No alphaArchived · long · 2026 (3-month regime window)
Pre-registered 2026-09-05 · 1 run · 8 variants
The idea
In a sell-off, margin calls, redemptions and cascading stops force people to sell what can be sold rather than what should be. Prices of names with no problem of their own fall away from value, and a patient buyer is paid to take the other side (Shleifer and Vishny on fire sales). In calm markets the seller usually knows something. If that is right, mean reversion is not a permanent edge: it should live in stress and be about zero outside it.
The idea came from an audit of the deployed RSI-bounce signal: 3 of 13 weeks carried 115% of its alpha, and SPY averaged −0.99% in those weeks against +0.07% in the others.
How it was tested
Pre-registered on 2026-09-05, together with a stress-episode holdout. The RSI-bounce signal stayed unchanged; only a stress gate was added. Two gates: VIX above its 252-day 70th percentile, or SPY at least 3% below its 20-session high. Each was crossed with holds of 5 or 10 sessions and 10 or 20 slots, 8 variants in all. Bonferroni set the bar at t ≥ 2.50, and the best gated variant also had to beat the ungated base.
Where it died
No gated variant passed: 0 of 8. The best reached t 2.19, below both the 2.50 bar and the ungated base at 3.60. Alpha was measured against SPY with beta assumed to be 1, the method in use before the lab adopted Jensen alpha on 2026-09-08.
The direction of the thesis survived; its strong form did not. The base it was trying to sharpen did not last either: on 2026-10-19 the RSI-bounce signal was archived, with no alpha over full history in its own stress regime.
The record
Thesis
Mean reversion is not a persistent edge but one conditional on market stress. In a sell-off, forced selling (margin calls, redemptions, cascading stops) sells what can be sold rather than what should be, decoupling price from value in names with no problem of their own; in calm markets the seller usually knows something. Restricting entries to stress should concentrate the alpha — and the alpha should be about zero outside it.
Prior: Motivated by an audit of the deployed RSI-bounce signal: 3 of 13 weeks carried 115% of its alpha, and SPY averaged −0.99% in those weeks against +0.07% in the others.
Shleifer, A. & Vishny, R. (2011). Fire Sales in Finance and Macroeconomics. Journal of Economic Perspectives.
Specification
signal
the deployed RSI-bounce signal, unchanged, plus a stress gate
gates tested
G1 — VIX above its 252-day 70th percentile; G2 — SPY in a drawdown of at least 3% from its 20-session high
Bonferroni for 8 variants → t ≥ 2.50; the best gated variant must also beat the ungated base
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)fail
Does the strategy beat SPY over the same holding windows, assuming beta = 1?
best gated variant t ≥ 2.50 (Bonferroni, 8 variants) and above the ungated base
best gated t+2.19ungated base t+3.60variants passing0 / 8
Alpha vs SPY as measured before the lab adopted Jensen alpha on 2026-09-08.
Why it died
The strong form of the thesis is false. Under the VIX gate, alpha per trade was 2.6× larger in stress (+0.882% vs +0.333%) — the direction was right — but still positive outside it, so the gate threw away 495 profitable trades to keep 131 better ones and the t-statistic fell from 3.60 to 2.19. The SPY drawdown gate was worse than useless: the entries it accepted had lower alpha than the ones it rejected.
What it taught
Per-trade quality and statistical power moved in opposite directions, and power won. If an edge is stronger in stress but not absent outside it, the expression is variable sizing, not an on/off gate. Both proxies encoded severe, sustained stress while the alpha lived in milder weekly declines; testing softer thresholds afterwards would have been mining this very result.
Verdict history
2026-09-05Pre-registered with 8 variants and a stress-episode holdout
2026-09-05Single run: best gated t 2.19 < 2.50; gate degrades the signal No alpha
2026-10-19The underlying RSI-bounce signal itself was later archived: in its stress regime it has no alpha over full history