How we forecast equity returns

These are ten-year forecasts

Valuation is a slow signal. How expensive a market is today tells you a good deal about the average annual return you can expect over the following decade, and close to nothing about the return you will get over the following year. This has been the central finding of the research on long-horizon return predictability since Campbell and Shiller's work in the late 1980s, and it holds across markets.

A cheap market can get cheaper for years before it recovers. An expensive market can carry on getting more expensive for a decade, as the US has done. Over one year, sentiment, earnings surprises and rate moves swamp valuation entirely. Over ten years, those effects wash out and the starting price does most of the work.

The forecaster offers three-year and five-year views as well. Those apply a proportional share of the same ten-year valuation journey, so the arithmetic stays consistent. The predictive content at those horizons is genuinely weak, and the tool says so on screen when you select them.

The formula

An equity investor is paid in three ways over a long holding period. Earnings grow. The multiple the market pays for those earnings changes. Dividends arrive along the way. The forecast is the sum of those three things:

Written out with the annualisation, where H is the horizon in years:

Growth and the multiple compound with each other, because earnings grow first and are then capitalised at whatever multiple applies at the end. The dividend is income received along the way, so it adds rather than compounds. This is the same decomposition used by Research Affiliates and GMO for long-horizon capital market assumptions.

Why CAPE rather than a forward P/E

CAPE divides a market's price by its average inflation-adjusted earnings over the previous ten years. A conventional P/E divides price by one year of earnings, either the year just gone or analysts' estimate for the year ahead.

One year of earnings is a poor anchor for a long forecast. Profits are cyclical, so a single year can flatter or wreck the ratio for reasons that have nothing to do with what the market is worth. A market at the bottom of a profit cycle looks expensive on a trailing P/E at exactly the moment it is cheap. Averaging a decade of real earnings removes most of that noise.

There is a second, more practical reason. Research Affiliates publish the complete historical CAPE distribution for every market they cover, going back to 1979 for most developed markets and to 1880 for US large caps. That is what allows the scenario bands on the forecaster to be observed history rather than a spread somebody chose. Nobody publishes equivalent forward P/E history for free, so a forward-P/E version of this tool would have had to invent its scenarios.

Forward P/E still appears in the forecaster table, taken from MSCI, as context. It is not an input to the forecast.

How far back the history should go

US CAPE data starts in 1880. Calling a market expensive against 146 years of its own history invites an obvious objection: the United States of 1881 was, in Research Affiliates' own words, an emerging market by modern standards. Accounting standards, index construction, sector mix and the way companies return cash have all changed since. A fair multiple in 1881 and a fair multiple today are not the same number.

Research Affiliates make this argument themselves. Their work on CAPE shows the best-fit line through US CAPE rising from roughly 12 times in 1881 to roughly 19 times by 2017, and they have long presented the series as an upward-sloping trend rather than a flat average, on the view that its equilibrium level is not static. Jeremy Siegel's critique in the Financial Analysts Journal runs alongside it: the 2001 change to goodwill impairment rules and the wider move to mark-to-market accounting deepened reported earnings declines in downturns, which inflates CAPE. Substituting national accounts profits or operating earnings lowers the measured ratio by roughly 15 to 30 percent in recent decades.

Why the long history does not drive these forecasts

The base case never reverts to the 146-year median. Research Affiliates' fair value comes from an exponentially weighted moving average with a 50-year window and a 20-year half-life, shrunk towards the market's group, which is why their expected CAPE for US large caps is 28.3 rather than the 16.6 the full history would suggest. Almost all of the concern about ancient data applies to an anchor this model does not use.

The long history is used for one thing only: the ratio of each market's 25th and 75th percentile CAPE to its median, which sets how wide the bear and bull bands are. Ratios are far more stable than levels. On Shiller's US series those ratios are 0.72 and 1.28 measured over the full history, 0.71 and 1.30 measured since 1950, and 0.64 and 1.28 over the last 50 years. Only a 30-year window differs much, at 0.87 and 1.16, and a window containing almost no cheap markets is a poor guide to how cheap a market can get.

Feeding those different windows through the model moves the US bear case between -0.9 and -3.9 percent a year and the bull case between 1.9 and 3.1 percent. The base case does not move at all, because it does not depend on the percentiles.

Adjusting fair value for interest rates

Equities and government bonds compete for the same money. When the real return available on a bond falls, investors will pay more for a given stream of real earnings, so the multiple a market deserves is not a fixed historical average. Vanguard identified this as the reason standard CAPE forecasts have performed poorly since the mid-1980s, and found that conditioning fair value on real bond yields cut out-of-sample forecast errors substantially.

The fair-value CAPE here is therefore tilted by how far a market's real bond yield sits from its own long-run norm. The sensitivity was fitted on US data, 810 monthly observations from 1956 to 2023, regressing the log of the cyclically adjusted earnings yield on the real 10-year government yield:

The tilt applies to the bear and bull targets as well as the base, since it moves the whole anchor.

One adjustment for the whole table

The same multiplier is applied to every market, taken from the US real yield, rather than each market using its own. That is a decision forced by arithmetic rather than a shortcut.

A bond yield and price series long enough to give a meaningful norm exists for 11 of the 45 markets here. An earlier version adjusted only those, and it broke the table. MSCI World is 72 percent US by weight, so lifting the US by a quarter while leaving World alone put an index below almost every market it contains, which cannot be true of a weighted average of those same markets. It also left the ranking as a mixture of two methods, and ranking markets is what this tool is for.

Scaling every fair value by one factor keeps aggregates consistent with their constituents by construction, and shifts the level of the table without reordering it. Because the uplift is multiplicative rather than additive, a market with a larger valuation term gains marginally more, so near-ties can swap; across the current 45 markets four move and none by more than one place.

The US yield is the one used because the sensitivity was fitted on the US series, so a deviation measured any other way would not match the elasticity applied to it. It is also the world's benchmark real rate and the dominant weight in the global indices. The other ten series are still collected, and they show European real yields sitting further below their norms than the American one, which is a reason to treat this adjustment as conservative rather than aggressive.

The test that justified shipping it

Forecasting 10-year US real returns from 1985 onward, re-estimating every month using only the data available at the time:

Two things matter in that table. The rate-conditioned version cuts the error against mechanical reversion by 17 percent. More importantly it is the only one of the three that beats the naive historical average, which is the benchmark most published return predictors fail. The naive average has a negative correlation with what actually happened, so while its error is respectable it can rank nothing and is useless for deciding how much of each market to hold.

Two limits worth stating

Deviations are clipped at two percentage points. Several European real yields currently sit three points below their own medians, which are lifted by the high-rate years of the 1970s and 1980s. Left unclipped that produces a 64 percent uplift to Germany's fair value and a 10 percent real forecast, which is not a number this method has any business producing. Two points is roughly the range over which the US fit is well populated.

The underlying series need a nominal government bond yield and a consumer price index both reaching back to 1985 or earlier, otherwise the norm is a median of the disinflation and zero-rate decades rather than a long-run level. Research Affiliates' own real yield series were the obvious source and were rejected for exactly this reason: they begin between 1997 and 2008, on which basis today's ordinary US real yield ranks at the 79th percentile and the UK's at the 96th, which would tilt every market down for no good reason.

Where every number comes from

Every figure in the forecaster is either taken from one of those sources or derived from them by the arithmetic on this page. There are no hand-entered estimates in the data files.

Bear, base and bull

The base case does not revert to the median

The obvious way to build this model is to assume every market drifts back to its own historical median valuation. That would be a much stronger claim than the evidence supports. US large-cap CAPE has sat above its long-run median for roughly thirty years. A model that assumed reversion to the median would have forecast a US crash every year since the mid 1990s, and would have been wrong every year.

The base case instead uses Research Affiliates' own expected CAPE for each market, which is a partial reversion they model rather than a mechanical return to the middle of the historical range. For US large caps at the time of writing that means a target well above the median.

The bands come from each market's own history

The bear and bull cases scale that base anchor by how wide the market's valuation range has actually been, using the ratio of its 25th and 75th percentile CAPE to its median:

A market whose valuation has historically swung hard gets wide bands. A market that has traded in a narrow range gets narrow ones. Nobody picks a spread by hand. Any target CAPE can be overridden in the tool by clicking the return figure, which is the honest way to test a view of your own.

Shorter horizons

Research Affiliates' expected CAPE is a ten-year expectation. Applying it whole to a three-year horizon would assume a decade of valuation change happens in three years. Instead the multiple travels a proportional share of the distance, so a three-year forecast moves three tenths of the way to the target and only a full ten-year horizon gets all of it.

Where the growth forecast comes from

Of the three terms in the formula, valuation and dividend yield are observed today. Long-run real earnings growth has to be forecast, and it is the term with the most room for error, so it is worth setting out exactly where the number comes from and what was rejected on the way.

Research Affiliates' per-market estimate

Growth is taken market by market from Research Affiliates, who publish the full decomposition behind their own capital market expectations. Their method fits a time trend to the natural log of real earnings per share over a rolling 50-year window, averages each country's estimate with its group average for developed or emerging markets, then scales every country back if the GDP-weighted world total would exceed world real GDP growth.

Those three steps matter more than the raw estimate. The 50-year window stops a single decade dominating. The averaging step pulls outliers back towards their peers. The cap stops the world as a whole being forecast to grow earnings faster than the economy that produces them. Across the 45 markets shown here the forecasts run from -0.7 percent a year for Portugal to 5.9 percent for Colombia, averaging 2.6 percent for developed markets and 3.7 percent for emerging ones.

The figures are identical across all six currencies Research Affiliates report in, which is the check that they are local-currency real growth rates rather than a view on exchange rates. The refresh script verifies that every month and refuses to publish if it ever stops being true.

Why not each market's own history

The obvious cheaper approach is to give each market a growth rate measured from its own past. It does not work, and testing it is what showed why the shrinkage step above is doing the real work.

The Jordà, Schularick and Taylor Macrohistory Database records annual equity dividend yields, capital gains and inflation for 18 advanced economies from 1870, and is free to use. Reconstructing real dividends from it, growth since 1950 ranges from 7.6 percent a year for Germany to -2.9 percent for Portugal. Almost all of that is history rather than signal. Germany's figure is post-war reconstruction and Portugal's covers the 1974 revolution.

Splitting the record and asking whether a market's growth in the first period predicts the second gives correlations of +0.26 for 1950-1985 against 1985-2020, +0.26 for 1900-1960 against 1960-2020, and -0.37 for 1970-1995 against 1995-2020. It is weak and it changes sign depending on the period. Forecasting each country's 1985-2020 growth from its own 1950-1985 record gives a mean absolute error of 3.79 percentage points against 1.92 points for a single number applied to everyone, so using raw history roughly doubles the error. Shrinking towards the cross-country mean is the standard fix, and across six rolling windows the best weight to put on a country's own 125-year record is about 0.1.

That is the same conclusion Research Affiliates reach by a different route. Their forecasts are per-country, but only after a 50-year window, a pull towards the group average and a global cap have taken most of the idiosyncrasy back out.

The same database sets the fallback growth rate of 1.5 percent, used only if a market ever appears without a Research Affiliates forecast. It is lower than their average because dividends per share have grown more slowly than earnings per share over the period, while payout ratios fell.

Three plausible substitutes that the evidence rules out

Scaling growth to each country's economic growth is the most intuitive option and the most clearly wrong. Emerging markets have grown far faster than developed ones and their equities returned 6.9 percent a year against 8.5 percent for developed markets over 1900-2025 in the Dimson, Marsh and Staunton data. Ritter found a negative cross-country correlation between economic growth and equity returns. Fast-growing economies issue more new shares, so the growth accrues to new capital rather than to existing holders.

Using the retention ratio, in the textbook form where growth equals return on equity multiplied by the share of earnings retained, has the sign backwards. Arnott and Asness found that higher payout ratios predict faster subsequent earnings growth, not slower.

Following that finding across countries gets closer but still fails for this purpose. ap Gwilym, Seaton, Suddason and Thomas confirmed in 11 international markets that higher payout ratios did go with higher subsequent real earnings growth. They also found it did not translate into return predictability in any persuasive way, which is what the forecast actually needs.

What this means for reading the numbers

Growth flows through to the forecast at roughly one for one, so a market whose earnings grow a percentage point faster than forecast returns about a point more a year. Realised long-run growth has varied across countries with a standard deviation of about 2.2 percentage points, which is large next to forecasts that mostly sit between 2 and 4 percent. Having a per-market forecast rather than a single number does not remove that uncertainty, and the bear and bull bands do not capture it either, since they vary only the valuation target. Every growth figure is editable in the forecaster so the sensitivity can be tested directly.

A worked example

US large caps, using the data as of July 31, 2026. Current CAPE is 40.4, which sits at the 99th percentile of its own history since 1880. Research Affiliates' expected CAPE is 28.3. The historical median is 16.6, the 25th percentile 12.0 and the 75th percentile 21.5. The MSCI USA dividend yield is 1.13 percent, and Research Affiliates forecast real earnings growth of 2.96 percent a year. The US real bond yield is 1.19 percent against a norm since 1956 of 2.59 percent.

The bear case scales the target by 12.02 / 16.58, giving a target CAPE of 25.6 and a real return of -0.5 percent a year. The bull case scales by 21.53 / 16.58, giving a target of 45.8 and a real return of +5.4 percent a year. Without the rate adjustment the base target would be 28.3 and the base forecast +0.5 percent, so low real yields are worth about 2.2 points a year on this market at present.

The reason these are modest is arithmetic rather than pessimism about American companies. A dividend yield of 1.13 percent and real earnings growth of 2.96 percent give about 4.1 percent a year before any change in valuation, and starting near the top of a 146-year valuation range means the multiple is more likely to take something away than to add to it.

Why the numbers look low next to history

US equities returned 6.6 percent a year in real terms from 1900 to 2025 in the Dimson, Marsh and Staunton data. This model expects 2.7 percent. Even setting the valuation drag aside, the yield and growth terms only add to about 4.1 percent. It is worth being clear about where the difference comes from, because it is not a view about American companies being worse than they were.

Take the historical return apart using Shiller's data. The US dividend yield averaged 4.04 percent from 1900 to 2025, and 4.56 percent across the twentieth century alone. Real earnings per share grew 1.87 percent a year over the same period. Those two together give about 5.9 percent, and the remainder of the realised 6.6 percent came from the multiple rising over the period.

Today the MSCI USA dividend yield is 1.13 percent. That is the whole story. The yield is roughly three points lower than the long-run average, and three points is most of the gap.

The other reason these figures feel low is that they are real. Adding expected US inflation of 2.8 percent puts the yield-and-growth total near 6.9 percent in nominal money before any change in valuation, which is an unremarkable number for equities. What takes the US base case down to 0.5 percent real is the valuation term alone, and that is a claim about starting at the top of a long valuation range rather than a claim about company performance.

Real and nominal returns

The forecaster shows real returns by default, meaning after inflation. That is the honest output of the method, because CAPE's denominator is inflation-adjusted earnings and the growth forecast is a real one. Real returns are also comparable between markets, which nominal returns in 45 different currencies are not.

Switching to nominal asks which currency you spend in and adds that market's expected inflation, from Research Affiliates, to every forecast in the table. A British investor sees every market in pounds and an American sees every market in dollars, so the markets stay comparable with each other while the numbers become the ones you would recognise from a fund factsheet.

This rests on exchange rates moving to offset inflation differences over a decade, which is the standard long-horizon simplification and is roughly true over long periods and unreliable over short ones. Research Affiliates model the deviations from it explicitly, which is part of why their published dollar forecasts differ slightly from the local-currency figures here.

An earlier version added each market's own inflation instead. That produced a nominal Turkish forecast in lira sitting next to a nominal Japanese one in yen, which compares nothing useful, and it left regional indices like MSCI World blank, since a basket of countries has no inflation rate of its own.

How well has this kind of forecast worked?

Valuation-based forecasting is the best-evidenced approach available for country-level equity returns, and it has a documented record of being wrong in one particular direction. Both halves of that sentence matter and the second half is usually left out.

The case for it

Keimling tested valuation measures across 17 MSCI country indices from 1979, which is close to what this page does. CAPE explained roughly 48 percent of subsequent 10 to 15 year returns, and only CAPE and price to book produced reliable forecasts at all. Vanguard examined around 15 commonly used signals and found about half had no forecasting power whatsoever, with CAPE the strongest of those that did. Dahlquist and Ibert collected the published assumptions of the largest asset managers and found their expectations move with CAPE at a sensitivity of 68 basis points per 10 percent change, which is almost exactly the sensitivity realised 10-year returns have shown. The professionals do not disagree about whether valuation matters.

The case against it

Vanguard also tested how well the standard approach actually forecast. For 10-year US forecasts made since 1985, CAPE-based predictions correlated 91 percent with what subsequently happened and still had a root mean squared error of 7.8 percent against 6.2 percent for simply assuming the historical average return. The direction was almost perfect and the level was worse than not forecasting at all. Goyal and Welch found the same pattern across most published predictors: they fail to beat the historical mean out of sample.

What that means for the numbers on this page

The method is good at ordering markets from cheap to expensive and has been too pessimistic about expensive markets, the US above all, for about three decades. The real-rate adjustment described earlier is the one change the out-of-sample evidence supports, and it lifts the US rather than lowering it. It does not make the record disappear.

What other forecasters say

The forecaster shows five professional houses beside its own number because the spread between them is the most honest thing on the page. For US large caps, in nominal dollars over roughly a decade, they currently range from BlackRock at 7.7 percent to GMO at -4.0 percent. That is a gap of nearly 12 percentage points between firms with enormous research budgets looking at the same market on the same day.

J.P. Morgan publish their building blocks, so the disagreement with them can be taken apart rather than guessed at. Their 6.7 percent for US large caps is nominal and their own inflation assumption is 2.5 percent, so the comparable real figure is about 4.1 percent. Their blocks are revenue growth of 6.0 percent, buybacks 3.0, dividends 1.7, gross dilution -1.5, margins -0.5 and valuations -2.0.

The growth gap is the substantive one. J.P. Morgan assume revenue grows at 6.0 percent against their own nominal GDP forecast of 4.3 percent, and that buybacks shrink the share count by a net 1.5 percent a year for a decade. Both are contestable, and both cut against the finding that per-share growth tends to lag the economy because new issuance dilutes existing holders. Their dividend yield of 1.7 percent is also above today's 1.1 percent, since they forecast the yield rising as the multiple falls.

Neither approach is the rigorous one. J.P. Morgan build up from corporate fundamentals and are exposed to whether those assumptions hold. This page starts from valuation and is exposed to whether valuation reverts. The evidence above says valuation is the better documented signal and that the level it implies has been too gloomy.

What this model cannot tell you

Every forecasting method has failure modes. These are the ones that matter here.

Coverage and updates

The forecaster currently covers 45 equity markets and regional aggregates, which is every market where both a CAPE distribution and a published index yield are available. Markets missing either one are left out rather than filled with an estimate.

Both datasets refresh automatically. Research Affiliates publish month-end CAPE figures in the first days of the following month, and MSCI refresh their factsheets a few working days after month end. Portfolio Lab checks both weekly and raises an alert if either dataset falls behind, so the dates shown on the forecaster are the dates the numbers actually carry.