Not pre-registered — see caveat · data snapshot 2026-09-19T090800Z
The idea
Buy expensive, very liquid stocks after a sharp short-term drop, but only when they are still close to their 200-day average. The implicit reasoning is that selling in a name whose long-term trend is intact is more likely liquidity than information, and that it is undone within a week. Nobody wrote that mechanism down when the configuration was found.
It belongs to the lab's canonical set: configurations from the early history, selected on Sharpe rather than pre-registered, then frozen and re-measured each time an instrument changed. Its closest relative is the deep-oversold big-cap configuration, which uses a deeper oversold reading and a looser liquidity floor.
How it was tested
Signal: RSI(14) below 30, price no more than 8% below its 200-day average, and a 5-day return worse than −5%. Universe: US stocks priced at $50 or more with at least $200M of daily dollar volume. Buy at the next open, exit after 5 sessions, with a 1.5×ATR stop and a 3.5×ATR target. The re-measurements use a neutral default: $5,000, 1% risk per trade, 10 slots. Because it was not pre-registered, only the later, locked re-measurements count as verdicts.
Where it died
In May 2026 it looked deployable. After the engine was limited to one position per symbol on 2026-05-14, in-sample trades fell from 1,533 to 577, and the out-of-sample Sharpe was still 2.929. It was archived on 2026-05-27 on Gate 8, with t_captured 0.42, using a portfolio simulator later found to leak slots; that reason was voided on 2026-09-19.
It did not need Gate 8 to fail. Jensen alpha on full history gave t +1.34 on 2026-09-08 and t +1.204 on the frozen snapshot. The gap-fill fix of 2026-10-17, which fills a stop at the open when the bar gaps past it, took it to +0.25. Since the calendar standard error of 2026-10-30 the record reads t +0.265: 0.031% per trade over 1,401 trades, with a beta of 0.92.
The record
Thesis
Expensive, very liquid stocks that become short-term oversold while still close to their 200-day average bounce within a week. The ficha never wrote down a mechanism; the implicit one is liquidity-driven selling in names whose long-term trend is intact.
Specification
signal
RSI(14) < 30; within 8% of the 200-day MA (dist_ma200 > −8%); 5-day return < −5%
universe
US stocks, price ≥ $50, 20d dollar volume ≥ $200M
entry
next open
exit
5 sessions, 1.5×ATR stop, 3.5×ATR target
sizing
neutral default for the re-measurement — $5,000, 1% risk per trade, 10 slots
caveat
Not discovered by pre-registration: an early-history configuration selected on in-sample / out-of-sample Sharpe, later frozen and re-measured under locked rules with the rest of the canonical set.
Gates
Gate 7 — Jensen alphafail
After regressing out market exposure, is there alpha left?
t ≥ 1.65
t (calendar SE)+0.27alpha / trade %+0.03beta+0.92trades1,401
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 — the re-measurement protocol computes Gate 8 only for strategies that pass Gate 7.
Why it died
Beta 0.92 and almost no alpha left: about 0.03% per trade after regressing out the market. In May it looked like a 10-of-12-positive-years strategy with an out-of-sample Sharpe near 3; that was the market's return in big caps, measured without a benchmark. Its May archive reason (Gate 8 failure) was decided with the portfolio simulator that was later found to leak slots, so that reason is void — but the strategy fails one gate earlier anyway.
What it taught
The trail of this page is the trail of the lab's instruments: a raw Sharpe said "deploy", beta-1 alpha said "weak but real", Jensen alpha said "+1.34, not enough", and fixing gap-through stops took it to +0.25. Every fix moved the number in the same direction — down — which is what a long mean-reversion signal whose stops were being filled at impossible prices should look like.
Verdict history
2026-05-14Engine changed to one position per symbol: in-sample trades fall from 1,533 to 577; out-of-sample Sharpe 2.929; 10 of 12 years positive
2026-05-27Archived on Gate 8: t_captured 0.42 (verdict later voided by the 2026-09-19 simulator slot-leak fix) Not capturable
2026-09-08Re-measured with Jensen alpha on full history: t +1.34, beta 0.868 — Gate 7 fail No alpha
2026-09-19Re-derived on the frozen snapshot: t +1.204, still Gate 7 fail
2026-10-17Gap-fill engine fix (stops filled at the open when the bar gaps past them): t +1.20 → +0.25 No alpha
2026-10-30Standard error by calendar (audit): Gate 7 t +0.249 → +0.265, same trades and alpha