Pre-registered 2026-10-23 · 2 runs · data snapshot 2026-09-19T090800Z
The idea
Retail investors buy what catches their eye, and they cannot sell what they do not own. Barber and Odean (2008) showed that this asymmetry makes attention-grabbing stocks net-bought by individuals. If the buying overshoots, a stock trading at unusually high volume should underperform once the attention fades. The counterparty is the attention buyer.
The lab had already measured the long side: buying high relative volume lost money, t −8.58. The short was its mirror, so some alpha was expected by construction. What was unknown was whether a stop, borrow costs and commissions would leave any of it.
How it was tested
Two horizons. A 20-session version was pre-registered on 2026-10-14 (commit d36a1c8). After borrow killed it, a 2-session version was pre-registered on 2026-10-23 and declared expected to fail, since the horizon was chosen after seeing the first result; the family threshold counted the earlier runs (k = 4, t ≥ 2.24). Signal: relative volume above 1.3 times the 20-day average, price at least $5, $3M of daily dollar volume. Short at the next open, out after 2 sessions or at a 3×ATR stop, up to 20 new names a day ranked by relative volume. Borrow at 20%/yr, 5 bp slippage per side and per-share commission charged trade by trade.
Where it died
The first 2-session run gave no verdict: the negative control failed, because the archived number it was compared against was mislabeled. Rerun the same day, the strategy passed four gates in a row: Gate 7 at t 4.79 net of borrow (5.46 without two trades above 300%; the worse governs), Gate 8, both halves (0.75 and 5.06), and own alpha, with only 6.9% of its trades overlapping the lab's surviving short.
It stopped at Gate F. Mean alpha of 0.236% per trade is below the 0.30% floor, and at the small ticket tested each trade lost $0.60 after costs. The 20-session version had died the other way round: t 7.09 with free borrow, −1.66 at 20%/yr.
The record
Thesis
Retail investors buy what grabs their attention (abnormal volume, news, gaps) and do not sell symmetrically, so attention-grabbing stocks end up overbought and underperform afterwards. Short the names with unusually high relative volume and let the overbuying revert.
Prior: Expected to fail. The 20-session version had been strongly negative in 2014-2019 (t −7.13 at 20%/yr borrow), and the 2-session horizon was chosen after seeing that borrow killed the 20-session version — a derived path, so the family threshold was raised to count the earlier runs (k = 4).
Barber, B. & Odean, T. (2008). All That Glitters: The Effect of Attention and News on the Buying Behavior of Individual and Institutional Investors. Review of Financial Studies 21(2).
Specification
signal
relative volume (vs 20-day average) > 1.3
universe
US stocks, price ≥ $5, 20d dollar volume ≥ $3M
entry
next open, short
exit
2 sessions or 3×ATR stop, no target
ranking
relative volume, highest first, max 20 new positions/day
costs
borrow 20%/yr, 5 bp slippage per side, tiered per-share commission computed trade by trade
caveat
Mirror image of a long already measured in the lab (relative-volume continuation, t −8.58). A positive Gate 7 for the short was therefore partly expected by construction; the stop, the borrow cost, the halves and the economics are what had not been seen.
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Gate 7 — Jensen alphapass
After regressing out market exposure, is there alpha left?
t ≥ 2.24 (Bonferroni, k = 4 including the 20-session runs), net of 20%/yr borrow
t (HAC)+4.79alpha / trade %+0.24beta+0.80trades50,438
Two trades above 300% made the plausibility check flag the run; without them t is +5.46. The worse of the two governs.
Gate 8 — Capture-preserved alphapass
Does the subset of trades a real portfolio can actually take keep the alpha?
t_cap ≥ 0.65 (scaled) and ratio ≥ 0.50
t captured+5.49capture ratio+4.26
Gate F — Frictionfail
Is the alpha per trade large enough to pay for its own execution?
alpha/trade ≥ 0.30% and net $/trade > 0 at the declared small-account sizing
alpha / trade %+0.24net usd per trade-0.60
Stabilitypass
Is the alpha present in both halves of history?
t > 0 in both halves
t 2014 2019+0.75t 2020 2026+5.06
Own alphapass
Is the alpha new, or already captured by a strategy that passed?
≤ 30% of trades overlap the lab's surviving short
overlap pct+6.90
Why it died
At 2 sessions the alpha exists: t +4.79, positive in both halves, and still t +2.75 at 50%/yr borrow. It dies on economics. Mean alpha is 0.236% per trade, against roughly 0.41% of round-trip commission at the small ticket size tested, plus slippage — net −$0.60 per trade. The 20-session version had died on borrow instead: t +7.09 at 0% borrow, −1.66 at 20%/yr, because about 28 nights of borrow cost ~1.5% per trade, the size of the whole gross alpha.
Open questions
Gate 8 says the 3 highest-ranked names per day carry 4.26× the mean alpha, and the portfolio simulation with the same costs returns +16.9%/yr. That is the most extreme "meme" tail, where real borrow is probably far above 20%/yr and locates fail. No unseen history is left to test it; only a forward test with real locates and real borrow can.
What it taught
Holding period was the lever that decided whether borrow or commission ate the edge, and each horizon lost to one of them. It also exposed a design flaw in the criteria: Gate F and the economics test measure the mean of the 20-a-day population, while a 3-slot portfolio would trade only the top of the ranking. That flaw was declared after the result and does not rescue the verdict.
Verdict history
2026-10-1420-session version, pre-registered (commit d36a1c8): t −1.66 at 20%/yr borrow, +7.09 at 0%; 2014-2019 t −7.13 Eaten by friction
2026-10-232-session version, first run: no verdict — the negative control failed, candidates not run (the control's archived number turned out to be mislabeled)
2026-10-232-session version, second run: Gate 7, Gate 8 and stability pass; alpha 0.236%/trade fails Gate F and net $/trade Eaten by friction