short-term mean reversion / overbought
Extreme overbought short
Specification
Signal, universe, exits and sizing are withheld because they are close to a strategy that is still in use.
Gates
Gate 7 — Jensen alphapass
After regressing out market exposure, is there alpha left?
t ≥ 2.13 (Bonferroni, 3 shorts measured together), net of 20%/yr borrow
t (HAC)+2.80alpha / trade %+0.65beta+1.12trades4,799
Gate 8 — Capture-preserved alphapass
Does the subset of trades a real portfolio can actually take keep the alpha?
t_cap ≥ 1.25 (scaled) and ratio ≥ 0.50
t captured+2.80capture ratio+1.34
Gate F — Frictionpass
Is the alpha per trade large enough to pay for its own execution?
alpha/trade ≥ 0.30% net of borrow
alpha / trade %+0.65
Stabilitypass
Is the alpha present in both halves of history?
t > 0 in both halves
t 2014 2019+0.60t 2020 2026+2.75
Own alphafail
Is the alpha new, or already captured by a strategy that passed?
t ≥ 2.13 on the trades that do not overlap the surviving short, n ≥ 100
t own-0.34n own2,484overlap pct+48.20
Nearly half its trades overlap the surviving strategy; the rest have a negative point estimate.
Why it died
Highest alpha per trade of the three archived overbought shorts (0.653% net of borrow, t +2.80), and it passes capture, friction and stability. But all of its alpha sits inside the trades of the lab's surviving short strategy, whose details are not published: the 2,484 trades outside that overlap give t −0.34.
What it taught
In May the lab checked this against its sibling short, found a daily-PnL correlation of only 0.540, and concluded "complementary, not redundant". Correlation between two candidates was the wrong question. The right one — does the alpha survive outside an already-validated strategy's trades? — was only asked in October, and it answered no.
Verdict history
- 2026-05-14Yearly breakdown 11 of 12 years positive, VIX-regime and correlation checks pass; promoted to the demo at 50% size
- 2026-05-27Archived at Gate 8: t_captured 1.56 vs 1.65 (engine alpha t +8.69), under the capture formula and simulator later found broken Not capturable
- 2026-09-14Re-examined on unseen data (2026-05-27 to 2026-09-11) with the corrected Gate 8: n 199, t +0.61 vs 2.128 needed, capture ratio 3.06 No alpha
- 2026-10-23Full-history re-measurement, pre-registered: passes Gates 7, 8, F and stability; own-alpha t −0.34 Redundant