Valuation and the allocation: what a CAPE of 41 does and does not license

Question. US CAPE is at the 99th percentile of 145 years. Should that change the equity share, the US/international split, or the rebalancing rule?

Decision it informs. Whether any allocation number moves because of an August 2026 valuation reading, and if so which one and through which mechanism. It does not set the equity share — setting the equity share owns that — and it forecasts no market.

Out of scope. Trend and moving-average timing, which is a different signal and a different page. Sector or single-stock valuation. Any claim that a level “must” mean revert.

as of 2026-08-22. Measured figures regenerate from studies/valuation_conditioning.py and its cache companion _valuation_conditioning_tables.py, run with uv run python -m portfolio_edge.studies.valuation_conditioning; the arithmetic is pinned in research/tests/unit/test_studies_valuation_conditioning.py. Everything measured here is exploratory: no specification was frozen before the numbers were seen and no experiment is registered, so nothing below may support a promoted claim.

---

Conclusion

1. The concern is correct about the level and wrong about what follows from it. US CAPE of 41.18 (as of 2026-08-01, Shiller) has been equalled or exceeded in 19 of 1,748 months since 1881, and 18 of those 19 are March 1999 to September 2000. The 1929 peak was 32.56. But a level is not a forecast, and the three claims bundled inside “we must consider valuations” — a return forecast, a risk statement, and a relative call — have very different evidence. Only two of them survive.
2. The forecast claim is much weaker than its usual presentation, and the weakness is in the standard errors, not the slope. At the ten-year horizon the CAPE-yield regression reports a Newey-West t of 4.84 and a Hodrick 1B t of 2.47 on the same coefficient. The sample holds 13.6 independent ten-year observations, not 1,628. Separately, the predictor’s monthly autoregressive root is 0.9966 and its innovation correlates −0.9975 with the return innovation, so the Stambaugh bias is 73.4% of the fitted slope: annualised, 0.0369 uncorrected against 0.0098 corrected. Neither correction is exotic and both are omitted from almost every published version of this chart.
3. Out of sample since 1990 the CAPE model has lost to a rolling mean at every horizon, and lost badly. Out-of-sample R**2 of −0.21 (1yr), −0.42 (5yr), −0.44 (10yr), −1.88 (15yr), with the model too pessimistic by +7.8, +6.3, +4.9 and +4.3 pp/yr respectively. An investor who acted on it in 1990 would have been wrong for a third of a century. That is not proof it is wrong now, but it is the record.
4. Conditioning the equity share on the CAPE level loses even before costs. Over 1921-01…2026-07 a rule that tilts 80/20 on the expanding CAPE percentile returned −9.0 bp/yr against a constant 80/20 gross, and −86 bp/yr net of 10 bp execution and a 15% effective capital-gains rate. “Halve above the CAPE median” lost −55 bp gross and −202 bp net. Both used a revised, non-point-in-time history, which flatters them.
5. One valuation rule does have a real gross edge, and tax eats it. Tilting on the excess CAPE yield rather than the level earned +49.5 bp/yr gross against a constant 80/20, of which +47.7 bp is timing rather than de-risking, and it was ahead in 98.5% of rolling 30-year windows. Net of execution and a 15% effective capital-gains rate it turns −15.6 bp, ahead in 39.4% of windows. The break-even effective capital-gains rate is 11.3%. The signal is real and the account decides whether it is worth having.
6. The risk claim survives, and it is the strongest thing valuation says. Buyers at CAPE above 30 spent a median 59.7% of the following fifteen years below their own real entry level, against 5.0% for buyers below 20, with a median worst real drawdown of −51.8% against −36.7%. That is a statement about holdability, not about return — the same cohort’s fifteen-year real return was still positive at a median +2.15%/yr and its twenty-year median was +4.31%/yr. It is drawn from two episodes (1929 and 1997-2002, 0.32 independent observations) and cannot be read as a probability.
7. The relative call is the best-posed of the three, and the strongest thing in it is not a CAPE reading. It is that 81% of thirty-five years of US outperformance was re-rating rather than earnings growth (+3.8 pp of 4.7 pp/yr, AQR, as of 2024-12). The US CAPE premium over developed ex-US is 1.70x (Siblis, as of 2026-06-30), it is about fifteen years old, and in the late 1980s it was inverted. Cross-sectionally on 1870-2020 and 18 countries a one-log-unit valuation gap bought +6.05 pp/yr of relative ten-year real return with t = 3.65 on 14 independent observations — and the relation is undetectable after 1990 (slope −0.0006, t = −0.03) on a design that could have found the historical 0.07 at 80% power. AQR’s own version of the same test reports a +0.5 correlation on “4+ independent observations.”
8. The 65/35 split is not a US overweight. Siblis’s global index is ~64% US (as of 2026-06-30), so 65/35 is a +1.0 pp active bet against global market-cap weight. The question is therefore not “should we cut an overweight” but “should we deliberately underweight the market portfolio”, which needs a stronger claim than “expensive”.
9. Every US-versus-international measurement here has a confound the size of the effect. Buybacks close about half the US-versus-Europe payout gap and explain about a third of the US’s own CAPE elevation — and correcting for them does not forecast better. Sector mix explains about half of the US’s relative richness. Currency contributed ≈−1.7 pp/yr to unhedged non-US returns over 2010-2024, ≈+7 pp in 2025 alone, and ≈−1.6 pp in 2026 H1. And Japan, Korea and Taiwan now carry CAPEs at or above the US’s.
10. What to do. Do not run a valuation-conditional weighting rule in a taxable account. Do not cut the equity share on the return forecast. Do widen the drawdown and drought assumptions the equity share is sized against, and do direct new contributions rather than realised gains toward the international side. A 10 pp US-to-international shift is worth about 14 bp/yr against 80-144 bp of tracking error and needs 55-178 years to demonstrate. It is not worth a tax bill; it is nearly free with new money. Hedging currency is worth ≈0.8 pp/yr of carry to a USD investor and has not been tested here — it is larger than the valuation edge and should be studied before the split moves.

---

1. What valuations actually are

1.1 The US level, from this repository’s own cache

Shiller’s ie_data workbook, sha256:71c3636d…, Last-Modified: Tue, 04 Aug 2026 15:29:32 GMT, refetched 2026-08-23 and byte-identical. Its final row is 2026-08 and its own footnote says “Aug price is Aug 1st close … Aug GS10 is Jul 31st value”, so this is an August 1st reading.

The last two rows are the finding. They are the same construction with a different real rate, and they disagree by nearly a percentage point, because the real yield implied by Shiller’s own column is 1.45% where the market prices 2.35% (2026-08-20) to 2.41% (the Aug 1-20 average). On Shiller’s measure today sits at the 19th percentile of 145 years and is unremarkable. On a market-priced real yield the excess CAPE yield is +0.02 pp on the August average and +0.08 pp on the latest daily reading — the lowest level of the entire 23-year TIPS record, and it has now sat at that 0th percentile for four consecutive months.

For scale on how far this has moved: the TIPS-based measure averaged 2.82 pp over 2003-2026 with a standard deviation of 1.25, peaked at 5.80 pp in 2009-03, and by annual average has gone 2.83 (2022) → 1.72 (2023) → 0.94 (2024) → 0.74 (2025) → 0.52 (2026).

The “low rates justify a high CAPE” defence has expired. It was a good argument in 2021 and it is not one now.

Both prior CAPE peaks had a negative Shiller excess CAPE yield: −1.09 pp at 1999-12 and −0.61 pp at 1929-09. On that measure today is less extreme than either. On the market-priced measure the gap has closed to eight basis points.

1.2 The same level from the web, and why the readings differ

Four “current” CAPE readings disagree: 41.18, 41.58, 41.96 and 40.4. Scaling Shiller’s 41.178 from its 7,600.50 August 1st close to 7,674.37 on August 21 gives 41.58, which is what GuruFocus published for August 2026 — so the staleness explains part of the gap and the construction explains the rest. multpl’s 41.96 is not reproduced by that scaling. The test test_rescale_cape_reconciles_the_workbook_with_an_independent_reading pins this. No page may quote these interchangeably, and the difference between 41.2 and 42.0 changes no decision below.

The tension that matters is between CAPE and the forward multiple. CAPE is at the 99th percentile and trailing P/E is 29.6x, yet forward P/E is 20.0 and fell over 2026, because forward earnings rose 24.9% year-to-date against a 12.1% price gain (Yardeni, 2026-08-22). FactSet reports Q2 2026 blended earnings growth of 50.4% (32.0% excluding Alphabet and Amazon) and an aggregate earnings surprise of 29.2%, the highest since it began tracking in 2008, which it flags as “heavily influenced by the two tech giants’ gains related to equity investments and valuations” — mark-to-market on equity stakes, not operating income. A ten-year smoothed denominator does not get that relief; a forward multiple gets all of it. Which of those two is right is the question, and this repository cannot answer it.

1.3 International

Source: Siblis Research and its world CAPE page, read 2026-08-22. US premium: 1.70x over developed ex-US, 1.85x over EM, computed inside one methodology. Siblis’s own US CAPE is 35.82, not 41.96; dividing multpl’s US figure by Siblis’s ex-US figure would give a spurious 2.00x.

Two qualifications that cut against the simple story. Japan, Korea and Taiwan are at or above US levels — the developed-ex-US discount is concentrated in Europe, the UK, Australia and Greater China. And AQR reports that while “the U.S. CAPE ratio of nearly 40 is at the 96th percentile since 1980, […] non-U.S. global developed CAPE is near the historical median”, the YE2025 MSCI EM CAPE of 23 sits at the 98th percentile since 2001 (AQR 2026 Capital Market Assumptions, as of start of 2026). EM is cheap against the US and expensive against itself.

1.4 Published forecasts, labelled as forecasts

Not measurements. Each is a model output on that firm’s assumptions, and AQR states there is a 50% chance realised ten-year returns fall outside its own error bars.

Every forecaster with a public number ranks US large-cap last. They do not agree on anything else: GMO’s US large-cap forecast is 12 pp below AQR’s, and GMO and Vanguard put EM below the US while AQR puts it above. AQR’s global 60/40 expected real return is 3.4%, about 1.5 pp above its 2021 low and still well below the long-run US average. Research Affiliates reportedly forecasts 3.1% nominal for US large cap against 4.7% for US aggregate bonds, but that came through a search snippet only and is unverified.

1.5 Sources that could not be reached

Recorded as evidence about the data contract, not as an excuse.

• FRED returned HTTP 403 to the web-fetch tool and to curl, but answered the research workspace’s own requests client normally on 2026-08-23. The block is on the fetch path, not on the source, and DGS10/DFII10 remain usable from research/.
• Research Affiliates Asset Allocation Interactive has moved: interactive.researchaffiliates.com/asset-allocation now 301-redirects to interactive.syzygyassetmanagement.com, which renders nothing without JavaScript. It is not usable as a programmatic source.
• 403 or paywalled: GuruFocus, Morningstar (both editions), Barclays indices.cib.barclays, MacroMicro, Slickcharts, CNBC, Schwab, MSCI factsheets, YCharts, FinanceCharts.
• Timed out: home.treasury.gov real yield curve, superseded by H.15.
• Binary-only: the GMO, AQR and J.P. Morgan PDFs needed local text or image extraction. If any becomes a recurring input, that step is part of the contract, and figures read off a chart rather than printed carry roughly ±0.5 pp of reading error — every such figure on this page is marked with ≈.
• AQR’s “Exceptional Expectations” is gated on aqr.com. Every -/media/ path returns HTML and ?aqrPDF=1 returns a cover stub. §5.1’s decomposition was read from a third-party mirror whose authors, disclosures and exhibit sources are intact. A primary-source copy should be obtained before this figure supports a decision record.
• Vanguard’s currency-valuation page is three years stale, dated 2023-09-30. It is the only free source found for a dollar over/undervaluation estimate with a stated method, and it cannot be used for a 2026 decision. This is the binding gap on the hedging question in §6.1.
• BIS/FRED real effective exchange rate: not reachable, so the dollar’s real-effective percentile in §5.2 is a chart-read from a J.P. Morgan slide rather than a computed figure.
• Shiller’s own excess-CAPE-yield publication is not reachable. Three secondary sources quote “current” ECY values spanning 3x (0.49%, 0.97%, 1.46%) on three different real-rate definitions. This repository uses the workbook’s own Excess_CAPE_Yield column and its own TIPS-based recomputation, and reports both, because there is no single number to quote.

---

2. What valuation predicts, and with what resolution

2.1 The regression, and the standard error that changes the answer

Shiller real total return, 1881-01…2026-08, regressor log(1/CAPE), response the annualised subsequent real log total return.

Three readings of the same data. Newey-West with a lag equal to the horizon is asked for 120 autocovariances from 13.6 independent windows and answers confidently; Hodrick’s 1B estimator sums the regressors backward instead of the residuals forward, so the number of quantities estimated does not grow with the horizon, and it halves the statistic. At five years the non-overlapping regression — which needs no HAC at all, and has 29 observations — gives t = 1.48.

The rising R**2 is not rising information. It rises because averaging returns over a longer window removes variance from the denominator, and the independent-observation count falls in exact step. An R**2 of 0.372 computed on 8.7 independent observations is not stronger evidence than an R**2 of 0.032 on 144.7.

2.1a Robustness: the averaged price series, and what it is worth here

Shiller’s P is the average of the month’s daily closes, not a month-end price, and the evidence base records the source-fitness finding that this disqualifies it for rules that depend on serial correlation. An overlapping long-horizon regression is not a moving-average rule, but it is in the same family — induced autocorrelation inflates HAC standard errors’ cousins — so the check is owed rather than assumed.

Same predictor, same 1926-07…2026-06 window, two responses: Shiller’s monthly-averaged real total return, and Ken French’s month-end Mkt-RF + RF deflated by the same CPI.

Three readings. The averaging triples the first-order autocorrelation (+0.087 to +0.265) exactly as the trend work found on its own window. It inflates the Hodrick statistic by about a quarter, and on month-end data the ten-year t is 1.88 — below the conventional threshold the averaged series clears. And the Newey-West statistic barely moves (4.37 vs 4.31), which is its own indictment: the estimator is so oversized in this design that a tripling of the response’s autocorrelation is invisible inside it.

Two things change at once and this design cannot separate them: the dating, and the universe, since French covers all of CRSP (18.4% annualised volatility) where Shiller covers the S&P composite (15.3%). The AR(1) column isolates the dating, which is the part at issue.

The direction of the effect is against the headline result, not for it. Every conclusion below is therefore quoted on the Shiller series, which is the more favourable of the two to the case that valuation predicts.

2.2 Stambaugh bias, which is unusually large here

Both ingredients are at their theoretical maximum here for a mechanical reason: price sits in the denominator of the valuation ratio and inside the return, so the predictor’s innovation is very nearly the negative of the return’s. This is not a defect of CAPE specifically; it applies to every price-scaled predictor. The consequence is that the uncorrected slope overstates by roughly a factor of four, and almost no published version of this chart corrects it.

2.3 The out-of-sample record

Expanding-window coefficients and an expanding-window benchmark mean, both using only outcomes already observed at the forecast origin.

Over the full sample the model is roughly a draw with a rolling mean at ten years and beats it at fifteen. Since 1990 it loses at every horizon, and the sign of the error is consistent: it forecast too little return, by between four and eight percentage points a year. The model’s mean error is positive in every row of the table, full sample included — it has been systematically pessimistic for a century, which is what a persistently rising valuation level does to a mean-reverting model.

The 1990-onward rows carry 1.4 to 6.3 independent observations. They are a record, not a test.

2.4 Is the level comparable across eras?

Two mechanical adjustments, in opposite directions, and neither is small.

Payout has moved out of dividends. Shiller’s own D/E ratio is 0.307 (as of 2026-03) against a 0.646 average before 1980. More than half the historical dividend payout now leaves as buybacks or retention, which raises earnings growth per share and means a given CAPE supports a higher return than the same CAPE did in 1950. Shiller publishes a total-return CAPE for exactly this reason; it stands at 43.98, at the 98.7th percentile — so the adjustment moves the level up and the percentile barely at all.

Real rates are the other adjustment, and it now points the other way. §1.1 has the arithmetic: the TIPS-based excess CAPE yield is at the bottom of its 23-year range.

Two more are named and not quantified here. Post-2001 goodwill impairment accounting depresses reported E in recessions, raising CAPE mechanically. And index concentration is at a record: the Magnificent Seven are 34.03% of S&P 500 market capitalisation (as of 2026-07, Voronoi), against a historical top-ten average near 24% and a prior peak near 28% in 1970. Yardeni puts the Mag-7 forward P/E at 23.7, its narrowest premium to the broad market since April 2025 — concentration currently rising on delivered earnings rather than on multiple expansion. This page does not adjust the CAPE for any of these. Their existence is the reason the level’s percentile should not be read as a probability.

---

3. Does conditioning the allocation on valuation beat not conditioning it?

The decision-relevant test. Shiller real equity total return against Shiller’s modelled real ten-year bond total return, 1921-01…2026-07 (1,267 months after a 40-year burn-in), monthly rebalancing, all percentiles expanding-window with no look-ahead.

Net figures charge 10 bp of execution and a 15% effective capital-gains rate on the equity sold. “Timing vs matched” compares each rule with a constant mix held at the rule’s own average weight, which strips out the part of any answer that is really just “held less equity.”

Three things follow.

The CAPE level is the wrong signal. Every level rule loses gross. Its timing component is mildly positive (+16 to +37 bp), so it is not that the level is uninformative — it is that acting on it costs more exposure than the timing is worth, because a level that drifts upward makes an expanding percentile rule sell continuously.

The excess CAPE yield is a genuinely different object. +49.5 bp gross, of which +47.7 bp survives the matched-weight control, at an average weight of 0.803 — it barely de-risks at all, it reallocates when. The signal is joint, and neither leg works alone. A rule tilting on the implied real yield alone, with no valuation term at all, lost −80.5 bp gross and was ahead in 0 of 908 rolling 30-year windows; the CAPE-yield-alone rule managed only +16 bp of timing. So the excess-CAPE-yield rule is not a bond-cheapness bet wearing a valuation label, and it is not the CAPE level with extra steps.

Tax decides it. Charging the rule only its turnover in excess of the constant-mix control’s own 6.1% — because a constant 80/20 also rebalances and also pays tax — the break-even effective capital-gains rate is:

The effective rate is the statutory rate times the embedded unrealised gain fraction. A long-held taxable position at a 15% federal long-term rate and 75% embedded gain sits at about 11% — exactly on the break-even. Add state tax or the 3.8% net investment income tax and it is comfortably below water. In a tax-advantaged account the effective rate is zero and the rule nets +48.6 bp.

3.1 The distribution, which is what the investor actually faces

An investor gets one draw, so the mean is the wrong summary. Rolling 30-year windows, annualised difference against the constant 80/20:

908 windows from about 2.5 independent 30-year blocks. The 98.5% is not a 98.5% probability of anything; it is one century seen 908 times.

Neither is the edge stable across eras. Gross, against the constant 80/20:

The CAPE-level rules earned their entire historical reputation in one era and have lost in every era since 1950. The excess-yield rule has the wrong sign in one of four. Each cell is about 0.7 independent 30-year blocks, so the dispersion is not itself evidence of regime change — but it is evidence that the full-sample mean is not a stable expectation.

3.2 Resolution

The instrument’s limits, stated before any verdict is read from it.

• The excess-CAPE-yield rule’s gross edge of 49.5 bp against 146 bp of tracking error clears its own 30-year minimum detectable effect of 34 bp and would take 14 years to reach 90% confidence — on the 106-year sample. Net, the edge is negative, so no horizon demonstrates it.
• Every CAPE-level rule has an MDE between 53 and 119 bp/yr and a measured edge of the wrong sign. Those rules are not merely unproven; they lost by more than the design’s resolution in the taxable case.
• Every backtest above uses revised, non-point-in-time Shiller data. The whole workbook is rebuilt on each release, so the “expanding window” hands each historical decision a history that had not yet been written. This biases the conditional rules in their own favour, and the level rules still lose.

---

4. Valuation as a risk statement rather than a forecast

This is the part of the investor’s concern that the evidence supports.

Subsequent annualised real total return by entry CAPE:

And the holdability statistic, over fifteen years from entry:

Read the two tables together. A buyer at CAPE above 30 did badly over ten years, recovered by fifteen, and did fine by twenty. What was different the whole way through was the path: the median such buyer spent six of every ten months of the next fifteen years below their own real entry level, against one month in twenty for a buyer below CAPE 20, and watched a median 52% real drawdown along the way.

That is an argument about whether the portfolio gets held, not about what it returns. It is the mechanism by which a high entry valuation can legitimately lower an equity share — via the drawdown constraint that setting the equity share §1.2 identifies as the thing that actually binds, rather than via a return forecast.

The resolution caveat is severe and unavoidable. The CAPE > 30 bucket is 57 monthly observations spanning seven distinct calendar years and two episodes — 1929, and 1997-2002 — which is 0.32 independent fifteen-year observations. Current CAPE is above every month in that bucket except the 1999-2000 window. The conditioning set for the question being asked is one episode. No probability in the table above should be quoted as one.

---

5. The relative call, which is a different and better-posed question

A cross-sectional comparison does not depend on any market’s level drifting, which is the defect that sinks §2. Jordà-Schularick-Taylor R6, 18 countries, 1870-2020, local-currency real returns, valuation measured by dividend yield. US minus the panel median.

This is the better-supported estimand. It has more independent observations than the US time series at the same horizon, and its identification is cross-sectional. Demeaning both variables inside each year — so that every global shock is absorbed by a year fixed effect and only within-year variation identifies the slope — leaves it intact, with Driscoll-Kraay standard errors clustered by year and HAC-corrected over the overlap:

The within-year slope at ten years, +0.038, is well below the US time-series slope of +0.062 — a one-log-unit valuation gap between two countries buys less than a one-log-unit change in one country’s own level appears to. That is what you would expect if part of the time-series slope is the drift §2.2 and §2.3 identify.

And it does not survive the modern era either.

The 1990-2010 design could have detected the historical slope of ~0.07 at 80% power and found zero. That is a more informative null than §2.3’s, and it is still built on 2.1 independent ten-year observations. The honest statement is that the cross-sectional relation held for 120 years and has not been detectable in the only era anyone is asking about — which is exactly the era in which the “US is expensive relative to the world” call has been made and been wrong.

5.1 How old the US premium actually is, and what built it

The US premium is about fifteen years old and its sign was inverted thirty-five years ago. AQR’s Ilmanen and Maloney track MSCI USA CAPE against MSCI World ex-USA CAPE from 1980: in the late 1980s the US CAPE was less than half the non-US CAPE (the Nikkei bubble), the two converged through the 1990s and sat near parity until the financial crisis, and by end-2024 the US was “nearly twice” non-US — a relative re-rating of 4x since 1989 and 2x since 2009 (“Exceptional Expectations: U.S. vs. Non-U.S. Equities”, May 2025, data to 2025-04, read 2026-08-22 from a third-party mirror because the paper is gated on aqr.com).

And the premium is almost entirely a price change, not an earnings change. Their decomposition of 35 years of MSCI USA outperformance over MSCI World ex-USA to 2024-12, currency-hedged and real, totalling 4.7 pp/yr:

About 81% of it was re-rating. For scale, AQR puts the US’s long-run growth edge at about 1 pp/yr over the past century and its long-run return edge near 2 pp/yr, with the cumulative-to-date figure trough at 1.1 pp (end-1988) and peak at 2.2 pp (end-2024).

This is the strongest available argument for a valuation tilt and it is not a CAPE argument. A premium built out of multiple expansion is a premium that a reversal of multiple expansion takes away; a premium built out of earnings growth is not. AQR’s own arithmetic for what reversion would look like: a 45% fall in US prices from December 2024 levels returns the two CAPEs to parity, or equivalently a positive-but-disappointing US growth edge of ~1 pp/yr with 5 pp/yr of relative repricing over a decade.

It is also the argument with the most obvious counter, and AQR supplies that too: “about half of US outperformance in the past 15 years and of its relative richness can be attributed to its different sector composition than the rest of the world (mainly a larger tech sector).” Half. A sector-neutral version of the spread is roughly half the headline.

5.2 Three confounds, now measured rather than named

Buybacks close about half the payout gap, and about a third of the US’s own CAPE elevation. The US-minus-panel log dividend-yield spread by decade — −0.093 (1990s), −0.498 (2000s), −0.516 (2010s), −0.259 (2020s) against a full-sample mean of −0.003 — moved to a large discount exactly as buybacks became the majority of US payout, and exactly when the cross-sectional relation stopped working. Total shareholder yield tells a different story (J.P. Morgan Guide to the Markets — UK slide 49, as of 2026-06-30, chart-read):

The US-versus-Europe dividend gap is ≈1.45 pp; the total shareholder yield gap is ≈0.75 pp. Buybacks close roughly half of it. And EM’s total shareholder yield is below the US’s, because EM dilutes — Capital Group puts 2025 US buybacks at 147% of US dividends, Europe ex-UK at 44%, Japan at 72%, and EM at under one-twelfth (Global Equity Study, March 2026, FY2025, read 2026-08-22).

The adjustment does not improve prediction. Keimling’s payout-adjusted CAPE attributes about one-third of the US’s elevated recent valuation to the change in dividend policy — and finds the adjusted measure shows “no signs of superiority to ordinary CAPE either in terms of R**2 or correlation”, with R**2 deteriorating in 9 of 16 countries (StarCapital, 2016, read 2026-08-22). So the confound is real, it is roughly a third of the level, and correcting for it buys no forecasting power.

Currency is the largest single term and this repository’s panel does not price it. JST returns are local-currency real. J.P. Morgan’s decomposition of USD total returns (Guide to the Markets — U.S. slide 45, as of 2026-06-30, chart-read) puts the currency contribution to unhedged non-US returns at ≈−1.7 pp/yr over 2010-2024, then ≈+7.2 pp in 2025 alone (Eurozone ≈+14.7 pp), then ≈−1.6 pp in 2026 H1 (Eurozone ≈−4.3 pp). Currency, not the multiple, drove 2025’s international outperformance. Across every dollar cycle since 1973 the currency contribution and the direction of US relative equity performance have agreed in sign, and in the current cycle roughly three-quarters of the US’s relative underperformance is currency (Guide to the Markets — UK slide 13). For a USD investor, AQR’s 2026 assumptions put the carry from hedging Eurozone equity back to USD at ≈+0.8 pp/yr.

The unconditional US premium in the JST panel is large. The intercept is +1.84 pp/yr at ten years: at an equal valuation spread the US beat the panel median by nearly two points a year over 150 years. Some of that is genuine and some is that the panel includes markets that closed. Either way, a relative-valuation model must overcome a large positive constant before it recommends underweighting the US.

5.3 What the published evidence says about this exact estimand

The most honest published statement found, from AQR on the relative US/non-US CAPE against next-decade relative performance, 1979-12 to 2025-04:

“the relative US/Non-US CAPE has predicted quite well the next-decade relative performance. The predictive correlation is +0.5 over an admittedly short sample (only 4+ independent observations).”

A correlation of 0.5 is an R**2 near 0.25 on four independent draws. That is the same shape as §5’s result and the same shape as §2’s: a real relationship with almost no statistical resolution. AQR add, in the same paper, that “this prediction has not panned out well for the past decade” and that consensus capital-market assumptions “since 2011 consistently assumed lower future return for US equities than for Non-US” — the gap has been forecasting US underperformance for fifteen years and been wrong throughout.

The closest published analogue to §3’s timing test reaches §3’s conclusion independently. AQR’s Market Timing: Sin a Little (2017) runs a realistic CAPE contrarian rule on US equity with no look-ahead:

“The results are disappointing. Even before costs, the timing strategy cannot beat the Sharpe ratio of buy-and-hold over either the full 116-year sample or the latter half… Neither return differences… nor Sharpe ratio differences… are statistically significant.”

And on the pooled cross-country level relation, Keimling’s 16-country panel gives R**2 0.49 and correlation −0.67 against subsequent 10-15 year real returns, with a median subsequent real return of +0.5%/yr from a starting CAPE above 30 — accompanied by his own caveats that the sample “comprises only two independent 10-15-year periods”, that removing Japan cuts the pooled R**2 by 0.07, and that per-country R**2 ranges from 0.01 (Canada) to 0.90 (Hong Kong).

Three independent literatures, three different designs, one answer: the relationship is there and the resolution is not.

5.4 Where the spread sits now, and whether it has already narrowed

It has narrowed, and mostly for a reason that is not a US de-rating. Siblis’s own three-year table (as of 2026-06-30):

The US did not de-rate — Asia re-rated violently. Korea’s CAPE roughly tripled in eighteen months. “International is cheap” is not true of the Asian half of EAFE and EM at these levels, and the narrowing of the aggregate ratio is largely arithmetic on that.

On forward multiples the picture is milder. J.P. Morgan’s relative forward P/E of MSCI ACWI ex-US against the S&P 500 stands at a −31% discount against a 20-year average of −20% (Guide to the Markets — U.S. slide 46, as of 2026-06-30) — on the US/ex-US convention, 1.45x now against 1.25x on average. That is 11 pp wider than average and marginally narrower than the −32% at 2026-01-01. The same slide shows ACWI ex-US at a wider-than-average discount in every sector it lists, so sector mix is not the whole of it on JPM’s own data, even though AQR attributes about half.

And relative performance has already delivered part of the trade and then stopped. In 2025 EAFE returned +31.9% in USD against the S&P 500’s +17.9% — a 14 pp gap, and 41.3% for the Eurozone. In 2026 H1, EAFE returned +9.8% against the S&P 500’s +10.2%: the developed relative trade stopped working, and ex-US only beat the US in aggregate because of Korea (+118.9% USD YTD) and Taiwan (+62.6%). JPM’s rolling two-year EAFE-minus-USA series was continuously negative from about 2009 through 2025, bottomed near −10%/yr in late 2024, and has since climbed only to roughly 0 to +2% — “fifteen years of US outperformance” has been given back for about eighteen months and the two-year window has barely reached zero.

5.5 Sizing the decision

Taking the ten-year fit at face value and plugging in the current spread — an assumption, not a measurement, because it applies a dividend-yield elasticity to a CAPE ratio, and §5.2 has just shown the two differ by roughly half the gap:

Siblis’s US/ex-US CAPE ratio of 1.704 is a log yield spread of −0.533. The fit implies the US underperforms by −1.38 pp/yr over ten years, with a residual standard deviation of 2.95 pp — a 95% band of −7.2 to +4.4 pp/yr. The interval is four times the point estimate and comfortably contains zero and US outperformance.

Two independent estimates land close to it. AQR’s December 2025 capital market assumptions give US large-cap 3.9% real against global developed ex-US 4.9% — a 1.0 pp gap; their December 2024 version gave 4.2% against 6.1% hedged, a 1.9 pp gap. AQR reach that number from the payout yield alone: they set the valuation-reversion term to zero and assume expected spot FX equals the inflation differential, so their ex-US edge contains no currency alpha and no mean reversion. Read the other way round, their framing says the US needs a 2.2 pp/yr growth edge merely to match non-US, assuming the valuation gap persists.

A JST-panel fit built on 150 years of cross-country data, and a practitioner model built on payout yields with no valuation term, both land near 1 to 1.9 pp/yr. That is meaningful agreement on the point estimate and says nothing about the interval, which remains four times as wide.

What a shift is worth, at the JST point estimate:

Tracking errors bracket an 8%/yr US-versus-developed relative return volatility against the 14.4%/yr the JST panel actually shows for US-minus-panel-median. The information ratio does not change with the size of the shift, which is the same structure setting the equity share §1.3 found on the equity ladder: the size of the step barely changes how long it takes to prove.

5.6 The reframing that matters most

Siblis’s global index is ~64% US (as of 2026-06-30). So:

65/35 is not a US overweight. It is approximately the global market portfolio. The question in front of the investor is therefore not “should I unwind an overweight” — there isn’t one — but “should I deliberately underweight the market portfolio on a signal whose cross-sectional relation has been undetectable for 35 years and whose implied edge has a confidence interval four times its point estimate.”

---

6. What to do

6.1 The US/international split

Keep 65/35 as the default. If it moves, move it to 60/40 with new contributions and not by selling.

The reasoning, in order of weight:

1. 65/35 is already market-cap neutral, so no correction is owed. A move below it is a new active bet, not the removal of an old one.
2. The direction of the tilt has the best argument on this page, and it is not a CAPE argument. It is §5.1’s decomposition: 81% of 35 years of US outperformance was re-rating, 23% was earnings growth. A premium built out of multiple expansion is the kind a reversal takes back. That argument is about the composition of a realised return, which is a measurement, rather than about a forecast.
3. Its size is small and its interval swamps it. The JST fit implies −1.38 pp/yr with a 95% band of −7.2 to +4.4; AQR’s payout-yield model implies 1.0 pp (2025-12) or 1.9 pp (2024-12) with no valuation-reversion term at all. Two very different methods agreeing on ~1-2 pp is real corroboration of the point estimate and no help at all with the interval.
4. It is far too small to justify realising a capital gain. ≈14 bp/yr per 10 pp shifted against 80-144 bp of tracking error, needing 55-178 years to demonstrate. At a 15% effective rate, moving 10 pp of a portfolio with 75% embedded gain costs about 113 bp of one-off wealth against an edge of 14 bp/yr — an eight-year payback on a signal whose sign is not established.
5. It is nearly free with new contributions, which pay no tax to direct. If the investor wants a valuation tilt, this is where it belongs, and it is where structural and tax edges put every other tilt.
6. Much of the trade has already happened, and part of it was currency. EAFE beat the S&P 500 by 14 pp in USD in 2025 and by −0.4 pp in 2026 H1. Currency contributed ≈+7 pp of the 2025 non-US result and has been negative again in 2026. An investor moving now is buying after the re-rating and after the dollar’s fall from its early-2025 peak — the broad real dollar is roughly 7-8% off that peak and still in the upper quartile of its 29-year range, so there is less currency tailwind left than there was.
7. Do not go to EM on the valuation argument. EM’s CAPE of 19.36 is cheap against the US and, per AQR, at the 98th percentile of its own history since 2001. EM’s total shareholder yield (≈2.15%) is below the US’s (≈2.8%) because EM dilutes — buybacks are under one-twelfth of dividends there. And the forecasters disagree on EM more than on anything else: AQR +5.1% real, GMO −1.8% real, Vanguard 2.0-4.0% nominal. There is no consensus to follow.
8. “International is cheap” is not true of Asia. Japan 38.59, Korea 40.76 and Taiwan 46.44 are at or above the US’s 35.82. The developed-ex-US discount is a Europe, UK, Australia and Greater China phenomenon, and a cap-weighted EAFE purchase buys a lot of the expensive half.

Confidence: high on “do not sell to move”; moderate on the direction; low on the size. The moderate rating on direction comes from the re-rating decomposition and from two independent models agreeing on the sign. The low rating on size comes from the 1990-2010 null, from AQR’s own +0.5 correlation on four independent observations, and from the fact that consensus capital-market assumptions have made this exact call every year since 2011 and been wrong every year. If the investor wants a number, 60/40 funded by new money is defensible. 50/50 is not supported by anything measured here.

On hedging, which the split raises and this page has not tested: AQR’s 2026 assumptions put the carry from hedging Eurozone equity into USD at ≈+0.8 pp/yr for a USD investor — larger than the entire valuation edge in point 4. That is a contractual, rate-differential quantity rather than a forecast, and it deserves its own study before the split moves at all.

6.2 The equity share

Do not cut it on the return forecast. Do widen the assumptions it is sized against.

The forecast route fails at every step: the corrected slope is a quarter of the raw one, the Hodrick statistic is 2.47 at ten years on 13.6 independent observations, and the out-of-sample record since 1990 is negative at every horizon with a consistent four-to-eight point pessimistic bias. And the regret of being wrong is not small, because it is linear in the weight cut and needs no forecast to state:

For scale, the realised figure on this repository’s own Shiller sample is +4.59 pp/yr (equity 6.50% real log, modelled ten-year bond 1.91%, 1881-2026). The investor should choose the row they actually believe, and notice that the cut only pays in the bottom two.

The route that does work goes through the constraint. §4 establishes that entry above CAPE 30 historically meant a median 52% real drawdown and six of every ten months below real entry for fifteen years. Setting the equity share §1.1-1.2 establishes that under the zero-leverage rule the growth objective alone returns a corner solution, and every bond in the portfolio is there because of a drawdown constraint nobody has supplied a number for. If the tolerable-drawdown number is set against a −37% real assumption and the conditional history says −52%, then the constraint binds harder and the implied weight falls — as a constraint decision, not a market call, and the difference matters because it tells the investor what would reverse it.

Concretely: this page recommends no change to the equity share and instead recommends that the drawdown assumption used to set it be widened from the unconditional −37% to the high-valuation-conditional −52%, and the “months underwater” assumption from ~5% to ~60% of a fifteen-year window. Whether that widening changes the share depends on a number the investor has not yet supplied. That number — the tolerable peak-to-trough real drawdown — is the single most valuable missing input in this repository, and it is worth more than any further valuation research.

6.3 A dynamic, valuation-conditional rule

No, in a taxable account. Defensible but unpromoted in a sheltered one.

• Rules on the CAPE level lose gross and net, in every specification tested. There is no version of this worth running.
• The excess CAPE yield rule is the only one with a real gross edge (+49.5 bp, +47.7 bp of it timing), and its break-even effective capital-gains rate is 11.3% — below where a long-held taxable position sits.
• In a tax-advantaged account it nets +48.6 bp against 146 bp of tracking error. That is a real but modest number, on a rule that had the wrong sign in one of four eras, was measured on revised data that flatters it, and rests on ~2.5 independent 30-year windows. It is exploratory and no specification was frozen. It should not be implemented on this evidence; it is a candidate for a registered experiment.
• The signal’s current reading, for the record and not as a recommendation: the excess CAPE yield sits at the 18.7th percentile of its expanding history on Shiller’s own measure and the 0th percentile of the TIPS era. A k=0.4 rule tilting on the former would set 0.675 against a 0.80 base; a CAPE-level rule would set 0.604.

6.4 The one-line answer to the investor

The evidence does not support timing, it does support a modest tilt funded by new money rather than by sales, and it strongly supports a wider expected-drawdown assumption. The return-forecast case has a corrected slope a quarter the size of the advertised one, a Hodrick t of 2.47 on 13.6 independent observations — 1.88 on month-end data — and a 35-year out-of-sample losing record. The risk case has a median 52% real drawdown and six of fifteen years underwater, from two episodes. The relative case is the best of the three, rests on the fact that four-fifths of the US premium is re-rating rather than earnings, and is worth 14 bp/yr per 10 pp shifted against 80-144 bp of tracking error — less than the ≈80 bp/yr of carry available from hedging the currency, which nobody here has studied.

---

7. Open questions, and the next informative test

Open.

• Is the current earnings surge durable? FactSet flags a record 29.2% aggregate surprise as heavily driven by mark-to-market on equity stakes rather than operating income, and hyperscaler capital spending is reportedly running near 100% of operating cash flow while buyback activity at those firms is down 64% year on year. CAPE and the forward multiple disagree by a factor the answer to this question would resolve. This repository has no instrument for it.
• What does the excess CAPE yield rule do on a point-in-time history? Every result in §3 used a revised workbook. No vintage archive is published, so this needs a different source.
• Does the cross-sectional relation hold on cap-weighted regional indices with currency? The JST panel is local-currency, equal-treatment, dividend-yield-based and ends in 2020. A test on MSCI USA against MSCI EAFE and EM in USD, with a CAPE or composite valuation measure, would answer the actual question. research/ already holds regional French factor files used by Experiments 005 and 009.
• What is currency hedging worth, and should the international sleeve be hedged? AQR puts the carry at ≈+0.8 pp/yr for a USD investor — larger than the valuation edge in §6.1 — and §5.2 shows currency has been the largest single term in relative returns. This page did not test it and it is now the most under-researched input to the split.
• How much of the US/ex-US gap survives sector neutralisation? AQR says about half of the US’s relative richness is sector composition; J.P. Morgan’s slide shows ACWI ex-US at a wider-than-average discount in every sector. Those two are not obviously compatible and a sector-neutral relative CAPE computed here would settle it.
• Two published readings could not be reconciled and are recorded as data-contract findings. Four “current” US CAPE readings span 40.4 to 41.96 on three constructions (§1.2). And J.P. Morgan’s Guide to the Markets — UK slide 59 prints an MSCI Europe ex-UK CAPE of 41.3x against Siblis’s Germany 23.28 and France 20.91 for the same date. It is irreconcilable and this page does not use it.

Not open. Whether the CAPE level alone should drive an allocation rule. Every specification tested lost, gross and net; the timing component that does exist is swamped by the exposure it gives up; and AQR’s independent 116-year test of a realistic contrarian CAPE rule found a Sharpe of 0.37 against buy-and-hold’s 0.38 with a −32% relative drawdown. Reopening condition: a point-in-time history, or a measure that removes the upward drift that makes an expanding percentile rule sell continuously.

Next informative test. The cap-weighted, USD, currency-decomposed version of §5 on regional indices, reporting the hedged and unhedged legs separately. It targets the estimand with the most decision leverage (the split), it is the one of the three claims with a defensible identification strategy, the currency term turns out to be larger than the valuation term, and the data is already partly in the cache. Second priority is the investor’s tolerable-drawdown number, which is not a research task at all and is worth more than the first.

---

Status

§1 is read, with sources and read dates; §§2-5 are measured and exploratory; §6 is interpretation. Nothing here is promoted, no specification was frozen before the results were seen, and no experiment is registered. Under decision 0010 the nulls in §2.3, §3 and §5 are scoped to their designs: “not detected here” is not “does not exist,” and the independent-observation counts beside every table are there so that no verdict outruns its instrument.