Buying SPY's oversold RSI(2) close, below and above its 200-day average
On SPY, does buying the close when its 2-day RSI is below 5 pay after costs, separately below the 200-day average and above it, with two exit rules?
No. None of the three pre-registered trials passed. The main rule - buy SPY when its 2-day RSI is below 5 and the close is below its 200-day average, sell on the first close above the 5-day average - made +0.83% a trade after costs over 64 separate episodes from 1993 to May 2013, winning 68.8% of them. Its month-clustered lower bound was +0.04% and its deflated Sharpe 0.85, so it cleared the return and deflated-Sharpe checks, but it failed the pre-registered crash check: in October 2008 a trade sat 15.4% under water at a close while open before it exited at -2.0%. The 5-day-hold version below the average made +0.47% a trade, but its lower bound was -0.38% and it also failed the crash check and the deflated Sharpe. The above-average 5-day-hold arm passed discovery under the approved gate (+0.91% a trade over 71 episodes) and failed validation: 28 trades from 2013 to 2020 (under the 30-trade floor), mean +0.27%, lower bound -0.41%. The holdout, from February 2020 on, was not read.
Backtested by QS Research
Pre-registration
Three trials were fixed before any return was read, on SPY from January 1993: buy at the next open after a close with a 2-day RSI below 5, either below the 200-day average or above it, and sell at the next open after either the first close above the 5-day average or the fifth session held. One position at a time, so each trade is a separate episode. The 1993-2026 window is split 60/20/20 by signal day into discovery, validation and holdout. Costs are 5 basis points a round trip, doubled as a stress. Each stage read must clear 30 trades, a positive mean and month-clustered lower bound at a one-sided alpha of 0.05 divided by 3, a positive stress mean, a worst exit loss and a worst close while open, each no worse than 15%, and a drawdown no deeper than buy-and-hold SPY; discovery also needs a deflated Sharpe of 0.80, half the years positive and at most two losing years. The holdout opens only for a trial that passes validation and only when the owner opens it. The classic above-average rule with the first-close-above-the-5-day-average exit was already registered in another study and was not priced here. QQQ and IWM were replicated descriptively, never judged, from January 2015.
Registered 2026-11-19Amendments
- 2026-11-19 Data pins and the rule-only signal table recorded before the counted run; no rule, stage boundary, cost or trial changed. approved: Recorded 2026-11-19 before any return was read
- 2026-11-21 Erratum after the counted run: the QQQ and IWM history starts in January 2015, not 2016, and the 531 signal rows are 370 distinct firing sessions. No rule, bar, trial, stage or cost changed. approved: Dated erratum from the independent review, 2026-11-21
- 2026-11-21 The erratum also corrects the remaining 2016 wording of the secondary window. No rule, bar, trial, stage or cost changed. approved: Dated erratum from the independent review, 2026-11-21
A holdout sample is sealed: its results stay unread until the study is final.
Trials · 3 trials · 0 passed · judged at N = 3 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| bounce-below-sma5-spy | FAIL | counted |
| bounce-below-hold5-spy | FAIL | counted |
| bounce-above-hold5-spy | FAIL | counted |
| bounce-above-hold5-spy | PASS | not counted |
Limitations
- SPY, not the S&P 500 index, and not an option: an option also pays time decay and its spread.
- No span is blind: a 45-look descriptive base rate and an informal 2017-2026 check had already looked at this rule before the counted run; both are disclosed in the pre-registration.
- The counted fills are the next open after the signal close; the signal-close version is reported beside them but never judged.
- Daily closes only: the intraday low inside a trade is not read, so a trade's true worst point can be deeper than the 15.4% close shown.
- Cash earns 0% between trades.
- The below-average arms' validation span (2013-2020) had only 18 signals, so it could not reach 30 trades and was not priced; their verdicts stop at discovery.
- QQQ and IWM are descriptive and never judged. Their history runs from January 2015 (the pre-registration said 2016; a dated erratum corrects it) to November 2026, so they read 2020 and 2022, which are SPY's holdout years; different instruments, but the same rule.
- The above-average 5-day-hold trial is counted once, as a failure at validation; its discovery pass is shown separately and not counted again.
- The study was judged on the gate approved when it was pre-registered. The live gate now raises the discovery floor to 120 trades and validation to 60; under it none of the three trials is measurable. The verdict follows the approved gate, as pre-registered.
- No research-bet slot was confirmed in writing before the run; the pre-registration records that.
- The signal table has 531 rows but 370 distinct firing sessions: a below-average session is listed once for each of the two exits.
- The above-average arm is the classic rule that has been tried before; four earlier trials of it are counted in the deflated Sharpe's effective N of 5.
- The 16-19 November 2026 sessions are not in the pinned data, so the live setup that prompted the study is not one of its trades.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
2 published versions · current 468099ea2668b4e9, published 2026-11-21