every hypothesis tested raises the deflated-Sharpe hurdle for every sleeve already in the book, including the ones that survived. That is why the kill count is published beside the survivor count rather than behind it.