UnstableArchived · long · 2026 (3-month regime window)
Pre-registered 2026-09-05 · 1 run · 6 variants
The idea
The deployed RSI-bounce signal held every trade for 5 sessions, a number chosen without evidence. If the edge is the correction of a price dislocation caused by indiscriminate selling, the dislocation takes a certain time to close, and alpha per trade should peak at some interior holding period rather than keep rising with the hold.
The origin was declared contaminated before the run: the idea came from a control arm of another study, where holds of 5-10 sessions had looked better. A pass could only have authorized a forward test, never a change to the live configuration.
How it was tested
Pre-registered on 2026-09-05. The signal stayed exactly as deployed: RSI(14) below 40, price within 10% of the 50-day average, at least $3M of daily dollar volume. Only the hold varied: 3, 5, 8, 10, 15 or 20 sessions, with 5 as the control. Six variants raised the bar to t ≥ 2.40. The current 3-month regime window was split into two halves, and the holding-to-alpha pattern had to correlate at least +0.5 between them, with the winner beating the control in both halves.
Where it died
Measured as alpha against SPY with beta assumed to be 1, the standard before 2026-09-08. The 15-session hold won: t 3.45, 0.950% per trade over 543 trades. In the first half it reached t 4.18; in the second, t 0.24. The first half held the stress episode, the second a rising market, and the ranking of holds flipped between them.
The deployed 5-session hold was kept. The split-half check was run inside a single 3-month window, not the 2014-2019 / 2020-2026 split the lab used elsewhere, so it tested stability across two regimes of a few weeks each.
The record
Thesis
If the RSI-bounce edge is the correction of a price dislocation caused by indiscriminate selling, it has a natural horizon — the time the dislocation takes to close — and the deployed 5-session hold, chosen without evidence, need not be it. Prediction: alpha per trade peaks at an interior holding period instead of rising monotonically.
Prior: Contaminated origin, declared before the run: the idea came from looking at a control arm of another study, where holds of 5-10 sessions scored better than the deployed configuration. A pass could therefore only authorise a forward test, never a change to the live configuration.
Specification
signal
the deployed RSI-bounce signal, unchanged (RSI(14) < 40; within 10% of the 50-day MA; 20d dollar volume > $3M)
grid
holding ∈ {3, 5, 8, 10, 15, 20} sessions; everything else fixed; 5 sessions is the control
window
current 3-month regime window, split into two halves for the stability check
validation
Bonferroni for 6 variants → t ≥ 2.40; split-half correlation of the holding→alpha pattern ≥ +0.5; winner must beat the control in both halves
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Market-adjusted alpha (legacy)pass
Does the strategy beat SPY over the same holding windows, assuming beta = 1?
best variant t ≥ 2.40 (Bonferroni, 6 variants)
best holding days15best t+3.45best alpha per trade pct+0.95trades543
Alpha vs SPY as measured before the lab adopted Jensen alpha on 2026-09-08.
Stabilityfail
Is the alpha present in both halves of history?
split-half correlation ≥ +0.5
split half correlation-0.63best t first half+4.18best t second half+0.24
Implemented as a split of the 3-month regime window, not the 2014-2019 / 2020-2026 split used elsewhere.
Why it died
Holding 15 sessions cleared the corrected threshold (t 3.45) with twice the alpha per trade of the control — and the ranking of holding periods inverted between the two halves of the window (correlation −0.631). Every long hold won on the first half, which contained the stress episode, and collapsed on the second, a rising market. Only the deployed 5-session hold was positive in both halves.
What it taught
Without the stability gate this would have been the second production change in two days made on noise; a slot-count change based on a pattern with split-half correlation −0.321 had already been deployed. After this study the split-half stability gate became mandatory for every parameter study. Longer holds did not capture more reversion — they captured more market.
Verdict history
2026-09-05Pre-registered with the contaminated origin declared up front
2026-09-05Single run: best t 3.45 passes, split-half −0.631 fails; deployed hold of 5 kept Unstable