Pre-registered 2026-10-09 · 1 run · data snapshot 2026-09-19T090800Z
The idea
A day of unusually heavy trading often means something real happened: institutions building a position, media attention, an index inclusion. If that information is absorbed gradually, the stocks with the highest volume relative to their own average should keep rising for weeks.
The idea came out of a failure. An earlier volume-reversal short had died, and its post-mortem concluded that high volume precedes rallies, not reversals. A corrected information-coefficient scan agreed, at t +7.6 to +11.6 out of sample. Rather than treat that as a finding, it was tested as its own hypothesis.
How it was tested
Pre-registered on 2026-10-09 as a clean factor test: 20-day relative volume above 1.3, nothing else. US stocks at $5 or more with at least $3M of 20-day dollar volume. Buy at the next open, ranked by relative volume, hold 20 sessions with no stop and no target. Neutral sizing: $5,000, 1% risk per trade, 10 slots. The bar was Gate 7, Jensen alpha t of at least 1.65.
Where it died
The first run, on 2026-10-10, was voided: 15 of 45,388 trades returned more than +300% and tripped the lab's plausibility guard. A data filter built on a false premise was added and then withdrawn. Checked by hand, those series were real meme-rally moves, not corrupt data.
The diagnostic run on 2026-10-14 used a rule committed before running: report the result with and without the outliers. Both failed, with the same sign: t −8.58 with them and −10.15 without. Alpha was −1.39% per trade over 20 sessions, with beta 0.72. Gate 8 was not evaluated.
The record
Thesis
Unusually high relative volume accompanies real new information — institutional buying, media attention, index inclusion — which is absorbed gradually, so high-volume names keep rising over the next 10-30 sessions instead of reverting.
Prior: An earlier volume-reversal short had failed, and its post-mortem concluded that high volume precedes rallies. The corrected information-coefficient scan agreed (t +7.6 to +11.6 out of sample). This tests that reading as its own hypothesis.
Blume, L. & Easley, D. (1983). Trading volume and returns.
Karpoff, J. (1987). The Relation Between Price Changes and Trading Volume. Journal of Financial and Quantitative Analysis.
Specification
signal
20-day relative volume > 1.3 — nothing else
universe
US stocks, price ≥ $5, 20d dollar volume ≥ $3M
entry
next open
exit
20 sessions, no stop or target (clean factor test)
ranking
highest relative volume first
sizing
neutral default — $5,000, 1% risk per trade, 10 slots
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Gate 7 — Jensen alphafail
After regressing out market exposure, is there alpha left?
t ≥ 1.65
t (HAC)-8.58alpha / trade %-1.39beta+0.72trades45,388
Without the 15 trades above +300% (real meme-rally moves, checked by hand) t is −10.148. Both fail with the same sign.
Gate 8 — Capture-preserved alphanot evaluated
Does the subset of trades a real portfolio can actually take keep the alpha?
Not evaluated once Gate 7 fails.
Why it died
A strong anti-edge: −1.39% alpha per trade over 20 sessions, t −8.58 on 45,388 trades. The strategy buys the top of the relative-volume ranking each day — names in an attention spike — and those underperform afterwards.
What it taught
Information coefficient +11 → alpha t −8.6. A factor can order the broad universe one way while its extreme tail behaves the opposite way. The first run was voided by the lab's 300% plausibility guard; checking the offending series by hand showed real explosions, not corruption — the guard gives false alarms on real moves, and a void needs a manual check before anything is "cleaned". The reverse (short) side was not pursued from here: flipping a sign after seeing the result is a new hypothesis, not a fix.
Verdict history
2026-10-09Pre-registered, sections 0-2 locked
2026-10-10First run VOID: 15 of 45,388 trades above +300%. An extra data filter based on a false premise was added and then withdrawn
2026-10-14Diagnostic with a rule committed before running: fails with and without the outliers (t −8.58 / −10.15) No alpha