No alphaArchived · long · 2026 (3-month regime window)
Pre-registered 2026-09-13 · 1 run
The idea
In the rotating market of mid-2026, stocks trading close to their 52-week low rebounded over the following 20 sessions. The pre-registration called it what it was: a regime artifact, with no counterparty forced to trade badly.
What made it worth a run was the scan. Its information coefficient against 20-day forward returns was −0.1149, t −8.97 over 42 dates, stronger than RSI in the same window. And it was almost uncorrelated with the deployed RSI-bounce signal, at 0.021: a candidate second leg rather than another copy of the first.
How it was tested
Pre-registered on 2026-09-13, with those scan statistics disclosed before the lock. Signal: price within 10% of its 52-week low, a threshold taken from the pooled 10th percentile of the discovery window rather than chosen on profit. Universe: US stocks above $5 with more than $3M of daily dollar volume. Buy at the next open, hold 20 sessions to match the scan horizon, with a 2×ATR stop and a 2.5×ATR target, closest to the low first, at a neutral $5,000 and 10 slots.
Where it died
The single run used the 3-month regime window: Gate 7 t +0.69 on 247 trades, Gate 8 t_captured 0.78. Archived, correctly as it turned out, but on far too little data to tell an edge from noise.
The method audit re-ran it on full history on 2026-10-30. Gate 7 fell to t −2.53 over 13,212 trades: −0.514% per trade with a beta of 1.037, negative in both halves. Gate 8 gave t_captured −1.68 and a capture ratio of 0.0.
The record
Thesis
In the mid-2026 rotating market, stocks trading close to their 52-week low rebound over the next 20 sessions. Declared as a regime artefact, not a permanent mechanism: there is no named counterparty forced to trade badly.
Prior: Disclosed before the lock: in the discovery scan the factor's information coefficient against 20-day forward returns was −0.1149 (t −8.97 over 42 dates), stronger than RSI in the same window, and it was almost uncorrelated with the lab's RSI-bounce signal (0.021).
Specification
signal
price within 10% of its 52-week low (threshold = pooled 10th percentile of the discovery window, not chosen on P&L)
universe
US stocks, price > $5, 20d dollar volume > $3M
entry
next open
exit
20 sessions (the horizon of the discovery scan), 2×ATR stop, 2.5×ATR target
ranking
closest to the 52-week low first
sizing
neutral default — $5,000, 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 ≥ 1.96 (Bonferroni, 2 repairs), full history
t (calendar SE)-2.53alpha / trade %-0.51beta+1.04trades13,212
Re-measured 2026-10-30 on full history after the method audit. The original verdict used only a 3-month regime window (t +0.69 on 247 trades), which has too little power to decide anything.
Gate 8 — Capture-preserved alphafail
Does the subset of trades a real portfolio can actually take keep the alpha?
t_cap ≥ 1.65 and ratio ≥ 0.50
t captured-1.68capture ratio+0.00
Why it died
No alpha on full history: t −2.53 over 13,212 trades, negative in both halves. The original archive was right, but it was decided on a 3-month window that could not have told a real edge from noise.
What it taught
A strong information coefficient in a scan did not become a tradeable alpha once the strategy bought only the tail. The same disconnect later killed two more scan-derived candidates, and a month later the lab measured that none of its cheap triage tools predicts Gate 7.
Verdict history
2026-09-13Pre-registered with the scan statistics disclosed
2026-09-13Single run: Gate 7 t +0.69, Gate 8 t_captured 0.78 No alpha
2026-10-30Re-measured on full history after the method audit: t −2.53, negative in both halves No alpha