Index‑ETF RSI(2) Dip‑Buy

Status: Adopt candidate (paper‑trade first) · Library entry #2 · Found 2026‑08‑01 Instruments: SPY, QQQ (primary) · DIA, IWM (secondary) · broad index ETFs only Market/timeframe: stocks, 1d · Direction: long only · Frequency: ~8–9 trades/yr/ETF, ~3.5‑day holds

The edge

Buy short, sharp dips inside an uptrend on a broad index ETF, and sell the snap‑back a few days later. Multi‑day panic selling in an index overshoots — the index is diversified, so a dip is (almost always) repricing fear, not a business failing — and mean reversion has been the dominant short‑horizon behaviour of US index ETFs since the 1990s.

Two things make this entry unusually trustworthy for a backtest:

• The parameters are not ours. RSI(2) < 10 above the 200‑day SMA, exit above the 5‑day SMA, is Connors' published (2008) spec, taken verbatim. We fitted nothing — so every year since 2008 is genuine out‑of‑sample evidence, and the pre‑2008 windows are the literature's own.
• The sample is big. 57–77 trades per ETF per window, ~800 trades across the study — no single lucky trade can carry it (and strip‑best‑trade proves that below).

The setup (compilable spec)

Market: stocks, symbol SPY (or QQQ), trading timeframe 1d. Longs only. Entry (daily close): close > SMA(200) AND RSI(2) < 10. Exit: close back above the 5‑day SMA (close > SMA(5), condition exit at that day's close). No take‑profit, no stop‑loss (see guardrails — a stop measurably hurts this system). Fees: 0.03%/side (liquid‑ETF spread+slippage; still profitable on SPY/QQQ at 0.1%/side).

Compiled shape: entryConditions = close > sma(200) AND rsi(2) < 10; exitConditions = close > sma(5); exitStrategy = fixed with takeProfitPct = 0, stopLossPct = 0 (both disabled — the condition exit does the work); fees on.

Validation evidence

Held to the library vetting bar; run on the same engine the app uses, with the warm‑up run‑in so the SMA(200) is valid from bar one. Costs 0.03%/side. Returns are summed per‑trade percentages (not compounded) and exclude dividends (conservative for a long‑only system that's in the market ~11% of days).

• Two windows: ✅ three independent multi‑year windows (2000–2010, 2010–2018, 2018–2026), all positive on all four ETFs — two of them fully post‑publication (out‑of‑sample).
• Strip‑best‑trade: ✅ every cell stays positive without its best trade.
• Cross‑asset: ✅ 4 index ETFs. SPY/QQQ strongest; IWM weakest (see guardrails).
• Drawdown: ✅ 4–12% on SPY/QQQ. The 200‑SMA filter carried it through two crashes (2000–2010) at 4–6% DD.
• Cost stress (0.1%/side): SPY/QQQ stay clearly profitable (PF 1.4–2.3). DIA thins to PF ~1.1–1.2; IWM flips slightly negative — hence majors‑primary below.

Robustness (pre‑registered neighbours — a plateau, not a spike)

Every neighbouring variant tested is also profitable in both main windows on SPY and QQQ: entry RSI(2) < 5 and < 15, exit RSI(2) > 65 instead of the 5‑SMA, trend filter SMA(150) and SMA(250) — PF 1.6–3.0 across all of them. The edge does not depend on the exact numbers 10/200/5.

Two ablations worth knowing:

• The SMA(200) filter is a drawdown device, not a return enhancer. Without it the system makes more (dip‑buying kept working in bear markets) but drawdowns double or triple (SPY 2018–2026: 7.8% → 23.8%; QQQ 2000s: 27.4%). Keep the filter.
• A stop‑loss makes it worse, not safer. A 15% disaster stop raised drawdown (QQQ 2010–2018: 10.5% → 21.7%) — it sells the flash‑crash low and misses the snap‑back that mean reversion depends on. Control risk by position size, not by a stop.

Guardrails (read before trading)

• Broad index ETFs only. The mechanism is "a diversified index overshoots and reverts" — a single stock (or a crypto alt) plunging on RSI(2) < 10 may simply be dying. Untested and untrusted outside index majors.
• SPY/QQQ first. IWM carries 21–22% drawdowns and dies at stressed costs; DIA is thin.
• No stop‑loss means gap risk is real. Worst single trades observed: −10% (SPY, Aug‑2015 flash crash), −13% (IWM). Size the position so that a −13% trade is survivable — e.g. a dedicated sleeve of ≤25% of account equity risks ≤3.3% of the account on the worst observed trade. Max losing streak observed: 5.
• It's an overlay, not a moonshot. ~+2.5–4.5%/yr per ETF on the allocated sleeve (non‑compounded), in the market only ~11% of days. Its value is the quality of the edge (high PF, huge sample, shallow DD), not the raw return.
• Expect ~65–75% winners but average loss ≥ average win — psychologically it wins often and loses bigger; that is the shape of mean reversion. Don't "fix" it mid‑drawdown.

Extension: sectors & international (2026‑08‑06, sweep #3)

The spec above, verbatim, zero knobs, on new instruments (research log). Five of seven pass the full bar; the 2000–2010 stress decade is positive on every sector tested (XLK +25.8% through the dot‑com crash):

Read: the mechanism generalizes well beyond the index majors — XLK/EEM join SPY/QQQ as first‑rank instruments; XLF/XLE/EFA work at liquid‑ETF costs but thin at stressed costs; XLV and XLI are excluded honestly (one failed a window, one breached the drawdown bar by 0.1).

Fill honesty: what next‑morning entry costs (2026‑08‑06)

The tables in this write‑up fill at the signal bar's close — the engine's default, and what the early (before‑close) alert lets you approximate. The engine can now also fill at the next morning's open (fillMode: "next_open"); the same spec re‑run that way stays profitable everywhere but gives back a real share of the edge:

Translation: entering near the close matters for this system — a third to half of the edge lives in the last hours of the panic day. Use the early signal; if you can only trade next‑open, the edge survives but expect the thinner column.

Trading it in the app

Build it as a Sensei plan (/ai/plan): market stocks, symbol SPY, timeframe 1d, the spec above. The live entry finder watches a plan's own symbol for stocks too, so signals fire in the app. Set signalMode: "early" — the daily bar "closes" only after the session, so the early (before‑close) alert is what lets you enter near the close the backtest assumes; the confirmed signal lands after the bar completes (entering next open is a slight, untested deviation from the tested fills). Universe scanning is crypto‑only today, which doesn't matter here — the symbol is fixed.

How it was found

Pre‑registered as a hypothesis from the mean‑reversion literature (strongest prior available: documented for ~30 years, parameters frozen since 2008), then validated on the app's own engine — not discovered by searching parameter space. The only exploration run was the pre‑registered neighbour/ablation grid above, reported in full: nothing tested is hidden, nothing was tuned. Candidate hypotheses that were not needed (crypto volatility‑squeeze breakout, daily trend‑pullback) were left untested rather than mined for a second entry.