Not pre-registered — see caveat · data snapshot 2026-09-19T090800Z
The idea
Large, liquid stocks rarely reach an extreme oversold reading on both a fast and a slow RSI at once. When they do while still not far below their 200-day average, the bet is that the drop was capitulation rather than news, and that it rebounds within a week. No mechanism or counterparty was written down when the configuration was found.
It is one of the lab's canonical set: early configurations chosen on Sharpe, then frozen and re-measured under locked rules whenever the instruments changed. It is the looser sibling of the strict big-cap oversold configuration: a deeper oversold reading, a lower price floor and half the liquidity floor.
How it was tested
Signal: RSI(5) below 20, RSI(14) below 30, and price no more than 10% below its 200-day average. Universe: US stocks priced at $30 or more with at least $100M of daily dollar volume. Buy at the next open, hold 7 sessions, with a 2×ATR stop and a 5×ATR target. The re-measurements size it at a neutral $5,000, 1% risk per trade, 10 slots.
Where it died
Its record is a staircase down. On 2026-05-14 the engine stopped buying the same symbol again while a position was open, and 2020 became a losing year with a Sharpe of −1.73; it was cut to half size. On 2026-05-27 it was archived on alpha against SPY with beta assumed to be 1, at t −0.67.
The later instruments only confirmed it. Jensen alpha on full history gave t −0.47 with a beta of 0.955 on 2026-09-08. The gap-fill fix of 2026-10-17, which fills a stop at the open when the bar gaps past it, took t to −1.87. With the calendar standard error of 2026-10-30 the record reads t −1.669: −0.157% per trade over 4,392 trades, beta 1.015.
The record
Thesis
Liquid large caps hitting an extreme short-term oversold reading (rare deep capitulation) while not far below their 200-day average rebound over the following week. No mechanism was written down when it was found.
Specification
signal
RSI(5) < 20; RSI(14) < 30; dist_ma200 > −10%
universe
US stocks, price ≥ $30, 20d dollar volume ≥ $100M
entry
next open
exit
7 sessions, 2×ATR stop, 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 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)-1.67alpha / trade %-0.16beta+1.01trades4,392
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
Beta about 1 and negative alpha: buying deep-oversold big caps is buying the market, minus a little. The first strong warning came from the engine fix of May 2026: once overlapping re-entries in the same symbol were no longer counted as independent trades, March-April 2020 turned from a winning year into the worst one, because the first entry of a crash holds through the rest of the fall.
What it taught
A backtest that lets the same symbol be bought again and again during a crash is simulating a trader who catches every bounce and none of the first leg down. Removing that artefact took the strategy from 12 of 12 positive years to 10 of 12, and the later benchmark-adjusted measurements took it below zero.
Verdict history
2026-05-14Engine changed to one position per symbol: 2020 becomes a losing year (Sharpe −1.73); downgraded to half size
2026-05-27Archived on alpha vs SPY with beta assumed = 1: t −0.67 No alpha
2026-09-08Re-measured with Jensen alpha on full history: t −0.47, beta 0.955
2026-10-17Gap-fill engine fix: t falls to −1.87
2026-10-30Standard error by calendar (audit): Gate 7 t -1.868 → -1.669, same trades and alpha