Not pre-registered — see caveat · data snapshot 2026-10-19T100333Z
The idea
In a choppy, rotating market, liquid stocks that become moderately oversold while still near their 50-day average tend to bounce within a week. No counterparty was named. It was treated from the start as a regime edge, expected to last weeks rather than years.
This is the parent of a family. The holding-period study, the stress-gated study, the tightening to a more oversold reading, the long-term trend filter and the low-volatility conjunction were all pre-registered variations of this one signal.
How it was tested
It was not pre-registered. On 2026-09-03 a factor scan of the previous two months, followed by a four-combination parameter test, found it with a Sharpe of 2.082 on July-September 2026, and it went into the demo the same evening. Signal: RSI(14) below 40, later 35, price within 10% of the 50-day average, at least $3M of daily dollar volume. Buy at the next open, hold 5 sessions, with a 2×ATR stop and a 2.5×ATR target, most oversold first. In the demo it only took entries while the lab's regime classifier read STRESS. Every later measurement was pre-registered.
Where it died
The warnings came early. On 2026-09-04 an audit found that 3 of 13 weeks carried 115% of its alpha; without them, t was −0.44. Restoring the full universe on 2026-09-07 cut alpha per trade from 0.629% to 0.364%. A pre-registered tightening to RSI below 30 the same day made it worse: 0.123% per trade, t 0.69.
The number that kept it alive was a Jensen t of +3.03 on 2026-09-08, beta 0.553, measured on every signal, not only on the stress days the sleeve actually traded. After the gap-fill fix of 2026-10-17, full history gave t −0.815. On 2026-10-19 the regime gate was corrected to read the previous day's regime, removing a look-ahead, and full history inside its own regime came out at t −2.38 over 11,030 trades. It was archived and removed from the demo that day.
The record
Thesis
In a choppy, rotating market, liquid stocks that become moderately oversold while still near their 50-day average rebound within a week. Treated from the start as a regime edge — expected to work for weeks, not years — and traded only while the market was in a stress regime.
Specification
signal
RSI(14) < 35 (originally < 40); within 10% of the 50-day MA; 20d dollar volume > $3M
entry
next open
exit
5 sessions, 2×ATR stop, 2.5×ATR target
ranking
most oversold first
regime
entries only while the market regime classifier reads STRESS
caveat
Not pre-registered. Found on 2026-09-03 by a factor scan of the previous two months followed by a four-combination parameter test, and deployed in the demo the same evening. Every later measurement was pre-registered.
Gates
Gate 7 — Jensen alphafail
After regressing out market exposure, is there alpha left?
t ≥ 1.65
t (HAC)-2.38alpha / trade %-0.11trades11,030t hac current window-1.58n current window64
Full history restricted to the stress regime the strategy trades, with the regime read from the previous day (no look-ahead). Overlapping holds make the HAC standard error too small, so the honest reading is "no alpha", not "negative alpha". The current 3-month window has only 4 entry days.
Gate 8 — Capture-preserved alphanot evaluated
Does the subset of trades a real portfolio can actually take keep the alpha?
The capture ratio is 0 by construction when alpha is ≤ 0 and carries no information.
Why it died
The number that justified deploying it was never measured on the population it could trade. The often-quoted t +3.03 was for all signals; restricted to the stress regime the sleeve actually traded it was −0.51 (n=166). Re-measured on repaired 2026 data the current-window result was t −1.43, and once the regime filter was corrected to use the previous day's regime the full history in its own regime came out at t −2.38 over 11,030 trades.
What it taught
An edge defined by "only in this regime" must be measured only in that regime — the ungated number described trades the strategy was forbidden to take. Two parameter studies on it also taught the lab two permanent rules: a split-half stability gate for every parameter study (the holding-period study), and that the most extreme reading is often the most justified one (tightening RSI from 35 to 30 cut alpha per trade to a third).
Verdict history
2026-09-03Found by a 2-month factor scan + 4-combination test (Sharpe 2.082 on July-September 2026); deployed in the demo the same day
2026-09-04Audit: 3 of 13 weeks carried 115% of its alpha; without them t −0.44 — motivated the stress-gated hypothesis
2026-09-07Re-measured after the universe was restored (~4,550 → ~7,443 symbols), alpha vs SPY with beta = 1: t +2.62, 0.364%/trade (was 0.629%), Gate 8 ratio 1.53 Passed all gates
2026-09-07Pre-registered tightening to RSI < 30: alpha falls to 0.123%/trade, t 0.69 — more oversold is worse; not tradeable at any size No alpha
2026-09-08Jensen alpha re-confirmation on the regime window: t +3.03, beta 0.553 (population not filtered by regime) Passed all gates
2026-10-17Gap-fill engine fix; full history t −0.815 (never had durable alpha)
2026-10-19Re-measured on repaired 2026 data: current window in its regime t −1.43 (n=45), FAIL by its own archive rule No alpha
2026-10-19Regime gate corrected to the previous day's regime: full history in regime t −2.38 (n=11,030); archived and removed from the demo No alpha