Positioning/Sentiment and Option-Implied Risk-Pricing Timing Signals: Out-of-Sample Evidence
Contents11 sections
- Which variables on this topic does GWZ 2024 cover, and how did they do? (The overall verdict)
- Short interest index (RRZ 2016)
- The Baker-Wurgler sentiment index and HJTZ aligned sentiment (PLS)
- Variance risk premium (BTZ 2009, BMXZ 2014, Drechsler-Yaron)
- SVIX / the Martin 2017 equity premium lower bound and its tests
- Other positioning/sentiment signals (eqis, surveys, NVIX, margin debt, fund flows, AAII, put-call)
- Term spread / credit spread / VIX (only the parts GWZ evaluates)
- Transaction costs, turning points versus average returns, and applicability to Nasdaq
- Free data available to retail investors, and publication delays
- Further reading
- Frequently asked questions
Scope: S&P 500 / the overall US market (Nasdaq is added only where data exists). The prior default is "timing does not work" (Goyal & Welch 2008; Goyal, Welch & Zafirov 2024 RFS, hereafter GWZ). Every number is tagged with its source page or table number; numbers seen only in an abstract or a secondary source are flagged separately; anything that could not be found is marked "insufficient evidence / not verified". Research date: 2026-10-26. The full text of GWZ 2024 (Open Access, RFS 37(11):3490–3557, https://doi.org/10.1093/rfs/hhae044) was downloaded and checked paragraph by paragraph, so the GWZ page numbers and figures in this document are first-hand.
Site note: This is one of five articles in the "VOO / QQQ timing methods ranked by evidence" series (this one: positioning/sentiment and risk pricing). It is the full, unabridged text of the site owner's research note: content, numbers, tables, citations and the "not found / not verified" passages are published as in the original, without deletion or rewriting; only the table-of-contents block at the top of the original draft was removed. The source file is
sentiment_risk_pricing.mdin the author'sresearch_notesfolder for this series (the folder name is in Chinese), research date 2026-10-26 (taken from the original). The default prior for the whole series is "timing does not work" (Goyal & Welch 2008; Goyal, Welch & Zafirov 2024). As stated in the original: every number is tagged with its source page or table number; numbers seen only in an abstract or a secondary source are flagged separately; anything that could not be found is marked "insufficient evidence / not verified". This is a research record, not investment advice.
Which variables on this topic does GWZ 2024 cover, and how did they do? (The overall verdict)
Takeaway
GWZ 2024 covers the four core variables of this topic: short interest (RRZ shtint), HJTZ PLS sentiment (sntm), the BTZ variance risk premium (vrp), and Martin SVIX (rsvix), plus the equity share in new issues from Baker-Wurgler 2000 (eqis). Result: vrp, sntm and rsvix are all explicitly "dismissed"; shtint is the only one still standing (highest OOS R² among monthly variables, 1.26%), but none of its four timing strategies significantly beats "always 100% in equities". eqis looks acceptable both IS and OOS when no publication delay is assumed, but its OOS turns negative once a 6-month publication delay is added.
Cited Findings
- Research design: Re-examines 29 variables from 26 post-2008 papers, plus the original 17 from GW 2008, with data extended to the end of 2021. The main OOS metric is the in-time prevailing OOSCT R² (Campbell-Thompson truncation at zero), benchmarked against the historical mean available at the time; significance uses McCracken 2007 / Clark-McCracken 2001 bootstrapped MSE-F. The OOS period always starts 20 years after the IS start, and no earlier than 1946 (p.3500–3504, Table 3 notes) — GWZ 2024
- Overall conclusion: More than one third of the new variables are no longer significant even IS; of those still significant IS, half perform poorly OOS (abstract, p.3490). Across 184 investment strategies on 46 variables, only three annual variables, accrul, gpce and gip, each have one strategy that significantly beats all-equity-all-the-time, and the t-values are only 1.31 / 1.54 / 1.35; 84% of strategies lose to always being fully invested, and about one third of those even lose money in absolute terms (p.3493) — GWZ 2024
- Classification: GWZ classifies HJTZ
sntmand RRZshtint(monthly) as Sentiment; BHvp, BPSimpvar, BTZvrpand Mrtnrsvixas Variance-related (Table 1 group notes, p.3498–3499) — GWZ 2024 - Table 3 interpretation, original text: "The standout IS/OOSCT performers were, in order of predictive IS and OOSCT performance: shtint, tby, tchi, and d/p"; "vp, impvar, ogap, ndrbl, fbm, dtoat, and rsvix lose IS significance in this synchronized log-equity premium specification" (Table 3 Interpretation, p.3504) — GWZ 2024
- Table 6 count: Of the 46 variables, 13 still hold after the authors' original samples are extended forward and have a non-negative OOSCT R²; among monthly variables only
tchiandshtint, and tby, and possibly lty and tms, also do well (Table 6 notes, p.3547) — GWZ 2024 - CEV for a risk-averse investor (γ=5): No variable statistically significantly beats a 50% equity or a 100% equity allocation; the best performers are quarterly
crdstd(2.52%–5.82% per year relative to the historical mean) and monthlyshtint(1.61%–3.92%) (Section 4, p.3544; Table 5, p.3545) — GWZ 2024 - A reminder that conflicts with the prior: GWZ itself writes "VIX-based predictors have performed extremely well in the most recent 10 years" (p.3493), and over the last 10 years the SVIX coefficient is far above 1, performing "too good", which in fact rejects Martin's original hypothesis (p.3527; Internet Appendix IA.V) — GWZ 2024
Inferences
- With GWZ as the referee, the ranking for this topic is: shtint far ahead of eqis; vrp, rsvix and sntm dismissed. But shtint's "statistically valid OOS" did not translate into "a timing strategy that significantly beats buy-and-hold", so its practical value for VOO / QQQ timing should still be discounted.
- GWZ's OOS design only starts 20 years after the IS start. For variables with short data histories, such as vrp (starting 1990) and rsvix (starting 1996), the OOS window is roughly only post-2010, and results can easily be swayed by the single COVID event in 2020.
Gaps
- The Table 3 columns I extracted with pdftotext are misaligned, so the exact IS R² and OOSCT R² values of each variable cannot be matched cell by cell. This document only uses GWZ's in-text descriptions of each variable (for example shtint's 1.26%, and vrp "positive, but small and not statistically significant"). Other exact values are "insufficient evidence / not verified".
- GWZ does not directly evaluate the original Baker-Wurgler 2006 sentiment index (it evaluates only the HJTZ PLS version and BW 2000's eqis), nor AAII, margin debt, fund flows, put-call, NVIX or the Greenwood-Shleifer surveys.
Short interest index (RRZ 2016)
Takeaway
The signal on this topic with the strongest OOS evidence, and one that survives GWZ's independent replication. However, the effect is much weaker than originally reported (original annual OOS R² 13.24% versus GWZ monthly OOSCT R² 1.26%; the IS t-value falls from −2.15 to −1.79), and the timing strategies do not significantly beat full investment.
Cited Findings
- Mechanism / direction: SII is the log of the detrended, equal-weighted average of short interest as a share of shares outstanding across individual stocks; high SII predicts lower future returns (negative relation). A VAR decomposition shows the predictive power comes mainly through the cash-flow channel, which the authors interpret as short sellers being informed traders — RRZ 2016 JFE; the abstract text quoted by GWZ is at GWZ 2024 p.3530
- Original figures (abstract, as quoted by GWZ p.3530): annual IS / OOS R² of 12.89% / 13.24%; utility gain for a mean-variance investor above 300 bp per year — GWZ 2024 p.3530; RRZ 2016
- Sample: monthly data 1973–2014 (CFA Digest summary); the WashU press release says 1973–2013; CXO's summary of the January 2015 working version says 1973/01–2012/12 and reports an OOS CER gain of 3.5%–4.6% annualized for 1990–2012 (secondary source, working-version figures; original table number not verified) — CFA Digest; WashU Source; CXO Advisory
- GWZ replication: original sample (to 2014) IS t = −2.15; extended forward t = −1.88; homogeneous specification (to 2021) t = −1.79 (still significant); OOSCT R² is positive, the highest of all monthly variables at 1.26%; p(IS,OOSCT) "confidently suggests a predictive association" (p.3530) — GWZ 2024
- GWZ timing strategies: the $1 equal-amount strategies (whether or not long-biased) all lose to all-equity-all-the-time; the Z-value-weighted strategies are "zero or insignificantly better" (p.3531) — GWZ 2024
- GWZ CEV (Table 5, p.3545): relative to the prevailing mean, CEV 3.36% (t = 2.31); relative to a 50% equity allocation 2.25% (t = 1.22); relative to a 100% equity allocation only 0.10% (t = 0.09). The text gives shtint's range as 1.61%–3.92% (p.3544) — GWZ 2024
- Periods of failure / success: it worked mainly from mid-2008 to mid-2011, and in the COVID bull market (Figure 18, p.3531). GWZ concludes "we suggest actively watching its performance forward" — GWZ 2024
- Clark-West p, Sharpe, drawdown and hit rate of the original RRZ: insufficient evidence / not verified (the JFE full-text tables were not obtained).
- Later extension: Liu, Lu, Xu & Zhou (working paper, 2023/12 version, AEA 2024; Zhou is also an RRZ author, so this is not an independent test) combine SII with an option lower bound to get an "SII-based bound": OOS R² of 7.88%, 17.47% and 20.00% across three evaluation periods; Sharpe of 0.63 for 2006–2022 and 1.04 for 2011–2022 (benchmark 0.26); utility gain versus the historical mean of 10.09% per year (introduction, p.2–3) — Liu, Lu, Xu & Zhou, AEA 2024 PDF
Inferences
- From the original annual OOS R² of 13% to GWZ's monthly 1.26%: the frequencies differ (annual uses overlapping returns, which inflates R² by construction), so the two cannot be compared directly. But the IS t-value declines as the sample is extended (−2.15 → −1.79), consistent with a "post-publication decay" pattern that has not yet fully disappeared.
- The "SII-based bound" OOS R² of up to 20% should be viewed with strong suspicion until independently replicated, for three reasons: overlapping authors; an R² that is too high, wildly apart from GWZ's 1.26%; and SII's detrending step introduces look-ahead if it uses full-sample information (whether the original RRZ OOS used recursive detrending is not verified).
Gaps
- No post-2014 dedicated OOS update of SII by independent authors was found; McLean-Pontiff-style post-publication decay studies target cross-sectional anomalies and cannot be applied directly to SII — McLean-Pontiff related summary
- No Nasdaq-specific aggregate short interest timing study was found (insufficient evidence / not verified).
- Whether retail investors can replicate SII in real time: it requires market-wide single-stock short interest data (Compustat or exchange / FINRA semi-monthly data) plus detrending, which is a high practical bar.
The Baker-Wurgler sentiment index and HJTZ aligned sentiment (PLS)
Takeaway
The original BW index was designed for the cross-section, and its OOS predictive power for the overall market is weak (OOS R² about 0.15%). After HJTZ aligned it with PLS, the original OOS R² rose to 1.23%, but once GWZ extended it to 2018–2020 the OOSCT R² is negative, and "sntm always underperforms all-equity-all-the-time"; both IS and OOS have been poor since 2001, and it is dismissed.
Cited Findings
- HJTZ original (RFS 2015, 28(3):791–837): monthly IS / OOS R² of 1.70% / 1.23%; the BW index 0.30% / 0.15% (CFA Institute Digest summary, original table number not verified). Direction: high sentiment → low future returns — CFA Digest; HJTZ 2015
- Sample: the six BW proxy variables, 1965–2010 (546 months) — ideas.repec
- GWZ's critique of HJTZ: sntm is "the six BW 2007 variables + Kelly-Pruitt 2013 in-sample optimized weights", so the IS t-value (2.59 in the original sample, 2.66 extended) is "somewhat meaningless because sntm is optimized to maximize the IS performance"; the OOSCT R² is negative; the conjoint significance of 7% comes mainly from the IS that PLS artificially inflates; "sntm always underperforms all-equity-all-the-time"; Figure 8 shows the IS-optimized performance is good from the start until about 2001, the OOSCT deteriorates almost from the beginning, and after 2001 both IS and OOSCT are poor (p.3517–3519) — GWZ 2024
- A time-series anomaly flagged by GWZ: sntm was "not particularly optimistic" during the 1998–1999 dot-com bubble, yet very optimistic in 1974–1976; "sentiment does not have intuitive time-series behavior" (p.3517–3518) — GWZ 2024
- GWZ half-sample stability: in the GWZ sample, sntm's t-values for the first and second halves are 2.11/1.00 (Table A.1, p.3553) — GWZ 2024
- A nonparametric rebuttal: one nonparametric study concludes "neither S_BW or S_PLS predicts stock market returns or even its volatility" (after allowing for nonlinearity and structural breaks) — Univ. of Pretoria repository (journal of publication not verified)
- Original BW literature: Baker & Wurgler 2006 JF "Investor Sentiment and the Cross-Section of Stock Returns" (https://doi.org/10.1111/j.1540-6261.2006.00885.x) and 2007 JEP "Investor Sentiment in the Stock Market" (https://doi.org/10.1257/jep.21.2.129) are mainly cross-sectional and descriptive evidence; first-hand figures for overall-market OOS R², CER and Sharpe: insufficient evidence / not verified.
- CER gain, Clark-West p and Sharpe of the original HJTZ: insufficient evidence / not verified (the full-text tables were not obtained).
Inferences
- Whether the PLS weights in the original HJTZ were estimated recursively or on the full sample determines whether 1.23% is a true OOS figure. GWZ's in-time reconstruction yields a negative OOS R², and GWZ should be taken as authoritative.
- The BW index is updated with a delay, and its orthogonalization step also rewrites historical values, so real-time predictive power should be worse than in the papers (inference, not verified).
Gaps
- No first-hand figures for the original BW index's OOS performance within the GWZ framework.
- A dedicated OOS study of sentiment indices for the Nasdaq-100: insufficient evidence / not verified.
Variance risk premium (BTZ 2009, BMXZ 2014, Drechsler-Yaron)
Takeaway
In the original paper, VRP had strong predictive power over a quarterly horizon in 1990–2007, but once GWZ extended it to 2021 the IS t-value collapsed to 0.72 (not significant in the homogeneous specification), the half-sample coefficients flipped from positive to negative, the OOSCT R² is positive but small and not significant, and the timing strategies trail full investment by 2.6%–7.0% per year; it is dismissed. The original international evidence (BMXZ 2014) is IS inference, and no OOS R² figure for it was found.
Cited Findings
- Mechanism: VRP = implied variance (model-free VIX²) minus realized variance (high-frequency intraday data); high VRP → high future returns; strongest at a quarterly (about 3 months) horizon (BTZ abstract, as quoted by GWZ p.3513) — BTZ 2009 RFS; GWZ 2024
- GWZ replication: uses the updated vrp from the authors' website (GWZ did not rebuild it); the original sample (to 2007) is strongly significant IS; extended to 2021 the t-value is only 0.72; not significant in the homogeneous specification; the half-sample coefficients flip from positive to negative (Table A.1: authors' sample first half t = 4.29, second half 0.66; GWZ sample first half 1.92, second half −0.05, p.3553); the OOSCT R² is positive but small and not significant; investment performance trails all-equity-all-the-time by 2.6%–7.0% per year; it performed well after the Great Recession and deteriorated in early 2020; the pure OOS result without zero truncation "did far worse, dropping off the chart" (p.3513–3514) — GWZ 2024
- GWZ conclusion: "We dismiss vrp as a useful predictor of equity premiums, primarily based on poor IS performance, secondarily based on poor OOS performance" (p.3513) — GWZ 2024
- GWZ's treatment of BH (Bekaert-Hoerova 2014) vp: vp loses IS significance in the homogeneous specification (Table 3 Interpretation, p.3504) — GWZ 2024
- BMXZ 2014 JFQA: predictive power is strongest at a 2–4 month horizon; Monte Carlo shows finite-sample bias cannot explain this predictability; France, Germany, Japan, Switzerland, the Netherlands, Belgium and the UK show similar patterns, and the panel predictive power of a "global VRP" is stronger. All of the above is IS inference; OOS R² figures: insufficient evidence / not verified — Fed FEDS 2011-52; Cambridge Core JFQA
- Robustness support: an arXiv paper that handles outliers rejects "no predictability" at the 5% level and considers this consistent with BMXZ (IS test) — arXiv 1612.05072
- Drechsler & Yaron 2011 RFS "What's Vol Got to Do with It": a theoretical model explaining why VRP can predict returns; OOS figures: insufficient evidence / not verified — RFS
Inferences
- VRP is a classic case of "publication happened to catch the best period": the authors' sample first half has t = 4.29, and data after that barely supports it. For VOO timing, VRP should be classified as "no longer working / not recommended".
- VRP's main value may be in avoiding post-crisis rebound periods like 2008, i.e. it is event-driven rather than a stable average-return predictor (inferred from GWZ's description of Figure 3).
Gaps
- OOS research on VRP for Nasdaq (VXN² minus NDX realized variance): insufficient evidence / not verified.
- Whether the original BTZ paper ran a formal OOS test, and with what figures: insufficient evidence / not verified.
SVIX / the Martin 2017 equity premium lower bound and its tests
Takeaway
Elegant in theory, a failing grade as an empirical timing signal. GWZ finds an IS t of only 0.33 in the homogeneous specification, a negative OOSCT R², timing strategies that trail full investment by 1.1%–10.4% per year, and notes that Martin's original paper itself in fact "suggests no statistical significance". BCK 2022 cannot reject "the bound is valid" but rejects "the bound is tight", so using it directly as a forecast is not reasonable. Independent OOS by Liu-Lu-Xu-Zhou: the Martin bound's OOS R² is −0.59% for 2006–2022 and 2.14% for 2011–2022 (not significant), with a Sharpe of 0.18 below the benchmark's 0.26.
Cited Findings
- Mechanism: SVIX² is computed from index option prices as a lower bound on the equity premium; Martin argues the bound is approximately "tight", with a time-series average of about 5%, exceeding 20% during the 2008 crisis — Martin 2017 QJE; LSE Research Online
- GWZ: the 1-month rsvix has a 99.5% correlation with VIX², "functionally equivalent to predicting it with the squared VIX"; longer-horizon rsvix still has a 94% correlation with VIX²; original sample (to 2021) IS t = 2.19, extended 2.31; homogeneous specification t = 0.33, not significant; half-sample coefficients flip from negative to positive (Table A.1: GWZ sample −0.63/1.46); the OOSCT R² is negative; timing strategies trail all-equity-all-the-time by 1.1%–10.4% per year; modestly effective from the start until 2007, then broke down badly in the Great Recession (p.3527–3528) — GWZ 2024
- GWZ's rereading of Martin's original paper: "a closer reading of Martin 2017 shows that the paper itself suggests no statistical significance"; the difference lies mainly in the choice of frequency and the setting of the null hypothesis (Internet Appendix IA.V, p.3527) — GWZ 2024
- Back, Crotty & Kazempour 2022 JFE (144:732–760): under conditional tests, that the bound is "valid" cannot be rejected, but "tight" is rejected; because slackness is high, using it directly as a forecast "appears unreasonable in many cases"; adding past average slackness helps, but is limited by data that is too short (as relayed by Liu et al.: about 150 years of data would be needed for a material improvement) — ideas.repec abstract; Liu et al. p.7
- Liu, Lu, Xu & Zhou (2023/12 version): OOS R² (Clark-West test, benchmark the historical mean) for the Martin, BCK and CYL (Chabi-Yo-Loudis, restricted) bounds, for 2006–2022 of −0.59%, −0.54% and −0.76%; for 2011–2022 of 2.14%, 2.80% and 3.25%, the latter two only marginally significant at the 10% level; "predictive power mainly comes from the period after 2008/09"; none of the three bounds consistently beats the benchmark's Sharpe / CER, with Sharpe ratios of 0.18, 0.22 and 0.24 versus a benchmark of 0.26, worse during the 2008/09 crisis, with negative CER (introduction, p.2–3) — AEA 2024 PDF
- Liu et al. relaying the GWZ 2021 working version: for option bounds "the out-of-sample performance is never statistically significant" (p.7) — AEA 2024 PDF
- Chabi-Yo & Loudis 2020 (JFE, "The conditional expected market return"): after using the option lower bound as a constraint in predictive regressions, the abstract claims "considerable out-of-sample gains" relative to unrestricted regressions and forecast combination. Specific OOS R² / CER figures: insufficient evidence / not verified; DOI not verified — lead found in the Back-Crotty-Kazempour search result
- Kadan & Tang 2020 RFS "A Bound on Expected Stock Returns" (33(4):1565–1617): a bound at the individual-stock level; OOS figures for overall-market timing: insufficient evidence / not verified — ideas.repec
- Performance over the last 10 years: GWZ notes SVIX / VIX-type variables have been "too good" over the last 10 years, with coefficients far above 1 (p.3493, p.3527); Liu et al. also find the 2011–2022 OOS R² turned positive. The two agree: effective recently, but failing over the full period and in crises — GWZ 2024; Liu et al.
Inferences
- The SVIX lower bound gives a "high expected return" when volatility is high, which amounts to asking investors to add during a crash, and the 1-month horizon is short. The 2008 experience shows this leads to adding midway through the decline and suffering a larger drawdown, the opposite of the goal of "improving drawdown" (inferred from the Sharpe results of GWZ and Liu et al.).
- Because the correlation between rsvix and VIX² is as high as 99.5%, retail investors using VIX directly (VXN for Nasdaq) get the same signal, with no need for raw SVIX data.
Gaps
- OOS R², hit rate and drawdown of Martin's original paper: insufficient evidence / not verified.
- An SVIX bound computed from NDX options for QQQ timing: insufficient evidence / not verified.
Other positioning/sentiment signals (eqis, surveys, NVIX, margin debt, fund flows, AAII, put-call)
Takeaway
Apart from BW 2000's eqis (evaluated by GWZ; OOS turns negative under the delay assumption), no verifiable positive figures were found for any signal in this category within an independent OOS framework of GWZ's caliber. The "negative relation" of the Greenwood-Shleifer surveys and the "high returns after high NVIX" of NVIX are mainly IS evidence.
Cited Findings
- eqis (Baker & Wurgler 2000 JF, "The Equity Share in New Issues and Aggregate Stock Returns", https://doi.org/10.1111/0022-1082.00285): the mechanism is a high equity share in new issues → low future returns. GWZ calendar-year (December) specification: coef −5.54, t −2.73, IS R² 8.17%, OOSCT R² 1.49% (Table 3 Panel C); with a 6-month publication delay added (Panel D), IS R² 4.05%, OOSCT −6.99% (Table 3, p.3503–3504; the Panel C row appears aligned, but Panel D's PDF extraction layout is misaligned, so the exact values still carry a risk of mistranscription); GWZ's original text states: "Without reporting delay, eqis performed well, too" (p.3504). It was strongest in the 1970s oil crises, weakest in 1998–2009, and "largely nondescript" after 2009 (p.3543); the investment strategies "performed about the same as all-equity-all-the-time" (p.3549); CEV relative to the prevailing mean 1.26% (t 1.25), relative to a 50% equity allocation −2.06% (Table 5, p.3545) — GWZ 2024
- Greenwood & Shleifer 2014 RFS "Expectations of Returns and Expected Returns": six investor-expectation surveys (including AAII, Gallup and others) are highly correlated with one another, positively correlated with past returns, and negatively correlated with model-implied expected returns and with actual future returns. This is IS regression evidence; OOS R² / CER figures: insufficient evidence / not verified (the full text was not obtained in this search) — RFS DOI
- Manela & Moreira 2017 JFE "News Implied Volatility and Disaster Concerns": builds NVIX from WSJ front-page text, going back to 1890; in postwar US data, a 1 standard deviation increase in NVIX predicts annualized excess returns 3.3 percentage points higher over the next 1 year and 2.9 percentage points higher per year over the next 2 years (press release and Quantpedia summary). Whether an OOS test was done: insufficient evidence / not verified — WashU Source; Quantpedia; JFE DOI
- Margin debt, mutual fund flows, the AAII bull-bear spread, the put-call ratio: no first-hand strict OOS (OOS R² / Clark-West / CER) figures for the S&P 500 from top journals were found this time; GWZ 2024 does not include them either. Insufficient evidence / not verified.
Inferences
- The academic evidence for surveys (AAII) is "contrarian, and holding only IS", and weekly surveys are very noisy, so they are not recommended as a standalone timing rule (inference).
- eqis is an annual variable with a publication delay; once the delay is added in real time its OOS is negative, so it is no longer useful in practice.
Gaps
- OOS academic evidence on margin debt (FINRA monthly data), ICI fund flows and the put-call ratio needs a separate search; no verifiable source was found within this session's tool-call budget.
- Whether NVIX data is still being updated, and how it performs OOS: insufficient evidence / not verified.
Term spread / credit spread / VIX (only the parts GWZ evaluates)
Takeaway
In GWZ, the Treasury-yield variables (tby, and possibly lty and tms) do reasonably well; GW 2008's dfy (default yield spread) is not significant in the homogeneous specification; among the new variables, crdstd (credit standard deviation, quarterly) stands out on CEV. VIX-type variables have been very good over the last 10 years but are poor over the full period (rsvix is dismissed).
Cited Findings
- "The Treasury-yield variables tby, and possibly lty, and tms also did well" (Table 6 Interpretation, p.3547); "The standout IS/OOSCT performers were... shtint, tby, tchi, and d/p" (p.3504) — GWZ 2024
- dfy and dfr have very low IS t-values in the homogeneous specification of Table 3 Panel A (the extracted text reads about 0.12 and 0.91; because the columns are misaligned, the exact values are not verified) — GWZ 2024 Table 3
- crdstd: CEV relative to the prevailing mean 5.82% (t 2.71), relative to a 50% equity allocation 1.74% (t 0.79) (Table 5, p.3545); GWZ considers its value "likely due to their intelligence about prevailing risk", rather than predicting average returns (p.3545) — GWZ 2024
- VIX: rsvix has a 99.5% correlation with VIX²; the conclusion is the same as in the SVIX section (p.3527) — GWZ 2024
Inferences
- The CEV advantages of shtint and crdstd come mainly from "risk information" (avoiding high-variance periods) rather than return prediction, a mechanism similar to the effect of volatility targeting (inferred from GWZ p.3545).
Gaps
- Exact OOSCT R² for tms and tby: because of the misaligned table extraction, insufficient evidence / not verified.
Transaction costs, turning points versus average returns, and applicability to Nasdaq
Takeaway
All the literature on this topic (GWZ included) is about "average return prediction", not "turning-point (top and bottom) identification"; transaction costs and taxes are almost never included; dedicated Nasdaq evidence is lacking.
Cited Findings
- Average returns versus turning points: GWZ's metrics are all OOS R² (MSE improvement), timing strategy returns and CEV (p.3495–3496), with no turning-point or hit-rate metrics; HJTZ, RRZ, BTZ and Martin likewise all use a predictive-regression framework — GWZ 2024
- GWZ points out that OOS R² and timing performance are not equivalent: "OOSCT performance does not map one-to-one into investment performance"; making only tiny position adjustments might beat full investment, but "natural" timing strategies cannot (p.3493 note 2) — GWZ 2024
- The Campbell-Thompson threshold: a monthly OOS R² of 0.5% is already economically meaningful (as cited by Liu et al., p.3 note 2) — AEA 2024 PDF
- Transaction costs / taxes: none of GWZ's four timing strategies or CEV calculations deduct transaction costs or taxes (the text does not mention cost adjustment, p.3493–3496, p.3544–3545). Post-cost results from other papers: insufficient evidence / not verified — GWZ 2024
- Drawdown: no source provides maximum drawdown comparisons for these signals (insufficient evidence / not verified). Qualitative evidence: rsvix and vrp broke down in 2008 or early 2020 (GWZ p.3513, p.3528).
- Nasdaq: no top-journal OOS studies of VXN-VRP, Nasdaq SII or Nasdaq sentiment were found (insufficient evidence / not verified).
Inferences
- Since all the signals are monthly, average-return type, and shtint's CEV versus full investment is only 0.10%, the expected net advantage after VOO / QQQ taxes and spreads should be close to 0 or negative (inference).
- QQQ has higher beta and volatility than VOO; variance-based signals (vrp, SVIX) could in theory be swapped for VXN versions, but there is no empirical evidence, so this should be treated as unverified.
Gaps
- Figures from turning-point tests (for example Pesaran-Timmermann directional accuracy): none of the sources on this topic provide them.
Free data available to retail investors, and publication delays
Takeaway
VIX / VXN are free, daily and real-time, and the easiest to obtain; SII requires aggregating single-stock short interest and detrending it yourself, a high bar; BW sentiment, HJTZ and VRP are provided on the authors' websites, but updates are delayed and historical values may be rewritten; the GWZ dataset itself can also be downloaded from Goyal's website.
Cited Findings
- How GWZ handles data availability: vrp (and BH vp) directly use the series "updated regularly" on the authors' websites (p.3494 note 4, p.3513); sntm can only be extended to 2018–2020, not to 2021 (p.3500 note 6; Table 1 gives sntm as 1965:07–2018:12); rsvix 1996:01–2021:12; vrp 1990:01–2021:12 (Table 1, p.3498–3499) — GWZ 2024
- The effect of eqis's publication delay: with a 6-month delay added, the OOSCT R² goes from 1.49% to −6.99% (Table 3 Panel C / D) — GWZ 2024
- Data URLs (listed from known author homepages; each was not opened this time to confirm its current state and update frequency, so they are marked not verified):
- Baker-Wurgler sentiment index: Jeffrey Wurgler's NYU homepage https://pages.stern.nyu.edu/~jwurgler/
- VRP: Hao Zhou's personal homepage https://sites.google.com/site/haozhouspersonalhomepage/
- SVIX: Ian Martin's LSE homepage https://personal.lse.ac.uk/martiniw/
- VIX / VXN: FRED series VIXCLS https://fred.stlouisfed.org/series/VIXCLS , VXNCLS https://fred.stlouisfed.org/series/VXNCLS ; the CBOE website
- Single-stock short interest: FINRA Equity Short Interest (semi-monthly) https://www.finra.org/finra-data/browse-catalog/equity-short-interest
- AAII sentiment survey (weekly): https://www.aaii.com/sentimentsurvey
- GWZ predictor dataset: Amit Goyal's homepage https://sites.google.com/view/agoyal145
Inferences
- The only signals that are truly usable in real time with zero delay are VIX / VXN (and the SVIX approximation derived from them). And the academic evidence shows these signals fail over the full period and work only over the last 10 years.
- SII is the signal with the best evidence, but the most costly to replicate in real time (FINRA short interest is published semi-monthly, about 1–2 weeks after the settlement date, and it also requires market-wide aggregation). The exact number of days of delay is not verified.
Gaps
- The latest update month on each author's website (for example, through which year the BW index is updated, and whether Hao Zhou's VRP is still being updated): not checked this time; insufficient evidence / not verified.
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)
- Trend-following and time-series momentum (Trend-Following / TSMOM) timing: out-of-sample (OOS) evidence for the S&P 500 and Nasdaq
- Forecast combination and machine-learning timing: out-of-sample and replication evidence on the US equity premium
- Turning-point prediction, what is special about the Nasdaq-100 / QQQ, 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)
- Custom comparison: VOO versus QQQ one-year returns (Chinese)
- Quant investing in US stocks, a primer: before quant trading, quantify the "market state" first
- RSI oversold golden crosses buy low, but rarely at the bottom: a full-history QQQ / VOO backtest
- 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
Can short interest be used for market timing?
In GWZ 2024's independent replication, short interest (shtint) is the only signal on this topic still standing, with a monthly OOSCT R² of 1.26%, the highest of all monthly variables; but the effect is much weaker than the originally reported annual OOS R² of 13.24%, its CEV relative to a 100% equity allocation is only 0.10% (t = 0.09), and the timing strategies do not significantly beat full investment. Replicating it in real time is also a hurdle for retail investors (it requires market-wide single-stock short interest data plus detrending).
Can the VIX be used as an entry and exit timing signal?
GWZ points out that the 1-month rsvix has a 99.5% correlation with VIX², making it functionally equivalent to VIX²; in the homogeneous specification the IS t is only 0.33, the OOSCT R² is negative, and the timing strategies trail full investment by 1.1%–10.4% per year. However, GWZ also writes that VIX-type variables have performed "too good" over the last 10 years. The note's conclusion: effective recently, but failing over the full period and in crises.
Does the variance risk premium (VRP) still have predictive power?
Once GWZ extended it to 2021, VRP's IS t-value fell to just 0.72, the half-sample coefficients flipped from positive to negative, the OOSCT R² is positive but small and not significant, and investment performance trails full investment by 2.6%–7.0% per year; it is explicitly dismissed.
Can the investor sentiment index (Baker-Wurgler) predict the market?
The original BW index has an OOS R² of about 0.15% for the overall market; after HJTZ aligned it with PLS, the original OOS R² was 1.23%. But once GWZ extended it, the OOSCT R² is negative; the original text says "sntm always underperforms all-equity-all-the-time", and after 2001 both IS and OOS are poor.
Is there out-of-sample evidence for AAII retail sentiment, margin debt or the put-call ratio?
No first-hand strict OOS (OOS R² / Clark-West / CER) figures for the S&P 500 from top journals were found this time, and GWZ 2024 does not include them either; this is insufficient evidence / not verified. The academic evidence for surveys (including AAII) is contrarian, and holds only IS.
Can these sentiment and risk-pricing signals catch tops or bottoms?
That conclusion cannot be drawn from the literature. All the literature on this topic (GWZ included) is about "average return prediction", evaluated with OOS R², timing strategy returns and CEV, with no turning-point or hit-rate metrics; transaction costs and taxes are almost never included, and dedicated Nasdaq evidence is lacking.