Pre-registered 2026-10-14 (commit b80e123) · 1 run · data snapshot 2026-09-19T090800Z
The idea
A sharp drop in a normally calm stock is large relative to its usual moves, so it is more likely temporary selling pressure than news; in a volatile stock the same RSI reading is routine. The bet combines two signals the lab had already seen something in: the RSI bounce and low volatility. A 200-day uptrend is required so as not to buy a falling knife.
It was the sixth RSI variant the lab had tried. The pre-registration said so, and noted that its multiple-testing correction covered only its own batch, not the whole RSI family.
How it was tested
Committed on 2026-10-14 (commit b80e123) as one of a batch of three pre-registered conjunctions, before any of them ran. Signal: RSI(14) below 35, ATR(14) below 2.5% of price, price above the 200-day average, at least $3M of daily dollar volume. Universe: US stocks at $5 or more. Buy at the next open, hold 10 sessions, 2×ATR stop, 2.5×ATR target, most oversold first, at a neutral $5,000 with 1% risk per trade and 10 slots. Full history on the frozen snapshot. Three conjunctions set the Gate 7 bar at t ≥ 2.13.
Where it died
Gate 7 returned t +0.58: 0.04% per trade over 8,162 trades, with a beta of 0.342. Gate 8 was not computed. Unlike the earlier RSI variants, this one was measured over 2014-2026 rather than a 3-month window, so a lack of power is not the explanation.
The record
Thesis
An oversold RSI in a normally calm stock is more likely liquidity noise (temporary selling pressure) than information, because the drop is large relative to the stock's usual volatility; in a volatile stock the same RSI is routine. The volatility filter is meant to separate noise from news, with a 200-day uptrend required so as not to buy a falling knife.
Prior: Sixth RSI variant in the lab. The batch correction covers the batch of three, not the whole RSI family.
Connors, L. — volatility filters on short-term RSI (practitioner precedent).
Specification
signal
RSI(14) < 35; ATR(14) < 2.5% of price; above the 200-day MA; 20d dollar volume > $3M
universe
US stocks, price ≥ $5, 20d dollar volume ≥ $1M
entry
next open
exit
10 sessions, 2×ATR stop, 2.5×ATR target
ranking
most oversold first
sizing
neutral default — $5,000, 1% risk per trade, 10 slots
validation
one of a batch of 3 pre-registered conjunctions; Bonferroni k=3 → t ≥ 2.13; Gate 8; alpha ≥ 0.30% per trade
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 ≥ 2.13 (Bonferroni, batch of 3)
t (HAC)+0.58alpha / trade %+0.04beta+0.34trades8,162
Gate 8 — Capture-preserved alphanot evaluated
Does the subset of trades a real portfolio can actually take keep the alpha?
Not evaluated once Gate 7 fails.
Why it died
The low-volatility filter did not reinforce RSI reversion — it removed it: t +0.58 and 0.04% alpha per trade over 8,162 trades. The two ingredients that had each shown something on their own did not combine.
What it taught
Two components with apparent alpha are not additive. Neither survived on its own either: the low-volatility breakout had already been archived on friction, and the RSI-bounce signal it borrowed from was archived five days later for lack of alpha in its own regime. The portfolio simulator also produced negative equity on this run — impossible in a real account — which was logged as a separate defect.
Verdict history
2026-10-14Pre-registered with two other conjunctions, committed before running