Home › Articles

Short-Period RSI Oversold Signals and RSI Golden Crosses in US Index ETFs (QQQ/VOO/SPY): Literature and Empirical Background

2026-10-19 · Research note · Series: RSI oversold crossover backtest (part 3 of 3) · Not investment advice; figures are as of the sources and dates cited in the text · 中文
Contents7 sections
  1. 1 Published backtests of short-period RSI oversold entries
  2. 2 Claims and backtests of the RSI golden cross (short RSI crossing above long RSI)
  3. 3 Signal frequency and failure scenarios (stuck readings in bear markets)
  4. 4 Academic evidence: short-term reversal and medium-term momentum at the index level
  5. 5 Differences in RSI calculation: Wilder versus Cutler SMA
  6. Further reading
  7. Frequently asked questions

This note collects published backtests, academic research and practitioner research as reference context for interpreting the backtest of "RSI6 < 20, then RSI6 crosses above RSI10" on QQQ and VOO. Research date: 2026-10-19. Each subsection gives the year of its sources and marks which material is pre-2010 data and which comes from blog/vendor backtests that may involve data mining.

Site note: This is the literature and empirical background appendix to "RSI Oversold Golden Crosses Buy Low, but Rarely Buy the Bottom". The content is the full text of the research note literature_context.md (kept in the project's research_notes/ folder for the QQQ/VOO RSI oversold golden-cross backtest), published as is (only the table of contents at the top was removed). Research date 2026-10-19. Following the original, each section is split into Takeaway (key points), Cited Findings (findings with sources), Inferences (the author's inferences) and Gaps (data not found or not obtainable); the original already marks pre-2010 data, blog/vendor backtests that may involve data mining, and figures that come only from search snippets and could not be verified on the original page, and this site keeps those labels as in the original. This is a research record, not investment advice.

1 Published backtests of short-period RSI oversold entries

Takeaway

Published backtests of going long on short-period RSI (RSI2/RSI5) oversold readings in SPY/QQQ almost uniformly show "a high win rate (about 70%–80%), small per-trade returns (about 1%), short holding periods (about a week) and low time in the market (about 8%–35%)". These strategies can reduce drawdowns, but absolute returns generally lag buy-and-hold; an independent backtest of the last decade (2016–2026) including costs shows an annualized return of only 1.7%, versus 13.5% for buy-and-hold over the same period.

Cited Findings

  • The RSI(2) strategy was popularized by Connors and Alvarez in their 2008 book; the classic rule adds a trend filter: go long only when the close is above the 200-day moving average (2008; pre-2010 data) — LuxAlgo RSI 2 concept page
  • Oversold thresholds vary by source: one says Connors usually uses RSI(2) < 5, while some authors use < 10 because it produces more signals — LuxAlgo RSI 2
  • The original book is "Short Term Trading Strategies That Work" (TradingMarkets, 2008) — MQL5 product page (marketing page, low credibility)
  • Independent backtest of the last decade (Backtrex): S&P 500 daily, 2016-10-03 to 2026-10-01; rules are enter when RSI(2) < 5 and the close is above the 200-day moving average, exit when the close gets back above the 5-day moving average, costs included. Results: total return +18.1% vs +255.4% for buy-and-hold; CAGR 1.7% vs 13.5% for buy-and-hold; maximum drawdown -21.5% vs -35.0% for buy-and-hold; 73 trades, win rate 71.23%; average win +1.22%, average loss -2.10%; profit factor 1.43. Returns by year: 2018 -9.8% (worst), 2019 -3.0%, 2020 -1.3%, 2022 +0.9%, 2024 +8.6%, 2025 +6.4%. The page's own assessment is "profitable, but lags buy-and-hold" with a thin edge — Backtrex
  • A 2010–2022 SPY RSI(2) backtest on QuantConnect: net profit 158.251%, annualized 7.568% (in-sample) — cited from a QuantConnect forum post listed in a search snippet (the original page could not be verified directly)
  • Quantified Strategies (vendor blog, possible data mining): the "RSI pullback" strategy on SPY had 184 trades, a 78% win rate, CAGR 3.9%, about 8% time in the market and a 10% maximum drawdown, versus a 55% drawdown for buy-and-hold — Quantified Strategies: RSI pullback
  • Quantified Strategies' "Triple RSI": enter when RSI5 < 30, after three consecutive down days, with the close above the 200-day moving average; exit when RSI5 crosses above 50; claimed win rate about 90% (91.03%) — Triple RSI; 91.03% win-rate post
  • Quantified Strategies' QQQ RSI(2) mean reversion (with a second filter): CAGR 12.75%, with buy-and-hold listed in that article at 6.4%; 14% time in the market, 196 trades, 75% win rate, 19.5% maximum drawdown. These figures come only from a search snippet; the original page now returns 404, so the sample period and whether it covers 2022 cannot be confirmed — Quantified Strategies QQQ (dead link)
  • A Quantified Strategies SPY RSI post from 2024-07-31 says "RSI works well on SPY", with a backtest period from 1993 to the present, but the rules and data are behind a paywall — Quantified Strategies
  • Arthur Hill (StockCharts, 2016): RSI5, thresholds 30/70, trading only when the S&P 500's 50-day moving average is above its 200-day moving average, tested on SPY, QQQ and IJR from 2000-12 to 2016-12. Across 12 systems the average win rate was 78%, with 20%–35% time in the market. SPY and QQQ both had positive expectancy, but IJR (small caps) did best. Over the same period SPY buy-and-hold returned 5.22% annualized with a 55% maximum drawdown — StockCharts SystemTrader
  • Hill also found that widening the thresholds to 25/75 or 20/80 actually made drawdowns larger — StockCharts SystemTrader
  • Sentimentrader (2017): buy SPY when the share of S&P 1500 constituents with RSI2 < 30 is ≥ 70%; 2005–2017, 72 trades, 79% win rate, 1.06% average per trade, average holding 8 days, 6.54% annualized. Out-of-sample performance was "worse than in-sample", which the author speculates is related to the volatility regime — Sentimentrader
  • CXO Advisory (2008; pre-2010 data): an RSI14 30/70 strategy on SPY, 1993-01 to 2008-02, 34 round trips, 27 wins and 7 losses, +2.5% average per trade, 29% time in the market, lagging buy-and-hold over the full period; it lagged in bull markets and won in bear markets; results depend on the start date — CXO Advisory

Inferences

  • The high win rate comes mainly from the structure of "short holding, small profit target, no stop-loss", and does not mean the strategy beats buy-and-hold. Indexes rise over the long run while these strategies are in the market only 10%–35% of the time, so cumulative returns usually lag. To compare with the user's backtest, look at both "per-trade return × share of time in the market" and "buy-and-hold over the same period".
  • Almost all published backtests add a trend filter such as the 200-day moving average; versions without a filter carry clearly higher risk in bear markets (see Section 3).
  • Independent cost-inclusive tests after 2016 differ from the book/blog figures of 2008–2017, which suggests possible decay, or possibly in-sample overfitting in the original work; but Backtrex's year-by-year data show no monotonic decay.

Gaps

  • No first-hand figures were found for the win rate and average return of SPY RSI(2) in the original Connors/Alvarez book; recent follow-up articles from Alvarez Quant Trading were not obtained either.
  • No published backtest using an RSI6 (6-day RSI) threshold on SPY/QQQ was found, nor any forward-return statistics for fixed 1/3/6-month holding periods; the existing literature mostly uses holding periods of a few days to two weeks.
  • No year-by-year RSI backtest of QQQ during the 2022 bear market was obtained.

2 Claims and backtests of the RSI golden cross (short RSI crossing above long RSI)

Takeaway

Taiwanese and Chinese-language tutorials commonly treat "a short-period RSI crossing above a long-period RSI" as a buy signal and "RSI < 20" as oversold, but these are mostly teaching rules of thumb; the definition of the low zone is inconsistent (20 or 50), and no methodologically transparent RSI6/RSI12 golden-cross backtest on US index ETFs was found.

Cited Findings

  • Financial education sites treat RSI < 20 as oversold and the market as too cold; a short-period RSI breaking above a long-period RSI signals enough upward momentum and can be treated as a buy — CMoney RSI tutorial; StockFeel
  • An article on Wantgoo uses RSI5 and RSI10 as an example and defines the low zone for a "low-zone golden cross" as below 50, not 20 — Wantgoo
  • The table of contents of a Da-Yeh University thesis includes chapters on weekly RSI6/RSI12 and weekly RSI10/RSI20 trading strategies and cites an investment-performance comparison study by Li Su-hui; but only the table of contents was obtained, no performance figures were seen, and it covers Taiwan stocks on weekly bars — Da-Yeh University thesis
  • One tutorial source warns that RSI is "insensitive" above 80 or below 20 (the indicator gets stuck), and that entering in the low zone during a persistent decline can leave you trapped — Chiang Ming-cheng, Mingcheng Academy (Anue/cnyes)

Inferences

  • "After RSI6 < 20, wait for RSI6 to cross above RSI10" is essentially a combination of "oversold + confirmed reversal", close in logic to Hill's RSI5 "enter only after crossing back above 30" (his Test 3 and Test 4); Hill's results show these confirm-first versions had better CAR/MDD across the three ETFs (IJR's Test 3 was highest).
  • A cross between a short RSI and a long RSI is mathematically close to "recent up/down strength turning positive", which is highly correlated with "price crossing above a short moving average", so its effect may be similar to an ordinary short moving-average breakout rather than new, independent information. This is an inference; no literature testing it directly was found.

Gaps

  • No published quantitative backtest or academic literature was found on RSI6/RSI12 or RSI6/RSI10 golden crosses on SPY, VOO or QQQ.
  • Chinese-language tutorial sources generally give no sample period, trading costs or benchmark, so they cannot be evaluated.

3 Signal frequency and failure scenarios (stuck readings in bear markets)

Takeaway

No published statistics were found on how often RSI6 < 20 occurs in SPY/QQQ. The main known failure mode is that in bear markets oversold signals recur repeatedly while prices keep falling (RSI gets stuck), so most published strategies rely on a 200-day moving average filter or a 50/200-day "death cross" filter to stop trading during bear markets.

Cited Findings

  • Hill's system does "not trade" when the S&P 500's 50-day moving average is below its 200-day moving average (a death cross), precisely to avoid bear-market periods — StockCharts SystemTrader
  • The classic RSI(2) rule goes long only when the price is above the 200-day moving average — LuxAlgo RSI 2
  • A StockCharts forum comment says the original system was used only on SPY daily bars and once suffered a huge drawdown (an unverified forum claim) — StockCharts forum
  • In CXO's RSI14 test, the largest single-trade loss of -13.9% occurred in mid-2002 (the 2000–2002 bear market), and the largest single-trade gain of 13.3% came in early 2000 — CXO Advisory
  • Backtrex's last-decade S&P 500 RSI(2): 2018 was the worst at -9.8%; 2020 -1.3% and 2022 +0.9%, showing that with a trend filter the 2020 and 2022 bear markets did not cause large losses, but the sharp drops and whipsaws of 2018 still produced a loss — Backtrex
  • RSI readings at 80/20 or 90/10 occur less often than at 70/30 and indicate stronger momentum — Wikipedia: Relative strength index
  • Chinese-language tutorials note that when a stock keeps falling, RSI may stay below 30 for a long time (getting stuck in a bear market) — CMoney; Chiang Ming-cheng (Anue/cnyes)

Inferences

  • If the user's backtest has no trend filter, one can expect repeated consecutive signals in 2000–2002 (especially severe for QQQ), 2008 and 2022, followed by negative 1/3-month returns; splitting returns into two groups, "above/below the 200-day moving average", is the most direct way to test this.
  • The confirmation condition "wait for RSI6 to cross above RSI10" can filter out some "catching a falling knife" entries, but it will still trigger frequently during bear-market rallies, after which prices break to new lows.
  • QQQ is more volatile than the S&P 500, so the same RSI6 < 20 condition may occur more often and with deeper declines in QQQ; this is an inference that needs to be verified with the user's backtest data.

Gaps

  • No published statistics were found on how many times per year RSI6 < 20 occurs on average in SPY, VOO or QQQ.
  • No study was found that splits forward returns of RSI oversold signals by bear-market periods such as 2000–2002, 2008 and 2022.

4 Academic evidence: short-term reversal and medium-term momentum at the index level

Takeaway

The academic literature as a whole supports "time-series momentum in indexes over 1–12 months", while short-term negative autocorrelation in daily index returns is weak and unstable. Early evidence of profitable technical analysis (BLL 1992) has largely faded in major markets after 1986; the review by Park & Irwin also points out that most studies have data-mining problems. This is an unfavorable backdrop for "buying oversold readings beats buy-and-hold over 1/3/6 months": at medium horizons, indexes behave more like continuation than reversal.

Cited Findings

  • Brock, Lakonishok & LeBaron (1992, Journal of Finance; pre-2010 data): using Dow Jones index data from 1897–1986 to test moving-average and trading-range-break rules, the authors state that the results "strongly support" technical strategies — IDEAS/RePEc
  • Urquhart, Gebka & Hudson (2015): retesting data after BLL, the predictive power of moving-average rules weakened in all three markets after 1986 — Southampton ePrints; IDEAS
  • Ulku (2009): the profitability of the BLL rules has disappeared in recent DJIA data, but they performed better on the Istanbul index — IDEAS
  • Park & Irwin (2007, Journal of Economic Surveys 21(4):786–826): source of the paper — IDEAS. Summary of its 2004 working-paper version: of 92 modern studies, 58 were positive, 24 negative and 10 mixed, but most studies have problems such as data mining, ex-post rule selection and hard-to-estimate risk and trading costs; about 30%–40% of practitioners consider technical analysis important in the short term, within 6 months — AgEcon Search working paper
  • Moskowitz, Ooi & Pedersen (2012, Journal of Financial Economics): returns of 58 liquid futures (including equity index futures) persist for 1–12 months and then partly reverse over longer horizons; the authors attribute this to sentiment-driven initial under-reaction and delayed over-reaction — SSRN; the sample includes 9 equity indexes over 1985–2009 — AQR dataset
  • Autocorrelation of daily index returns: a 2011 R-bloggers rolling AR(1) analysis suggests the S&P 500 may have shown signs of mean reversion in recent years, but the evidence is weak; a 2009 Angry Bear post commenting on Nate Silver's claim that "the S&P 500 has shown day-to-day mean reversion since 2000" argues that it is over-interpreted — R-bloggers; Angry Bear
  • A summary of financial "stylized facts": autocorrelation of price changes is insignificant — Martin Sewell stylized facts

Inferences

  • The edge of oversold bounces lies mainly at horizons of a few days to two weeks (average holding in published backtests is 7–8 days); at 1–6 months, the index-level time-series momentum literature instead implies that "an index that has recently fallen sharply does not necessarily have higher future returns". So if the user's backtest looks at 1/3/6-month holding periods, the results may mainly reflect "whether pullbacks were bought during a bull market", not the predictive power of RSI itself.
  • QQQ versus the S&P 500: in Hill's 2000–2016 test, SPY and QQQ both had positive expectancy, but no comparable figures were given; QQQ went through a decline of roughly eighty percent in 2000–2002, so a backtest starting in 1999 would be dragged down severely. This is an inference; the user needs to verify it by splitting the period.

Gaps

  • The original texts of short-term reversal studies such as Lehmann (1990) and Jegadeesh (1990) were not obtained; those studies are mainly at the individual-stock level and cannot be applied directly to index ETFs.
  • No academic study was found that specifically compares the strength of short-term reversal in QQQ versus SPY.
  • The original abstract of the formal 2007 version of Park & Irwin was not obtained; the figures above come from the 2004 working paper.

5 Differences in RSI calculation: Wilder versus Cutler SMA

Takeaway

Wilder's original RSI uses a smoothed moving average (SMMA/RMA) with α = 1/n, while Cutler's version uses a simple moving average (SMA) instead; the two give similar but not identical values, and some commercial software uses a standard EMA. For a short period like RSI6, the SMA version is more sensitive to single-day big moves and drops old data faster, which may make the number and dates of "< 20" and crossover triggers differ from the Wilder version, so the formula should be standardized before comparing across software.

Cited Findings

  • Wilder's RSI: the up move U and down move D are each smoothed with an n-period SMMA (α = 1/n), RSI = 100 − 100/(1 + RS); Wilder recommended n = 14, with the first value seeded by the simple average of the first n periods — Wikipedia: Relative strength index
  • Cutler's RSI replaces the SMMA with an SMA because Wilder's values depend on where the data start (Data Length Dependency); Cutler's version gives the same result regardless of data length or starting point; the two differ slightly but are close — Wikipedia
  • Some commercial software (for example AIQ) uses a standard EMA rather than Wilder's SMMA — Wikipedia
  • Wilder's smoothing is the RMA, an EMA-type recursion with α = 1/length — LuxAlgo RSI

Inferences

  • A standard EMA has α = 2/(n+1): at n = 6 that is about 0.286, versus Wilder's 1/6 of about 0.167, so the EMA version of RSI6 is more volatile and hits < 20 more often; the SMA version (Cutler) drops data older than 6 days entirely, so its value jumps when a single-day drop leaves the window, and it may also trigger extreme readings more often than the Wilder version. This is an inference derived from the formulas; no literature quantifying the difference in trigger counts was found.
  • From memory (not verified in this search), Chinese-language software such as Tongdaxin (TDX) often writes the RSI formula as SMA(MAX(C−LC,0),N,1)/SMA(ABS(C−LC),N,1)×100, where SMA(X,N,M) is a recursive average with M/N weighting; with M = 1 it is equivalent to Wilder's α = 1/N, not Cutler's simple average. Implementations in Taiwanese brokers' charting software may differ; if a backtest is meant to match signals from Taiwanese charting software, sample a few dates and compare the values directly.
  • With the Wilder version, values over the first few dozen bars depend on the starting point; if a backtest starts from an ETF's first trading day, the earliest signals may be distorted, so it is advisable to reserve a warm-up period.

Gaps

  • No official documentation was found for the RSI formulas of major Taiwanese charting software (such as XQ or broker apps).
  • No study was found that quantitatively compares how often the Wilder, Cutler and EMA versions of RSI6 trigger < 20 on SPY/QQQ.

Further reading

Frequently asked questions

Does the RSI(2) oversold strategy still work on the S&P 500 over the last decade?

According to the independent Backtrex backtest cited in the original (S&P 500 daily, 2016-10-03 to 2026-10-01, enter when RSI(2) < 5 and the close is above the 200-day moving average, costs included): total return +18.1% vs +255.4% for buy-and-hold; CAGR 1.7% vs 13.5%; maximum drawdown -21.5% vs -35.0%; 73 trades with a 71.23% win rate. The page's own assessment is "profitable, but lags buy-and-hold" with a thin edge.

If buying RSI oversold readings has such a high win rate, why do returns still lose to buy-and-hold?

Published backtests almost uniformly show a win rate of about 70%–80%, per-trade returns of about 1%, holding periods of about a week, and time in the market of about 8%–35%. The original infers that the high win rate comes mainly from the structure of "short holding, small profit target, no stop-loss"; indexes rise over the long run while the strategy is in the market only 10%–35% of the time, so cumulative returns usually lag. A comparison should look at both "per-trade return × share of time in the market" and buy-and-hold over the same period.

Is there any published backtest or academic research on the "RSI golden cross" of RSI6 crossing above RSI10?

The original found none. No published quantitative backtest or academic literature was found on RSI6/RSI12 or RSI6/RSI10 golden crosses on SPY, VOO or QQQ; Chinese-language tutorial sources generally give no sample period, trading costs or benchmark, and the threshold for the "low zone" is inconsistent (20 or 50). The original also infers that such a cross may work much like an ordinary short moving-average breakout, but no literature testing this directly was found.

What does academic research say about technical analysis on equity indexes?

Brock, Lakonishok & LeBaron (1992), using Dow Jones data from 1897–1986, found results that strongly support technical strategies; Urquhart, Gebka & Hudson (2015) retested and found that predictive power weakened after 1986. The Park & Irwin review (2004 working-paper version) found that of 92 studies, 58 were positive, 24 negative and 10 mixed, but most had problems such as data mining. Moskowitz, Ooi & Pedersen (2012) found that index futures returns persist for 1–12 months, an unfavorable backdrop for "buying oversold readings beats buy-and-hold over 1/3/6 months".

What is the difference between Wilder RSI, Cutler RSI and the EMA version of RSI?

Wilder uses a smoothed moving average (SMMA/RMA) with α = 1/n, Cutler uses a simple moving average (SMA) instead, and some commercial software uses a standard EMA. The original infers from the formulas: an EMA has α = 2/(n+1), which at n = 6 is about 0.286, higher than Wilder's 1/6 (about 0.167), so the EMA version of RSI6 is more volatile and hits < 20 more often; but no literature quantifying the difference in trigger counts across the three methods was found.

How do RSI oversold signals fail in a bear market?

The main failure mode is that in a bear market oversold signals recur repeatedly while prices keep falling (RSI gets stuck), so most published strategies add a 200-day moving average filter or a 50/200-day death-cross filter and stop trading during bear markets. For example, in CXO's RSI14 test the largest single-trade loss of -13.9% occurred in mid-2002. The original also notes that no published statistics were found on how many times per year RSI6 < 20 occurs in SPY, VOO or QQQ.

More articles

← All articles Open today's market dashboard →

Contents