Trend-Following and Time-Series Momentum (TSMOM) Timing: Out-of-Sample (OOS) Evidence for the S&P 500 and Nasdaq
Contents10 sections
- 1. Mechanism and time scale: does a trend signal detect the market state, or predict turning points?
- 2. Sample period, OOS period, and whether parameters were chosen ex ante for each method
- 3. Quantitative performance: OOS R², CER, Sharpe, maximum drawdown, hit rate
- 4. Failure periods: repeated false signals, lagging at bottoms, lagging in long bull markets
- 5. Does it beat buy-and-hold after transaction costs and taxes?
- 6. Can it predict turning points, or only reduce drawdowns?
- 7. Applicability to QQQ / Nasdaq
- 8. Free data sources
- Further reading
- Frequently asked questions
Scope: the Faber 10-month SMA, TSMOM (MOP 2012, HOP 2017), moving-average (MA) rules and data-snooping tests, the trend factor / technical indicators, volatility management (risk timing), and direct evidence for the Nasdaq-100 / Nasdaq Composite. The default prior is "timing does not work" (Goyal & Welch 2008; Goyal, Welch & Zafirov 2024).
Site note: This is one of five articles in the series "Ranking the Empirical Evidence on VOO / QQQ Timing Methods" (this one: trend and momentum). It is the site owner's research note published in full: the content, numbers, tables, citations and the "not found / not verified" passages appear as in the original, without cuts or rewrites; only the table of contents at the top of the original draft was removed. The source file is
trend_momentum.mdin the series folder underresearch_notes/(the folder name is in Chinese). This note itself carries no research date; the other notes in the same series are dated 2026-10-26. The default prior throughout the series is "timing does not work" (Goyal & Welch 2008; Goyal, Welch & Zafirov 2024). Tags follow the original definitions: [OOS] = the paper itself ran a genuine out-of-sample or ex-ante (real-time) test; [IS] = in-sample results only; [Verified this time] = the original paper or a first-hand abstract / table was read during the research; [Not verified] = the original was not obtained, and the numbers or bibliographic details come from memory or a secondhand summary. This is a research record, not investment advice.
Tag legend:
- [OOS]: the paper itself ran a genuine out-of-sample or ex-ante (real-time) test
- [IS]: in-sample results only
- [Verified this time]: the original paper or a first-hand abstract / table was read this time
- [Not verified]: the original was not obtained this time; the numbers or bibliographic details come from memory or a secondhand summary and must be checked again before writing the report
1. Mechanism and time scale: does a trend signal detect the market state, or predict turning points?
Takeaway
All the major literature (including AQR on the supporting side) describes trend-following as a lagging mechanism that "lets the market fall for a while, then follows it onto the right side." Returns come from gradual bear markets that last several months (about 15 months on average), and the approach fails in crashes (1987) and V-shaped reversals. This mechanism does not predict tops.
Cited Findings
- HOP 2017 explains why trend-following does well in bear markets: most bear markets "develop gradually over several months rather than suddenly over a few days," letting trend followers build short positions after the initial decline and profit from the subsequent fall. [Verified this time, original text p.20–21, around Exhibit 6] — Hurst, Ooi & Pedersen 2017, JPM 44(1)
- Same paper: for the 10 largest drawdowns of a 60/40 portfolio over 1880–2016, the peak-to-trough period averaged about 15 months; the strategy "may perform poorly in bear markets that happen very quickly (such as the 1987 crash), because there is no time to build positions." [Verified this time] — HOP 2017
- Same paper, the "smile" in Exhibit 5: trend-following did best in years of "extreme gains or extreme losses" for US equities. [Verified this time, text description; the Exhibit itself is a chart and no numbers were extracted] — HOP 2017
- Same paper, Exhibit 2: when trades lag the signal by one month, performance "naturally deteriorates, especially for short-term signals," showing that profits are concentrated shortly after the signal appears. [Verified this time, text description] — HOP 2017
- Secondhand summary: using long-run S&P data, Zakamulin finds that MA rules generate many false signals but mostly win in bear markets. — Zakamulin, SSRN 2677212 "A Comprehensive Look at the Empirical Performance of Moving Average Trading Strategies"
- Zakamulin & Giner (IRFA 2022): the S&P Composite has short-term momentum that can be described by an AR(p) process; their theoretical model also shows that tests of trend-strategy profitability have very low statistical power, so a genuine small edge is hard to distinguish from noise. — Zakamulin & Giner 2022, UiA Brage
- Levine & Pedersen 2016 (FAJ 72(4), "Which Trend Is Your Friend?") show that common trend rules such as TSMOM and MA crossovers can all be written as "a weighted sum of past returns," differing only in the shape of the weights. [Not verified, original not obtained this time; DOI from memory: 10.2469/faj.v72.n4.3] — Levine & Pedersen 2016, FAJ
Inferences
- For a single index such as VOO / QQQ, a trend signal is in practice a lagging detector of "we are currently in a bear state." Its value depends on whether future bear markets fall slowly like 2000–02 and 2007–09, or crash and rebound quickly like 1987 and 2020.
- HOP's positive evidence comes mainly from a futures portfolio diversified across 67 markets, able to short, and volatility-weighted; it cannot be transplanted directly into an "S&P 500 long / cash switch."
Gaps
- HOP's Exhibits are images; table numbers such as per-decade Sharpe ratios and 2010s returns were not extracted.
2. Sample period, OOS period, and whether parameters were chosen ex ante for each method
Takeaway
Truly independent OOS evidence is scarce and falls roughly into three types: (a) extending history backward (HOP 1880–1984, Zakamulin from 1860), which is more credible but is not "post-publication" real performance; (b) examining later periods after a data-snooping correction (STW 1999, Bajgrowicz-Scaillet 2012), all of which are negative; (c) strict real-time tests (Zakamulin 2014 JAM, Cederburg et al. 2020), where the edge is small or negative. Faber's (2007) 10 months was selected in-sample, and no independent academic paper evaluating its OOS performance since publication (2007–2025) was found.
Cited Findings
Faber 2007 (10-month SMA)
- Published in the Journal of Wealth Management (a practitioner-oriented journal, not one of the top JF/JFE/RFS journals) 9(4), 2007. [Bibliographic details from memory: DOI 10.3905/jwm.2007.674809, page numbers not verified] — Faber 2007, JWM
- A criticism relayed on a forum: 10 months was the best parameter backtested on the same data, so the reported results carry data-mining bias and expected future performance should be worse than reported (relayed on a forum, not peer reviewed). — Mutual Fund Observer discussion thread
- Using monthly S&P Composite data for 1860–2009, Zakamulin estimates that the optimal SMA lookback averages 10 months (median 11 months), indirectly supporting that 10 months is not an isolated case, but he also finds that the average returns of all timing rules are lower than buy-and-hold, only with clearly lower risk (secondhand summary). — Relayed from a search summary, UiA Brage master's thesis page / CXO
- A 2012 Ivy Portfolio update mentioned that "long-term MA strategies have struggled and come under scrutiny" (practitioner article). — Investing.com, Ivy Portfolio Update
- No independent academic study evaluating the post-publication OOS performance of the Faber rule over 2007–2025 could be found → insufficient evidence / not verified.
Moskowitz, Ooi & Pedersen 2012 (JFE 104(2))
- 58 liquid futures, roughly 1985–2009; 12-month lookback, 1-month holding period, volatility-weighted. [Not verified, original not obtained this time; DOI from memory 10.1016/j.jfineco.2011.11.003] — MOP 2012
Hurst, Ooi & Pedersen 2017 (JPM 44(1))
- 67 markets (29 commodities, 11 equity indices, 15 bonds, 12 currency pairs), 1880-01 to 2016-12; the signal is an equal-weighted combination of 1, 3 and 12-month trends, following the methods of MOP 2012 and Hurst et al. 2013 (parameters carried over ex ante from the existing literature). [Verified this time] — HOP 2017
- The authors state explicitly that "the data for the first 10 decades are out-of-sample evidence relative to the existing literature," and Exhibit 4 separately reports per-market results for 1880–1984 (before MOP's start year of 1985), which are "also very consistent." [Verified this time] — HOP 2017
- Note: the authors are from AQR (which has a commercial interest in managed-futures products); 1880–1984 is "backward-extension" OOS, not post-publication OOS.
Brock, Lakonishok & LeBaron 1992 → Sullivan, Timmermann & White 1999 → Bajgrowicz & Scaillet 2012
- BLL 1992 (JF 47(5)): DJIA 1897–1986, 26 MA / trading-range breakout rules. [IS; bibliographic details from memory] — BLL 1992
- STW 1999 (JF 54(5): 1647–1691): expands BLL's rule universe and applies White's Reality Check, using about 100 years of daily DJIA data; a book review relays that on the DJIA over the following 10 years (1987–1996), the best BLL rule averaged 8.63%/year (before costs), not statistically significant. Other literature relays that STW and White (2000) find no significantly profitable rule on the DJIA, the S&P 500 index or S&P 500 futures. [Secondhand, table numbers not verified] — STW 1999, LSE Research Online; FMG DP303 working paper
- Bajgrowicz & Scaillet 2012 (JFE 106(3): 473–491, DOI 10.1016/j.jfineco.2012.06.001): daily DJIA data 1897–2011, controlling data snooping with FDR; persistence tests show investors could never pick the future best-performing rules ex ante; even in-sample, very low transaction costs are enough to wipe out performance completely. [Abstract verified this time] — IDEAS/RePEc; CFA Digest
Glabadanidis 2015 → Zakamulin 2018 rebuttal
- Correction: Glabadanidis "Market Timing with Moving Averages" was published in the International Review of Finance 15(3) (2015), not the Review of Finance. [Journal name verified this time] — Zakamulin 2018 rebuttal page
- Zakamulin 2018 (IRF 18(2): 317–327, DOI 10.1111/irfi.12132): Glabadanidis's "too good to be true" performance comes from a one-month look-ahead bias; after correction, MA strategies are "at best slightly better than buy-and-hold and statistically indistinguishable." Data: Ken French's 30 value-weighted portfolios, 1960-01 to 2011-12, one-way switching cost 0.5%, with R code for replication. [Abstract verified this time] — IDEAS; CXO summary
Zakamulin 2014 (JAM 15(4), DOI 10.1057/jam.2014.25): "The real-life performance of market timing with moving average and time-series momentum rules"
- Long-only MA and momentum rules on the S&P Composite, the DJIA, and long- and intermediate-term US Treasuries, 1926–2012; parameters chosen by rolling real-time OOS, with bootstrap to handle data snooping; transaction costs included, taxes not included. [Secondhand summary] — IDEAS; CXO summary
Huang, Li, Wang & Zhou 2020 (JFE 135(3): 774–794, DOI 10.1016/j.jfineco.2019.08.004) "Time series momentum: Is it there?"
- Regressions of "next 1-month return on past 12-month return" for individual assets show weak predictive power both in-sample and out-of-sample; the large t-values of the pooled regression are unreliable under bootstrap critical values; the TSMOM strategy is profitable, but performs about the same as a "historical mean" strategy (TSH) that requires no predictability at all. [Abstract verified this time; table numbers and OOS R² values not verified] — IDEAS/RePEc; WashU page
Kim, Tse & Wald 2016 (JFM 30: 103–124)
- Redoing MOP with similar data, they find the results come mainly from volatility weighting, not from momentum itself; TSMOM alpha falls from 1.27%/month with volatility weighting to 0.41%/month without it, lower than the 0.95% of cross-sectional momentum; most individual contracts do not beat buy-and-hold (31 of 55 win in the weighted version, 29 in the unweighted version). The abstract gives the sample as 1985–2009; one secondhand summary says 1985–2013 and claims the edge disappeared over 2009–2013 (the two accounts differ). [Secondhand summary, table numbers not verified] — bepress author page; Alpha Architect summary
Goyal & Jegadeesh 2018 (RFS 31(5)) "Cross-Sectional and Time-Series Tests of Return Predictability: What Is the Difference?"
- Key point (from memory): TSMOM's advantage over XSMOM comes mainly from TSMOM's implied "time-varying net long position" (essentially bearing the market risk premium); after controlling for it, the two do not differ. [Not verified, original not obtained this time; DOI from memory 10.1093/rfs/hhx131] — Goyal & Jegadeesh 2018
Han, Yang & Zhou 2013 (JFQA 48(5)) / Han, Zhou & Zhu 2016 (JFE 122(2)) / Neely, Rapach, Tu & Zhou 2014 (MS 60(7))
- HYZ 2013: applying MA timing to volatility-decile portfolios, the MA excess return is larger for high-volatility portfolios; the HZZ 2016 trend factor is a cross-sectional stock-selection factor, not index timing. NRTZ 2014: technical indicators (MA, momentum, volume) have OOS predictive power for the equity premium and complement macro variables in recessions (combination forecasts are covered by other researchers). [Numbers all not verified, originals not obtained this time; DOIs from memory: HYZ 10.1017/S0022109013000586, HZZ 10.1016/j.jfineco.2016.01.029, NRTZ 10.1287/mnsc.2013.1838] — HYZ 2013; HZZ 2016; NRTZ 2014
Moreira & Muir 2017 (JF 72(4)) → Cederburg, O'Doherty, Wang & Yan 2020 (JFE 138: 95–117, DOI 10.1016/j.jfineco.2020.04.015)
- See the figures in Section 3. [Original verified this time] — COWY 2020 PDF
Lempérière et al. 2014 "Two Centuries of Trend Following"
- From memory: uses two centuries of data (including equity indices) to show that the trend effect has persisted over the long run, with very high t-values. [Not verified, original not obtained this time; arXiv 1404.3274] — arXiv 1404.3274
Inferences
- Ranked by credibility: studies that "correct for data snooping or choose parameters in real time" (STW, BS, Zakamulin 2014/2018, COWY) consistently reach negative or neutral conclusions; studies with positive conclusions (Faber, Glabadanidis, HOP) respectively suffer from in-sample parameter selection, look-ahead bias, or being multi-asset futures (not a single index).
- 10 months or 12 months can in fact be viewed as "ex-ante conventions": 12 months comes from the Jegadeesh-Titman and MOP literature convention, and 10 months matches Zakamulin's estimated long-run optimum. But the conventions themselves were shaped by early backtests and are not fully independent.
Gaps
- Academic OOS evaluation of the Faber rule after publication over 2007–2025: insufficient evidence.
- Specific OOS R² and table numbers in Huang et al. 2020: not verified.
- Numbers from Goyal & Jegadeesh 2018, NRTZ 2014, the Han series, Levine-Pedersen and Lempérière: none of the originals were read this time.
3. Quantitative performance: OOS R², CER, Sharpe, maximum drawdown, hit rate
Takeaway
OOS figures from reliable sources show that MA / TSMOM timing on a single US equity index improves the Sharpe ratio by about 7–12% after costs (Zakamulin), returns are usually lower than buy-and-hold, and the main benefit is lower volatility and drawdowns; about 80% of "sell periods" lose to T-bills (low hit rate). The volatility-managed market portfolio is actually worse OOS, with Sharpe 0.42 vs 0.46 and CER 1.56% vs 1.75% (COWY 2020).
Cited Findings
- Zakamulin 2014 (JAM) real-time OOS: long-only MA / momentum timing on the S&P Composite and DJIA improves the Sharpe ratio by only 7–12% after costs (secondhand summary, 2013-11 working paper version). [Table numbers not verified] — CXO summary
- Hit rate in the same study: in about 80% of "sell" periods on average, T-bill returns exceeded stock returns, meaning that when the signal tells you to exit, it is mostly wrong; the excess performance is concentrated in a few relatively short periods. [Secondhand (Swedroe); the author cautions that the original should be checked] — ETF.com, Swedroe: Beware Lure Of Market Timing
- Zakamulin's long-run version from 1860: with a one-way cost of 0.25%, MA rules have a higher Sharpe ratio than buy-and-hold but lower average returns. [Secondhand summary] — CXO
- Zakamulin 2018 (IRF): after correcting the look-ahead bias, MA strategies and buy-and-hold are "statistically indistinguishable." [Abstract verified this time] — IDEAS
- COWY 2020 Table 1 (p.99), market factor MKT, 1926-08 to 2016-12 (this is a Moreira-Muir-style in-sample direct comparison): the raw MKT Sharpe is 0.42, the volatility-managed version 0.51, a difference of 0.09, Jobson-Korkie p = 0.30 (not significant). [Original verified this time] — COWY 2020
- COWY 2020 Table 2 and p.101 text: only 53 of 103 strategies have a positive Sharpe difference, binomial test p = 0.84, so "equal performance" cannot be rejected; under the null of "volatility clustering with no risk-return relation," 66 would be expected to be positive, bootstrap p = 0.01. [Original verified this time] — COWY 2020
- COWY 2020 Table 5 Panel A (p.107–108 text) [OOS]: the real-time combination strategy on MKT has Sharpe 0.42 vs 0.46 (raw market) and CER 1.56% vs 1.75%/year, so volatility management is actually worse; the reason is that MKT's in-sample advantage is concentrated around the Great Depression, i.e. early in the sample. Of the 9 factors, MKT, SMB, HML and CMA have negative OOS differences. The abstract concludes that reasonable OOS versions "generally" produce lower CER and Sharpe ratios than the raw portfolios, because of structural instability in the spanning regressions. [Original verified this time] — COWY 2020
- HOP 2017: the average Sharpe ratio per market is about 0.4 (Exhibit 3, before fees and costs); in the ten largest 60/40 drawdowns, returns were positive 8 times (Exhibit 6); allocating 20% of capital to TSMOM lowers maximum drawdown, lowers volatility and raises returns (Exhibit 7). [Text verified this time] — HOP 2017
- HOP's full-period Sharpe ratio after 2/20 fees: ETF.com (Swedroe) gives 0.76; CXO gives about 0.4 averaged per decade after 2%+20% fees. The two conflict, possibly because of different bases (full period vs per-decade average). [Not verified, Exhibit 1 is a chart] — ETF.com Swedroe; CXO
- Practitioner research (not a top journal): S&P 500 1988-07 to 2011-06, buy-and-hold annualized 9.5%, Sharpe 0.31; the best daily MA rule has Sharpe 0.54 (switching cost 0.2%). [IS, the best rule was picked after the fact] — CXO summary
- Journal of Investment Strategies: 1950-01 to 2019-04, buy-and-hold returns are significantly higher than TSMOM, but TSMOM has lower volatility and a better Sharpe ratio. — Risk.net summary
Inferences
- For a single US equity index, the reasonable OOS expectation for trend timing is: returns slightly below buy-and-hold, a small Sharpe improvement (within about 0.05 in absolute terms, or around 10% in relative terms), and a clearly smaller maximum drawdown. This is a change in risk profile, not alpha.
- The OOS evidence for volatility management (risk timing) on a single market index is negative (COWY); it cannot be treated as a robust substitute for trend timing.
Gaps
- OOS R² for single-index TSMOM (Huang et al. 2020 should have individual-asset figures): not verified.
- Maximum drawdown and hit rate of the Faber rule by period: no independent academic figures.
4. Failure periods: repeated false signals, lagging at bottoms, lagging in long bull markets
Takeaway
The literature clearly identifies three failure patterns: fast crashes (1987), V-shaped reversals after sharp falls (lagging at the bottom and buying back higher), and persistent underperformance in long bull markets due to false signals. For academic year-by-year breakdowns of specific years such as 2011, 2015–16, 2018, 2020 and 2022, no first-hand source was found this time.
Cited Findings
- In fast bear markets (such as 1987), the strategy may not have time to build positions. — HOP 2017
- TSMOM's own stress periods often coincide with "rapid simultaneous reversals across many markets" or the absence of clear trends; the historical maximum drawdown was about 25% (Exhibit 8, gross of fee, net of cost). [Text verified this time] — HOP 2017
- About 80% of sell periods lose to T-bills, implying many false signals, with excess returns concentrated in a few short periods. — ETF.com
- A 2022 paper finds that trend strategies failed to beat passive buy-and-hold over the past 10 years (secondhand summary, author and journal not verified). — CXO search summary
- A secondhand summary of Kim-Tse-Wald says TSMOM's edge disappeared over 2009–2013. [Not verified] — Alpha Architect
- Practitioner commentary: the 10-month rule "left a lot of money on the table" during periods when it was whipsawed out. — Commonwealth IMO blog
Inferences
- The long US bull market of 2010–2021, together with repeated sharp falls and sharp rebounds (2011-08, 2015-08 / 2016-01, 2018-Q4, 2020-03), is exactly the environment least favorable to monthly MA rules. This is inferred from the mechanism; no academic year-by-year figures were found this time.
- 2022 was a slow-decline bear market, which in theory favors trend-following.
Gaps
- Academic year-by-year performance of the 10-month SMA on the S&P 500 / Nasdaq for 2011, 2015–16, 2018, 2020 and 2022: insufficient evidence / not verified; you would need to backtest it yourself with free data.
5. Does it beat buy-and-hold after transaction costs and taxes?
Takeaway
After transaction costs, MA timing on a single index at best improves the Sharpe ratio slightly, with returns below buy-and-hold; Bajgrowicz-Scaillet find that low costs are enough to wipe out the performance of technical rules on the DJIA. No peer-reviewed study was found that includes capital gains tax (short-term capital gains) in comparing trend timing against the S&P 500 or Nasdaq. For investors in taxable accounts, this is a material and unquantified disadvantage.
Cited Findings
- Bajgrowicz & Scaillet 2012: low transaction costs can completely wipe out performance (daily rules, DJIA 1897–2011). — IDEAS
- Zakamulin 2014: includes transaction costs, excludes taxes; the Sharpe improvement after costs is 7–12%. — ETF.com; CXO
- Zakamulin 2018: with a one-way switching cost of 0.5%, statistically indistinguishable from buy-and-hold. — CXO
- Searching "after-tax moving average timing S&P 500" found no study that includes taxes; all results are pre-tax. — Related CXO pages (all pre-tax)
Inferences
- A monthly 10-month SMA usually switches only a few times a year, and ETF transaction costs are very low; the real cost is tax. Every sale may realize a gain, and holdings of less than one year are taxed at short-term capital gains rates. In a taxable account, an already small Sharpe advantage is very likely to be offset. This is an inference, with no figures in the literature.
- In retirement or tax-exempt accounts, the tax problem disappears, leaving only costs and tracking differences.
Gaps
- Academic research on the after-tax performance of trend timing: insufficient evidence.
6. Can it predict turning points, or only reduce drawdowns?
Takeaway
Clear conclusion: there is no evidence that trend or TSMOM signals can predict market turning points (tops or bottoms). The existing evidence only supports "reducing drawdowns and volatility in persistent bear markets," at the cost of false signals, lagging at bottoms, and lower long-run returns.
Cited Findings
- The supporters' own mechanism explanation is that short positions are built only "after the initial decline" and fail in crashes (a lagging mechanism). — HOP 2017
- Predictive power for next-month returns is weak both IS and OOS for individual assets; TSMOM's profits can be replicated by a historical-mean strategy that requires no predictability. — Huang et al. 2020
- About 80% of sell signals are wrong (measured by whether T-bills beat stocks). — ETF.com
- The rules have lower risk but lower average returns than buy-and-hold (Zakamulin 1860–2009). — CXO
- Kim-Tse-Wald: MOP's edge comes mainly from volatility weighting (risk management), not from momentum prediction. — bepress
Inferences
- Under the "timing does not work" prior, the most defensible role for trend-following is as a tail-risk and drawdown management tool, not a return-forecasting tool. If the ranking criterion is "raising long-run returns," it should rank near the bottom; if the criterion is "reducing maximum drawdown / improving Sharpe," it is one of the few methods that still has small positive OOS evidence, but the magnitude is small and it faces data-snooping criticism (STW, BS).
Gaps
- No study has evaluated trend signals OOS using a "turning-point hit rate" metric (for example, warning of a top N months in advance).
7. Applicability to QQQ / Nasdaq
Takeaway
No peer-reviewed paper was found that directly runs an OOS test of trend timing on the Nasdaq-100 / QQQ / Nasdaq Composite. Glabadanidis 2015 used Ken French portfolios, not the Nasdaq; the Zakamulin series mainly uses the S&P Composite and DJIA; Zakamulin & Giner 2022 also uses only the S&P Composite. Applicability to QQQ can only be inferred.
Cited Findings
- Zakamulin & Giner 2022 studies only the S&P Composite, not the Nasdaq. — UiA Brage
- Zakamulin 2014 uses the S&P Composite, the DJIA and US Treasuries. — CXO
- Glabadanidis 2015 / Zakamulin 2018 use Ken French's 30 portfolios (sorted by size, B/M and momentum), 1960–2011. — CXO
- A search for academic OOS studies of a Nasdaq 10-month MA found only Nasdaq's official NDXMAX30 drawdown study, which is an index product document, not academic research. — Nasdaq NDXMAX30 Drawdown Study
- Han, Yang & Zhou 2013: MA timing earns larger excess returns on high-volatility portfolios. [Numbers not verified] — HYZ 2013
- Zakamulin has an interactive OOS testing tool that lets you run OOS backtests on an index of your choice (a tool, not a paper). — UiA Shiny OOS test
Inferences
- Possibly more favorable: the Nasdaq's major bear markets (2000–2002, 2008, 2022) were all slow declines lasting several months or more, consistent with the trend mechanism; MA rules may earn larger excess returns on high-volatility assets (the direction of HYZ 2013).
- Possibly less favorable: high volatility means more frequent crossings around the MA and more false signals; in V-shaped reversals (2020) the Nasdaq rebounds faster, so the cost of lagging is larger; the Nasdaq's high long-run return also raises the opportunity cost of sitting in cash.
- The Nasdaq-100 index only began in 1985 (QQQ listed in 1999), so the independent major bear-market sample is essentially just three episodes, 2000–02, 2008 and 2022, giving extremely low statistical power (echoing Zakamulin & Giner's low-power conclusion). Any edge in a QQQ backtest may come mainly from the single 2000–02 episode.
Gaps
- Direct OOS academic evidence for the Nasdaq: insufficient evidence. Clare et al. (for example, the trend research by Clare, Seaton, Smith & Thomas) was not obtained this time, so it cannot be confirmed whether it covers the Nasdaq.
8. Free data sources
Takeaway
The S&P 500 has monthly data from 1871 (Shiller) and a market factor from 1926 (Ken French); the Nasdaq Composite has data from 1971 (FRED / Yahoo); the Nasdaq-100 starts in 1985. The start years below are generally known information and were not checked one by one this time.
Cited Findings
- Zakamulin's OOS tool and his research use long-run S&P Composite data (from 1860). — UiA Shiny; CXO
- COWY 2020 uses the MKT factor from the Ken French data library (from 1926-08). — COWY 2020; Ken French Data Library
- Other candidate sources (start dates not checked one by one this time):
- FRED NASDAQCOM (daily Nasdaq Composite data, from about 1971-02): FRED NASDAQCOM
- FRED NASDAQ100 (from about 1986): FRED NASDAQ100
- FRED SP500 (usually only the last 10 years because of licensing limits): FRED SP500
- Shiller S&P Composite monthly data (from 1871, including dividends): Shiller Data
- Yahoo Finance (^GSPC, ^IXIC, ^NDX, SPY, QQQ): Yahoo ^IXIC
- Stooq (^SPX, ^NDQ, ^NDX, downloadable CSV): Stooq
Inferences
- When backtesting the Nasdaq, use total returns including dividends; a price index understates buy-and-hold because, although Nasdaq dividends are low, they are not zero. For the cash leg, use Ken French's RF or FRED TB3MS.
- To avoid look-ahead bias (the problem Zakamulin 2018 caught), signals must be generated from month-end closes and applied to returns only from the following month.
Gaps
- Exact start dates of each data source and whether they include dividends: not checked one by one.
Further reading
The other four articles in this series:
- Out-of-Sample (OOS) Evidence on Valuation-Based Timing Signals (S&P 500 / US Large Caps; Nasdaq Covered Where Possible)
- Positioning / Sentiment and Risk-Pricing (Option-Implied) Timing Signals: Out-of-Sample Evidence
- Forecast Combination and Machine-Learning Timing: Out-of-Sample and Replication Evidence on the US Equity Premium
- Turning-Point Prediction, Nasdaq-100 / QQQ Specifics and the Realities of Retail Implementation
Related pages on this site:
- VOO returns and technicals (Chinese), QQQ returns and technicals (Chinese), CSPX (Chinese), CNDX (Chinese)
- Watchlist comparison: VOO and QQQ one-year returns (Chinese)
- Quantitative Investing in US Stocks: Before Quant Trading, Quantify the Market Regime
- RSI Oversold Crossovers Buy Low, but Rarely at the Bottom: A Full-History Backtest of QQQ / VOO
- Signal Combinations Can Bracket a Bottom, but Cannot Pinpoint the Low
- Leveraged ETF Return Convergence and the Market-Support Hypothesis: A Test Report
Frequently asked questions
Does 10-month moving-average (Faber) timing still work after publication?
Faber's (2007) 10 months was selected in-sample; this research found no independent academic study evaluating its post-publication OOS performance over 2007–2025, so the evidence is insufficient / not verified. Using 1860–2009 data, Zakamulin estimates that the optimal SMA lookback averages 10 months, but he also finds that the average returns of all timing rules are lower than buy-and-hold, only with clearly lower risk (secondhand summary).
Can trend-following predict market tops or bottoms?
The note's clear conclusion is: there is no evidence that trend or TSMOM signals can predict market turning points. It is a lagging mechanism that "lets the market fall for a while, then follows it onto the right side," with returns coming from gradual bear markets lasting several months (the ten largest 60/40 drawdowns averaged about 15 months from peak to trough), and in crashes like 1987 it may not have time to build positions.
Does moving-average timing beat buy-and-hold after transaction costs?
Zakamulin's (2014) real-time OOS shows that long-only MA / momentum timing improves the Sharpe ratio by only 7–12% after costs, with returns usually below buy-and-hold; in about 80% of "sell" periods on average, T-bill returns exceeded stock returns. Bajgrowicz & Scaillet (2012) find that very low transaction costs are enough to wipe out the performance of technical rules on the DJIA. The benefit is mainly lower volatility and drawdowns, not alpha.
Does volatility management (volatility targeting) work out of sample?
The results of Cederburg et al. (2020) lean negative: only 53 of 103 strategies have a positive Sharpe difference, binomial test p = 0.84; the real-time combination strategy on the market factor has a Sharpe ratio of 0.42 vs 0.46 (raw market) and CER of 1.56% vs 1.75%/year, so volatility management is actually worse.
Can trend timing be used on QQQ / the Nasdaq-100?
No peer-reviewed paper was found that directly runs an OOS test of trend timing on the Nasdaq-100 / QQQ / Nasdaq Composite, so it can only be inferred. The Nasdaq-100 index only began in 1985 (QQQ listed in 1999), so the independent major bear-market sample is essentially just three episodes, 2000–02, 2008 and 2022; statistical power is extremely low, and any edge in a QQQ backtest may come mainly from the single 2000–02 episode.
Do taxes need to be considered in trend timing?
This research found no peer-reviewed study that includes capital gains tax in comparing trend timing; all results are pre-tax. The note's inference is that a monthly 10-month SMA has very low ETF transaction costs, and the real cost is tax: every sale may realize a gain, and holdings of less than one year are taxed at short-term capital gains rates; in retirement or tax-exempt accounts the problem disappears.