Home › Articles

Out-of-sample (OOS) evidence on valuation-based timing signals (S&P 500 / US large caps; Nasdaq covered where possible)

2026-10-26 · Research note · Series: VOO/QQQ market-timing evidence (part 1 of 5) · Not investment advice; figures are as of the sources and dates cited in the text · 中文
Contents11 sections
  1. 1. Baseline verdict: how Goyal-Welch 2008 and GWZ 2024 view valuation variables
  2. 2. Shiller CAPE / CAPE yield (e10/p)
  3. 3. Excess CAPE Yield (ECY)
  4. 4. Fed model / ERP (earnings yield minus bond yield)
  5. 5. D/P, D/Y, E/P, B/M, D/E (Goyal-Welch valuation variables)
  6. 6. Buffett indicator (market cap / GDP)
  7. 7. Household equity allocation (AIAE) and Lettau-Ludvigson cay
  8. 8. How much the Campbell-Thompson 2008 restrictions can rescue
  9. 9. Cross-signal issues: failure periods, trading costs, turning points, Nasdaq applicability, free data
  10. Further reading
  11. Frequently asked questions

Research date: 2026-10-26. Default prior: "market timing doesn't work" (Goyal & Welch 2008 RFS; Goyal, Welch & Zafirov 2024 RFS). This note treats only results with a genuine OOS design (recursive / rolling estimation, real-time data) as "OOS evidence"; results that are in-sample (IS) only are marked "[IS only]". Blog and industry articles are treated only as leads and marked "[not peer-reviewed / lead]". Numbers that could not be checked against the original source are marked "insufficient evidence / not verified".

Site note: This is one of five articles in the series "Ranking VOO / QQQ timing methods by empirical evidence" (this article: valuation). It is the site author's research note published in full: the content, numbers, tables, citations and the "not found / not verified" passages are published as in the original, with nothing removed or rewritten; only the table-of-contents block at the top of the draft was removed. The source file is valuation_signals.md in the author's research_notes folder 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 doesn't work" (Goyal & Welch 2008; Goyal, Welch & Zafirov 2024). In the text, "[IS only]" means only in-sample results exist; "[not peer-reviewed / lead]" means a blog or industry article, used only as a lead; numbers that could not be checked against the original source are marked "insufficient evidence / not verified" (these definitions follow the original). This article is a research record, not investment advice.

Note on number sources: for Goyal-Welch 2008, Campbell-Thompson (NBER WP version) and Asness-Ilmanen-Maloney 2017, the full-text PDFs were downloaded this time and the tables were extracted with pdftotext and checked row by row, so table numbers and figures are relatively reliable. For GWZ 2024, only the variable summary table on the authors' website was obtained (sample through 2020, not the end-2021 sample of the official RFS version); table numbers in the official RFS version were not verified.

1. Baseline verdict: how Goyal-Welch 2008 and GWZ 2024 view valuation variables

Takeaway

In Goyal-Welch 2008's annual and monthly tests, the plain OOS R² of d/p, d/y, e/p, b/m and d/e is negative in every case. GWZ 2024 extends the sample to 2020 / 2021, and in the authors' website summary table the OOS R² of these valuation variables is still negative across the board, with cay at −11.57. Using valuation as grounds for "beating the historical mean by default" does not hold up in either benchmark study.

Cited Findings

  • Conclusion of the GW2008 abstract: most models are unstable or even spurious, and most are not significant even IS; early significance often "rests entirely on the years around the 1973–1975 Oil Shock"; OOS performance is poor in the later part of the sample, not the earlier part — Goyal & Welch 2008, RFS 21(4):1455–1508, abstract / introduction, DOI 10.1093/rfs/hhm014
  • GW2008 Table 1 (annual, S&P 500 log excess returns; OOS starts 20 years after the sample start, with a second set starting in 1965); the valuation variables checked row by row this time:
    • d/p (1872–2005): IS R² 0.49% (not significant); OOS R² −2.06%; OOS R² from 1965 −3.69%
    • d/y (1872–2005): IS R² 0.91%; OOS R² −1.93%; from 1965 −6.68%
    • e/p (1872–2005): IS R² 1.08%; OOS R² −1.78%; from 1965 −1.10%
    • d/e (1872–2005): IS R² −0.75%; OOS R² −4.33%; from 1965 −4.99%
    • b/m (1921–2005): IS R² 3.20%*; OOS R² −1.72%; from 1965 −12.71%
    • 1927–2005 only: d/y IS 2.71%, e/p IS 3.20%, b/m IS 4.14%*, with corresponding OOS R² from 1965 of −6.44%, −3.15% and −19.46%
    • Source: GW2008 Table 1, p.1470–1471
  • GW2008's significance tests use McCracken (2004) MSE-F with bootstrap critical values, not Clark-West. No Clark-West p-values reported by GW2008 for valuation variables were found this time — GW2008 Table 1 notes
  • GW2008 Table 2 (5-year returns, [overlapping observations]): d/p IS R² 10.24%, OOS R² −1.19%, from 1965 −26.09%; d/y OOS −4.45% (from 1965 −16.84%); e/p OOS −1.04% (from 1965 −3.03%); b/m OOS −13.06% (from 1965 −46.34%) — GW2008 Table 2
  • GW2008 Table 3 (monthly, using the Campbell-Thompson procedure, 1927/12–2004/12; risk aversion γ=3, equity position capped at 150%; the CEV column is in %/month, and the original's "a d/p investor loses 10 bp/month" corresponds to the table value −0.10):
    • d/p: plain OOS R² −0.30%; OOS R² after CT truncation 0.17%*; CEV −0.10
    • d/y: −1.12% → CT −0.04%*; CEV −0.14
    • e/p: −1.04% → CT −1.03%; CEV −0.04
    • b/m: −3.28% → CT −2.23%; CEV −0.22
    • e10/p: −2.21% → CT −0.48%; CEV −0.13
    • cay3: −2.05% → CT −0.34%*; CEV +0.06
    • Source: GW2008 Table 3 (around p.1480–1482)
  • GW2008 original text: under CT truncation the OOS R² of d/p is 0.17%, but "an investor trading on d/p would lose 10 bp/month in CEV", and "since January 1992 d/p has consistently predicted a negative equity premium" — GW2008 §5.2 (v)
  • Formal citation of GWZ 2024: Review of Financial Studies 37(11):3490–3557, data through end-2021; it examines 29 new variables from 26 papers plus the original 17; more than one third are not significant even IS, and about half of the rest perform poorly OOS — SFI citation page; SSRN 3929119
  • Valuation rows in the GWZ variable summary table on Ivo Welch's website (labeled the "II" version, with a sample that "includes the original discovery period but extends to 2020"):
    • CSdp (Campbell-Shiller 1988 d/p, monthly): IS coefficient 0.20, t=0.84, IS R² 0.14, OOS R² −0.12
    • CSdy (d/y): t=1.03, IS R² 0.22, OOS R² −0.32
    • CSep (e/p): t=1.72*, IS R² 0.31, OOS R² −1.24
    • CSde (d/e): t=−0.34, OOS R² −1.19
    • KS (Kothari-Shanken 1997 b/m): t=0.95, OOS R² −1.35
    • LL (Lettau-Ludvigson 2001 cay, quarterly): t=0.51, IS R² 0.17, OOS R² −11.57
    • Website conclusion: "most variables have lost empirical support, a few are still passable ... overall predictive performance remains disappointing"
    • Source: ivo-welch.org card-eqpredict
  • The same website summary table has no CAPE / e10p row — ivo-welch.org card-eqpredict

Inferences

  • Under the strict benchmark framework (linear forecasts, recursive estimation, compared with the historical mean), the monthly and annual OOS R² of valuation variables are consistently negative, and OOS after 1965 is worse. This directly supports the default prior.
  • The only one "rescued" is d/p after CT truncation (OOS R² 0.17%). But its CEV is negative, and from 1992 onward it kept telling investors to be bearish, so the tiny statistical positive does not translate into economic value.

Gaps

  • The table numbers for valuation variables in the official GWZ 2024 RFS version and the end-2021 figures: insufficient evidence / not verified (only the 2020 summary table on the authors' website was obtained; the website does not state the unit of OOS R², presumed to be %).
  • Whether GWZ 2024 reports updated figures for e10p (CAPE): not verified.
  • DOI of GWZ 2024: from memory 10.1093/rfs/hhae044, not verified online this time; please cross-check with SSRN 3929119.

2. Shiller CAPE / CAPE yield (e10/p)

Takeaway

CAPE has high explanatory power at the 10-year horizon [IS] (R² about 0.24), but once turned into an actually tradable timing rule, even before costs its Sharpe over 1900–2015 does not beat buy-and-hold (0.37 vs 0.38). Monthly OOS is roughly zero or negative; annual and 5-year results are positive only when counted from 1902, and deeply negative when counted from 1965. No turning-point evidence.

Cited Findings

  • Asness, Ilmanen & Maloney (2017), Journal of Investment Management 15(3):23–40 (S&P 500, reconstructed data before 1926, 1900/01–2015/12):
    • Table 1 [IS only]: Shiller EP (the inverse of CAPE) used to predict future returns. Next 10 years: β 0.8, t 3.4 (Newey-West corrected for overlap), R² 0.24; next 1 year: t 1.9, R² 0.05; next 1 month: t 2.3, R² 0.00
    • The authors also ran OOS forecasts, with results "generally weaker and highly dependent on method and parameter choices", but did not list the numbers
    • Table 2 (tradable timing rule): equity weight = 100% + (trimmed Shiller EP − median) / (95th-to-5th percentile range), bounded at 50%–150%; the percentile window expands from 1881 to 1941, then uses a rolling 60 years; monthly rebalancing; gross of trading costs and fees
      • 1900–2015: excess return 6.6% vs 7.4%; volatility 17.5% vs 20.0%; Sharpe 0.38 (buy-and-hold) vs 0.37 (value timing); max drawdown −83% vs −87%; max relative drawdown −32%
      • 1958–2015: excess return 5.5% vs 5.4%; Sharpe 0.37 vs 0.37; max drawdown −53% vs −51%; average position 89%
      • The return difference (difference-in-means test) and Sharpe difference (Jobson-Korkie test) are not significant in either sample
    • Original text: "Even before costs, the timing strategy cannot beat the Sharpe ratio of buy-and-hold" (around p.26)
    • Failure mechanism: the valuation anchor drifts upward over the long run; the rule flagged "overvalued" (cutting at least 25%) as early as 1991, not in 1996 or 1999, which the authors call "early equals wrong" (around p.28)
    • Table 3 (ex-post adjustment, which the authors themselves call "cheating"): after shifting the signal so the average position is 100%, the 1900–2015 Sharpe is 0.38 vs 0.37, still no win
    • Variants in Appendix Table A.1: "EP extremes only", acting only at extreme values, Sharpe 0.36–0.37 (1900–2015); "EP 10yr lock-in" 0.38; "EP annual rebalance" 0.39; none is a material improvement
    • Table 5 / Table 6: the value + momentum combination (VM) has a Sharpe of 0.43 (1900–2015) and 0.41 (1958–2015); max drawdowns −80% and −43%, versus −83% and −53% for buy-and-hold; gross of costs, and the difference is still not statistically significant
    • Source: AQR full-text PDF; JOIM page
  • GW2008 Table 4 (e10/p, i.e. CAPE yield; recursive OOS):
    • Monthly, OOS from 1965: IS R² 0.49%**; OOS R² −0.83%
    • Annual, OOS from 1902: IS 4.89%**; OOS R² +2.12%**; ΔRMSE +0.30
    • Annual, OOS from 1965: OOS R² −11.45%
    • 5-year [overlapping observations], from 1902: IS 16.47%**; OOS R² −2.85%*
    • 5-year, from 1965: OOS R² −25.65%
    • Source: GW2008 Table 4 (around p.1495)
  • GW2008 Table 3 (monthly, CT procedure): e10/p plain OOS R² −2.21%, after CT truncation −0.48%, CEV −0.13%/month — GW2008 Table 3
  • Campbell & Thompson (NBER WP 11468, 2005 version; sample 1881/01–2003/12, OOS from 1927/01):
    • Table 1 smoothed E/P OOS R²: unrestricted 0.098%; coefficient-sign restriction only 0.236%; non-negative forecast only 0.501%; both 0.538%
    • Table 2: despite the positive OOS R², the monthly utility gain is still −0.085%/month (constant variance, restricted)
    • The original attributes this to "the highly volatile monthly returns of the early 1930s"
    • Table 3 annual: OOS R² 6.643%, utility gain +0.170% (constant variance) or +0.648% (rolling 5-year variance)
    • Source: NBER w11468
  • Siegel (2016) FAJ 72(3):41–50: argues that GAAP accounting changes (for example mark-to-market impairments) depress 10-year average earnings, making CAPE overstated; using NIPA after-tax corporate profits improves forecasts and gives higher future-return forecasts (the direction of the argument is that "CAPE is too pessimistic") — CFA Institute summary page; RePEc
  • The counter-view to Siegel: a study comparing CAPE variants argues that "the original CAPE is still the best overall, including against the NIPA version" [not peer-reviewed / as relayed by Quantpedia; original author and source not verified] — Quantpedia
  • Campbell & Shiller (1988) "Stock prices, earnings, and expected dividends" JF 43(3):661–676; Campbell & Shiller (1998) "Valuation ratios and the long-run stock market outlook" JPM 24(2):11–26. These two are the original basis for CAPE and rely mainly on long-horizon [IS] regressions — citations in the GW2008 references (DOIs from memory are 10.1111/j.1540-6261.1988.tb04598.x and 10.3905/jpm.24.2.11 respectively, not verified online this time)

Inferences

  • CAPE is the valuation variable with the "least bad" OOS results: annual OOS R² is positive from 1902, and monthly is positive with CT restrictions. But the results are extremely sensitive to the OOS starting point; from 1965 they all turn deeply negative, which is exactly the era in which investors could actually use CAPE.
  • The 10-year R² of 0.24 is [IS] and based on overlapping observations, inflated by Stambaugh bias and overlap; it cannot be taken as evidence that a timing rule works. The contrast between Asness et al.'s Table 1 and Table 2 is a demonstration of the gap between IS and a tradable rule.
  • The persistence of the 1995–2000 bubble and the long-elevated CAPE after 2009 are both instances of "early equals wrong"; Siegel's accounting-change argument suggests the CAPE anchor itself may have shifted.

Gaps

  • Clark-West p-values for CAPE timing: no peer-reviewed literature reporting this figure was found this time (insufficient evidence).
  • E10/P figures in the official Campbell-Thompson 2008 RFS version: only the NBER 2005 working paper version (sample through 2003) was checked; the official version's figures were not verified.
  • Dimson-Marsh-Staunton's specific criticisms of CAPE and figures (for example cross-country OOS tests): the original was not obtained this time; not verified.
  • After-tax results for CAPE timing: no peer-reviewed estimate was found.

3. Excess CAPE Yield (ECY)

Takeaway

No peer-reviewed out-of-sample test of ECY could be found. ECY comes from an industry article by Shiller, Black and Jivraj; the only independent academic test found (Catanho & Saville 2022, Investment Analysts Journal) concludes that ECY does not improve on CAPE's predictive power, and it could not be confirmed from the abstract whether it uses an OOS design.

Cited Findings

  • ECY is defined as the CAPE earnings yield minus the 10-year real Treasury yield, proposed by Shiller, Black and Jivraj in the 2020 piece "CAPE and the COVID-19 Pandemic Effect"; that piece says ECY's correlation with the next 10 years' returns is +0.55 [not peer-reviewed / second-hand, not verified] — eco3min.fr
  • Catanho & Saville (2022) find that CAPE peaks when real rates are at 3%–5%; ECY fails to improve CAPE's predictive power — University of Pretoria repository
  • Industry ECY regressions using Shiller's data are single [IS] regressions, not OOS tests [not peer-reviewed] — ii.co.uk

Inferences

  • Conceptually, ECY is the Fed model plus CAPE's smoothed earnings. The closest peer-reviewed OOS analogue is the "EP − real bond yield" in the appendix of Asness et al. in Section 4, whose improvement is likewise not significant and gross of costs.

Gaps

  • OOS R², CER or Sharpe for ECY: no peer-reviewed evidence at all.
  • Whether Catanho & Saville 2022 uses an OOS design: not verified.

4. Fed model / ERP (earnings yield minus bond yield)

Takeaway

The Fed model (E/P minus the nominal yield) was criticized on theoretical grounds by Asness (2003, JPM "Fight the Fed Model"). He points out that comparing a nominal yield with a real E/P commits "money illusion", and that the Fed model has no predictive power for long-run returns. The timing-rule version's result: Sharpe rises from 0.37 to 0.42 over 1958–2015, but gross of costs, with an average position of only 78%, and it is an appendix variant with no significance test.

Cited Findings

  • Asness, Ilmanen & Maloney (2017) Appendix Table A.1 (gross of costs, monthly rebalancing):
    • "EP − Nom. Bond Yield": 1900–2015 excess return 7.8%, volatility 19.9%, Sharpe 0.39 (buy-and-hold 0.38), average position 98%; 1958–2015 excess return 5.4%, volatility 12.7%, Sharpe 0.42 (buy-and-hold 0.37), average position 78%
    • "EP − real bond yield": Sharpe 0.39 (1900–2015), 0.40 (1958–2015)
    • Source: AQR full-text PDF, Table A.1
  • In the main text the authors say that various simple or reasonable approaches "mostly produce similar results (see the appendix)" — same, around p.26
  • Asness (2003) "Fight the Fed Model" JPM 30(1):11–24: the full text was not obtained this time; figures not verified (DOI from memory is 10.3905/jpm.2003.319915).

Inferences

  • Part of the Sharpe improvement over 1958–2015 comes from holding only 78% equities on average, which lowers volatility to 12.7%, so it is not entirely timing skill; this period is also a single big cycle of yields first rising then falling, so the risk of extrapolating the sample is high.

Gaps

  • OOS R² and Clark-West p-values for the Fed model: no peer-reviewed figures were found.
  • Neither GW2008 nor GWZ 2024 tests "E/P − yield" as a single variable (GW has rate variables such as lty and tms, treated separately): based on the tables read this time, no such row was seen.

5. D/P, D/Y, E/P, B/M, D/E (Goyal-Welch valuation variables)

Takeaway

These are all original GW valuation variables. Their OOS R² is almost entirely negative across frequencies and OOS starting points; only CT restrictions pull d/p and e/p up to tiny positive values. b/m is negative in every setting.

Cited Findings

  • For annual, 5-year and monthly OOS figures see Section 1 (GW2008 Table 1 / 2 / 3) — GW2008
  • GW2008's description by period:
    • d/p: slightly poor from 1905 to WWII, good from WWII to 1975, flat until the mid-1990s, poor afterward; more than half of the good OOS performance comes from the Oil Shock; over the last 30 years (1976–2005) IS R² −4.80%, OOS R² −15.14%
    • d/y: large forecast errors in 1958–1965 and 1995–2000; OOS R² over the last 30 years −20.79%
    • e/p: OOS R² over the last 30 years −5.98%
    • Source: GW2008 §3.1, p.1473–1475
  • GW2008 Table 5 (Lewellen / Stambaugh bias correction, 1927/01–2005/12): persistence coefficient ρ is 0.9927 for d/p, 0.9879 for e/p and 0.9843 for b/m, close to a unit root; OOS R² remains negative under each corrected version, for example e/p −1.21% / −0.54% / −0.01%, b/m −2.45% / −1.61% / −0.31% — GW2008 Table 5
  • Campbell-Thompson (NBER WP version) Table 1:
    • D/P OOS R² unrestricted −0.331%, with both restrictions 0.001%
    • E/P: 0.258% → 0.249%
    • B/M (OOS from 1941): −1.328% → −0.988%
    • D/E: −0.287% → 0.008%
    • Original text: B/M "produces negative OOS R² in all four panels"
    • Source: NBER w11468, Table 1
  • Campbell-Thompson Table 2 (γ=3, weights 0–150%, gross of costs), monthly utility gains of the restricted versions: D/P −0.024%, E/P +0.053%, B/M −0.063% (constant variance). Table 3 annual: E/P OOS R² 1.731%, utility gain +0.122%; D/P OOS R² −0.159%, utility gain −0.474%; B/M OOS R² 0.640%, utility gain −0.256% — NBER w11468, Table 2–3
  • Campbell-Thompson's economic translation: an E/P OOS R² of 0.25% relative to a squared monthly Sharpe of 1.2% would raise the average return proportionally by about 21%; for a γ=3 investor, about 1% per year; the original also notes that Table 2 "does not take any transaction costs into account" — NBER w11468, §3, around p.9–11

Inferences

  • E/P is the only variable in this group whose OOS R² and utility are both positive in the CT framework (monthly +0.25%, annual +1%). But the same variable is negative in GW2008's annual test from 1965, and −1.24 in GWZ's website summary table. Results depend on the sample start and the form of restriction, so robustness is insufficient.
  • The ρ of these ratios is close to 1, and Stambaugh bias inflates IS t-statistics; the R² inflation of long-horizon overlapping regressions is another face of the same problem.

Gaps

  • Sub-period OOS results for these variables on the sample added after 2006 (for example 2009–2021): figures in the official GWZ version not verified.

6. Buffett indicator (market cap / GDP)

Takeaway

No peer-reviewed out-of-sample test of market cap / GDP could be found. The closest are an SSRN working paper (Umlauft 2020, long-horizon results, unclear whether OOS) and industry tests. The latter point out that the ratio has no stable mean, so its level is not a useful predictor in itself.

Cited Findings

  • Umlauft (SSRN, 2020): in the US over 1951–2019, the "Buffett ratio" has consistent and significant predictive power over long horizons of about 10 years; it cannot be confirmed from the abstract whether the design is OOS [working paper, not peer-reviewed] — SSRN author page
  • CXO Advisory's test of Wilshire 5000 / GDP: the ratio has no stable mean over the available history, and its level is not a useful predictor [not peer-reviewed / lead] — CXO Advisory
  • A related but different variable: Cooper & Priestley (2013, Review of Finance 17(3):1029–1064) find that the "global capital / output ratio" has strong OOS predictive power in almost every country, and note that financial-market variables "almost never have OOS predictive power" — RePEc (DOI 10.1093/rof/rfs014)

Inferences

  • The denominator of market cap / GDP is domestic GDP, while the numerator includes multinationals' overseas profits; the rising share of overseas revenue at US large caps, plus higher profit margins and changes in listed-company coverage, give the ratio a structural upward trend, the same kind of "anchor drift" problem as CAPE, and more severe.

Gaps

  • Peer-reviewed OOS R², CER, Sharpe: none found at all; should be treated as "no OOS evidence".

7. Household equity allocation (AIAE) and Lettau-Ludvigson cay

Takeaway

AIAE (Aggregate Investor Allocation to Equities) comes from the Philosophical Economics blog (2013); no peer-reviewed OOS test was found. The original "OOS" for cay used full-sample cointegration coefficients (look-ahead bias); when GW2008 switched to a genuinely real-time version, caya, the annual OOS R² fell from 16.78% to −4.33%; the cay OOS R² in the GWZ website summary table is −11.57.

Cited Findings

  • The original AIAE article claims it "predicts long-run returns better than any classic valuation measure" [not peer-reviewed / blog, used only as a lead] — Philosophical Economics, 2013
  • A ten-years-later review relays Micaletti's "Towards a Better Fed Model" as finding AIAE superior to CAPE, Tobin's Q and market cap / GDP; publication status and OOS design not verified [not peer-reviewed / lead] — Portfolio Optimizer blog
  • GW2008 original text: when constructing cay, Lettau-Ludvigson assume agents have ex-post information; their so-called OOS is "very unusual", as agents still know the full-sample CAY construction coefficients. GW call this cayp, and separately construct caya, which does not require knowing the construction coefficients in advance, representing a more realistic OOS experiment — GW2008 §3, around p.1477
  • GW2008 Table 1 (annual, 1945–2005): cayp IS R² 15.72%, OOS R² 16.78%; caya OOS R² −4.33%, ΔRMSE −0.14 — GW2008 Table 1
  • GW2008 Table 2 (5-year [overlapping]): cayp OOS R² 30.35%; caya 9.10% — GW2008 Table 2
  • GW2008 Table 3 (monthly cay3, 1951/12–2004/12): plain OOS R² −2.05%; after CT truncation −0.34%*; CEV +0.06%/month — GW2008 Table 3
  • GWZ website summary table: LL cay IS t=0.51 (not significant), H1 coefficient 2.64, H2 coefficient −0.54 (sign flips in the second half), OOS R² −11.57 — ivo-welch.org
  • Brennan & Xia (2005) "tay's as good as cay" Finance Research Letters 2(1):1–14: argues that cay's predictive power comes mainly from the look-ahead contained in estimating the cointegrating relation on the full sample; replacing it with a time trend, tay, gives similar IS results. The original could not be obtained online this time; the citation and argument are from memory and not verified (DOI from memory is 10.1016/j.frl.2004.10.002). GW2008 cites the Brennan & Xia (2004) working paper for the CEV method — GW2008 §5.3
  • Lettau & Ludvigson (2001) "Consumption, aggregate wealth, and expected stock returns" JF 56(3):815–849 (DOI from memory is 10.1111/0022-1082.00347).

Inferences

  • cay is a textbook case of "look-ahead making OOS look good": for the same variable, simply removing the ex-post coefficients drops the annual OOS R² from +16.78% to −4.33%.
  • AIAE's data come from the Fed Z.1, which has revisions and publication lags (quarterly data roughly 2–3 months behind); any backtest that does not use vintage / real-time data may have the same look-ahead problem.

Gaps

  • Peer-reviewed OOS R², CER, Sharpe for AIAE: not found.
  • Brennan-Xia's specific figures: not verified.

8. How much the Campbell-Thompson 2008 restrictions can rescue

Takeaway

CT's two restrictions (the coefficient sign must match theory, otherwise use the historical mean; the forecast equity premium is truncated at zero) can lift the monthly OOS R² of d/p, e/p and e10/p from negative values to 0.0%–0.54%. But in GW2008's replication, e10/p and b/m are still negative, and the CEV of most valuation variables is still negative.

Cited Findings

  • CT paper: with weak restrictions imposed, many predictive regressions beat the historical mean; although OOS explanatory power is small, it still has economic meaning for a mean-variance investor. Formal citation RFS 21(4):1509–1531, DOI 10.1093/rfs/hhm055 — NBER w11468
  • CT (WP version) Table 1, OOS R² for Panels A to D (OOS from 1927/01; B/M from 1941/03):
    • D/P: −0.331%, 0.001%, −0.043%, 0.001%
    • E/P: 0.258%, 0.264%, 0.243%, 0.249%
    • Smoothed E/P: 0.098%, 0.236%, 0.501%, 0.538%
    • B/M: −1.328%, −1.328%, −0.988%, −0.988%
    • Source: NBER w11468, Table 1
  • CT original text: Table 2 prohibits short selling, so the restrictions affect utility less than R²; the leverage cap and the high average premium limit the utility gains — NBER w11468
  • GW2008's replication using the CT procedure (Table 3, 1927–2004 monthly): OOS R² after truncation is d/p 0.17%, d/y −0.04%, e/p −1.03%, b/m −2.23%, e10/p −0.48%; the original says "even after truncation, ten of the models whose plain OOS R² was negative still have negative CT OOS R²" — GW2008 Table 3 and §5.2
  • Possible reasons CT and GW differ: CT predicts simple returns while GW predicts log returns; the sample end (2003 vs 2004), OOS start and implementation of the restrictions also differ (inferred from the two papers' table notes).

Inferences

  • The CT restrictions mainly keep the model from giving absurd forecasts when estimates are unstable, which amounts to shrinking forecasts toward the historical mean. They can turn "negative OOS R²" into "a positive value close to zero", but do not produce meaningful economic value; the correction to the default prior is small.

Gaps

  • Differences between figures in the official CT version (RFS 2008) and the WP version: not verified.
  • Performance of the CT restrictions on the 2005–2021 extended sample: whether the official GWZ version reports it, not verified.

9. Cross-signal issues: failure periods, trading costs, turning points, Nasdaq applicability, free data

Takeaway

Valuation signals stayed bearish and failed in both 1995–2000 and after 2009. Peer-reviewed tradable rules do not beat buy-and-hold even before costs. No evidence of predicting turning points (tops / bottoms) was found. Nor was any peer-reviewed OOS study found that times the market using Nasdaq-100 valuation.

Cited Findings

  • Failure periods:
    • d/y has large errors in 1995–2000 — GW2008 §3.1
    • d/p has consistently predicted a negative premium since 1992/01 — GW2008 §5.2
    • The CAPE rule started cutting exposure as early as 1991 — Asness et al. 2017, around p.28
    • The valuation anchor drifts upward over the long run during the sample period (the further back the calibration, the lower the neutral CAPE; this is the same phenomenon as the "long-run upward drift" in Section 2), and at the end of the sample CAPE was far above its long-run average, which the authors consider "a particularly unfavorable draw" for value timing — Asness et al. 2017, p.28
    • The early significance of valuation signals is concentrated in the 1973–1975 Oil Shock — GW2008 abstract
  • Accounting changes: GAAP impairment rules depress 10-year earnings, pushing CAPE higher — Siegel 2016 FAJ summary
  • Overlapping observations and Stambaugh bias:
    • The 10-year R² of 0.24 in Asness Table 1 uses Newey-West correction but is still an [IS] result
    • The ρ in GW2008 Table 5 is 0.984–0.993
    • Stambaugh (1999) "Predictive regressions" JFE 54:375–421 is the original reference on the bias; citation in the CT references
    • Source: AQR PDF; GW2008; NBER w11468
  • Trading costs and taxes:
    • All of Asness et al.'s tables are "gross of t-costs and fees"; the authors say momentum has higher turnover and costs, which are not deducted — AQR PDF, p.31–32
    • CT's utility difference can be read as the "trading cost or management fee the investor would be willing to pay each month"; Table 2 itself does not deduct costs — NBER w11468
    • GW2008 Table 3 uses h to mark turnover about 10%/month higher than the benchmark; most valuation variables carry no h mark — GW2008 Table 3 notes
  • Turning points: the three core papers above evaluate only average returns (R², CEV, Sharpe); none of them tests the ability to detect tops or bottoms. Conclusion: no turning-point evidence (limited to the literature read this time).
  • Free data:
    • Shiller's online data (CAPE, S&P earnings and dividends, from 1871) is the data source for Asness et al. and CT
    • The Goyal-Welch / GWZ variable data are provided via Google Drive links on Amit Goyal's personal website — Amit Goyal website
    • Household allocation and market cap / GDP can use FRED / Fed Z.1, with GDP from BEA / FRED; multpl.com is an aggregator site, used only as a convenience source (not primary data)

Inferences

  • Nasdaq-100 / QQQ extrapolation risk (inference, no direct literature):
    • The Nasdaq-100 index only began in 1985, so even at annual frequency there are fewer than 40 independent annual observations, far too few to estimate a 10-year valuation forecasting relationship (the S&P, with about 130 years, barely obtains unstable results)
    • Expensing of R&D at tech companies depresses accounting earnings, making P/E and CAPE systematically high, in the same direction as Siegel's accounting argument and more severe
    • Constituents are concentrated in a few industries and large caps; the history contains only one big bubble, in 2000, a single-event sample
    • Taken together, the S&P 500 valuation evidence should not be applied directly to QQQ
  • What this means for the ranking: valuation signals are weakly positive to negative for "OOS forecasting of average returns"; they do not win on "tradable rule vs buy-and-hold" (gross of costs); and there is no evidence on "turning points". CAPE and E/P have faint value only from specific starting points and with CT restrictions; ECY, the Buffett indicator and AIAE have no peer-reviewed OOS evidence at all.

Gaps

  • Peer-reviewed studies of OOS timing using Nasdaq-100 or Nasdaq Composite valuation: not found (insufficient evidence).
  • After-tax results of valuation timing for taxable accounts: not found.
  • Dimson-Marsh-Staunton's cross-country valuation-timing evidence: the original was not obtained this time.
  • Data URLs such as Shiller's data and FRED series codes: availability not verified online one by one this time.

Further reading

The other four articles in this series:

Related pages on this site:

Frequently asked questions

Can CAPE (the Shiller P/E) be used for market timing?

On the evidence compiled in this note, only weakly. When Asness, Ilmanen & Maloney (2017) turned CAPE into a tradable timing rule, even before costs its Sharpe over 1900–2015 was 0.37 versus 0.38 for buy-and-hold, a difference that is not statistically significant. In Goyal-Welch 2008's annual test, OOS R² is positive only when counted from 1902 (+2.12%); counted from 1965 it is −11.45%. The rule flagged "overvalued" as early as 1991, which the authors call "early equals wrong". There is no turning-point evidence.

Is there out-of-sample evidence for the Excess CAPE Yield (ECY)?

No peer-reviewed out-of-sample test of ECY was found. ECY comes from an industry article by Shiller, Black and Jivraj; the only independent academic test found (Catanho & Saville 2022) concludes that ECY does not improve on CAPE's predictive power, and whether it uses an OOS design was not verified.

What is the out-of-sample R² of valuation measures like dividend yield and P/E?

In Goyal-Welch 2008 Table 1 (annual), the OOS R² is −2.06% for d/p, −1.93% for d/y, −1.78% for e/p, −1.72% for b/m and −4.33% for d/e, and worse from 1965. After GWZ 2024 extended the sample, the OOS R² of the valuation variables in the authors' website summary table is still negative across the board, with cay at −11.57.

Can the Buffett indicator (market cap / GDP) predict the stock market?

No peer-reviewed out-of-sample test could be found. Umlauft (SSRN, 2020) finds predictive power over long horizons of about 10 years in the US over 1951–2019, but it cannot be confirmed from the abstract whether the design is OOS, and it is a working paper; CXO Advisory's test notes that the ratio has no stable mean and its level is not a useful predictor in itself. This note treats it as having "no OOS evidence".

Is the Fed model (earnings yield minus Treasury yield) useful?

Asness (2003) criticizes it for comparing a nominal yield with a real E/P, committing "money illusion". The timing-rule version raises the Sharpe from 0.37 to 0.42 over 1958–2015, but gross of costs, with an average position of only 78%, and it is an appendix variant with no significance test.

Can the S&P 500 valuation-timing conclusions be applied directly to QQQ?

They should not be applied directly. No peer-reviewed OOS study timing the market with Nasdaq-100 valuation was found this time. The note's inference: the Nasdaq-100 index only began in 1985, so even at annual frequency there are fewer than 40 independent annual observations; expensing of R&D at tech companies depresses accounting earnings, making P/E and CAPE systematically high; and the history contains only one big bubble, in 2000.

More articles

← All articles Open today's market dashboard →

Contents