US stock valuations are near a 100-year high. This research never tries to predict the next drop. It asks: how should you hold the stocks you already own, at these prices, so that you can stick to your plan through the next decline? Results are judged on staying with the plan, then on the reward for the risk, not on raw returns.
Berkshire's cash. At June 30, 2026, Berkshire Hathaway held $365.5 billion in cash and Treasury bills, more than the $323.8 billion of listed stocks it owned (10-Q filing). That is 53% of its cash plus listed stocks, but only 29% of its $1,263 billion of total assets: most of what Berkshire owns is whole companies, such as its railway, utilities and insurers. Warren Buffett's 2024 letter (February 2025) says "the great majority of your money remains in equities. That preference won't change." Greg Abel is now CEO. Berkshire is not predicting a crash. It is refusing to pay prices it thinks are too high, and waiting.
Your version of the question. Almost every stock investor holds US stocks, directly or through funds, so record prices affect whole portfolios, not just new money. This work began with a narrower question: should a bonus or an inheritance go into the market all at once when prices are high? The bigger issue turned out to be the money already invested. US stocks have fallen 20% or more times since 1926, about once a decade. The useful question is not "will I time the market?" but "is my stock allocation right for my goals and my tolerance for losses, at these prices?"
What the usual advice misses. "Invest now" wins about two-thirds of the time, averaged over cheap and expensive markets alike. It answers "all now, or in pieces?", not "how much in stocks, at these prices?". It is incomplete at today's extremes.
The analogy. Would you bet the same amount on a hockey game before the puck drops as with one team up by three goals late in the third period? The score is new information, and it changes the odds. Valuation is the market's score: it is strongly linked to the next decade's returns, and today it sits near a record. But no lead is safe. In Game 7 in 2013 the Leafs led Boston 4-1 with 10:42 left, and lost in overtime. High prices raise the odds of a poor decade; they don't guarantee one. That is why the practical solution reduces stocks rather than abandoning them.
How expensive is "expensive"? This report uses the Shiller price-to-earnings ratio (CAPE): the price of the market divided by its average earnings over the past 10 years, after inflation. It is the most-used measure of how expensive stocks are over the long run. Today it is , higher than in of months since 1881.Valuation is the Shiller CAPE ( today, the percentile since 1881), ranked in real time against the history known at each month. The rate-adjusted version, the Excess CAPE Yield (ECY), is used where interest rates matter.
What "cash" means here. The research standard is Treasury bills (1-month US, 3-month Canadian). Investors hold high-interest savings accounts and cashable GICs instead. These usually track T-bills within about a point; in Aug 2025 big Canadian banks paid 2.3% on investment savings accounts against a 2.66% 3-month T-bill (Globe and Mail, 22 Sep 2025). The longer Length shows how the answers move if cash earns up to a point less, or half a point more.
What is tested. US stocks since 1926, Canadian stocks since 1956, and Japan's 1989 bubble as a case study. Everything is total return (dividends reinvested). Every valuation ranking uses only the history known at the time, so no result relies on hindsight.
Terms. Every term is defined in Terms, at the end.
The claim tested. At extreme valuations, "stay fully invested, whatever the price" is incomplete advice. A simple rule, written in advance, that trims stocks only when valuations are extreme (the CAPE in its highest 10%) and the market's price trend breaks, can help investors stick to their plan through the next decline, with a smaller worst case, at little long-run cost.
What would disprove it. The hypothesis fails if, with the practical solution, investors were less likely to stay with their plan, if the risk-adjusted value were worse than simply holding stocks, or if the result were no better than simply owning less stock all the time. All three were tested. Six Red Team reviews (one by Claude, five by OpenAI's GPT-6 model) re-ran the key numbers from the raw data and attacked the conclusions. Their corrections are built in.
A positive average is not enough. A 50/50 bet where $100,000 becomes either $150,000 or $75,000, while cash stays at $100,000. The average outcome is $112,500, above cash. But for a moderately cautious investor, losing $25,000 hurts more than gaining $50,000 helps: they would swap the bet for a sure $91,000. That sure amount is the bet's risk-adjusted value (its certainty equivalent). Because it is below $100,000, this investor is better off in cash, even though the average is higher.
History at the dearest prices looked much like that bet. From the 10% of starts where stocks were dearest relative to bond yields (lowest ECY), stocks beat cash over 10 years of the time; on average they ended versus cash, and their risk-adjusted value was . That group is thin: start months, essentially two episodes (1929–30 and 1997–2001), so treat it as an illustration, not a rate. Today's reading sits in the group, where the risk-adjusted edge over ten years was only ; measured on each asset's real ending wealth instead, it is , so treat it as small either way.
Reward per unit of pain. From the highest-CAPE 5% of starts, the next decade's reward per unit of pain fell from about to (median Martin ratio). Today the rate-adjusted measure is , lower than of months since 1881.
How to read it. Each 10-year outcome is stocks' ending wealth divided by cash's, so 1.0× means "same as cash". Expected value is the plain average of those outcomes. Risk-adjusted value is the sure result an investor would accept instead of the whole spread of outcomes. It depends on how much the investor dislikes losses: in the example above, an aggressive investor values the bet at exactly $100,000, the same as cash, and a moderate one at $91,000. When risk-adjusted value falls below 1.0×, that investor prefers cash to being 100% in stocks, despite a positive average; a partial stock allocation may still suit them. Technically: constant relative risk aversion, γ = 2 (aggressive) and 4 (moderate). Bands use ECY's real-time percentile; overlapping windows, so only a few independent decades per band. The certainty equivalent weighs the spread of ending outcomes; the Martin ratio weighs the pain along the way.
Left axis: the rate-adjusted valuation known at each month. Right axis: the reward per unit of pain realized over the next 10 years from that month (so it ends in 2016). When the valuation measure was low (1929, 1965–72, 1996–2000), the following decade paid little or nothing per unit of pain.
Rank correlation between each gauge (known at the start month, oriented so higher = cheaper) and what the next 10 years delivered, same sample for all ( to starts). ECY ranked first on reward-to-risk and on real return, but the margin over the investor-equity-allocation measure is small and in-sample, and ECY is not statistically distinguishable from the plain CAPE yield (1/CAPE). What sets ECY apart is that it accounts for interest rates. Two other gauges are at record highs today: US investors' equity allocation () and market value relative to GDP (×). Sources: Shiller data; Federal Reserve Z.1 via FRED (NCBEILQ027S, FBCELLQ027S, TCMDO, GDP), used with a 3-month publication lag.
Three numbers, in order: staying with the plan, the risk-adjusted value, and the four big crashes. The risk-adjusted value is from starts where stocks were dear relative to bond yields, like today.sf (2-year Stick-to-Plan Factor, the gate), then the 10-year certainty equivalent from ECY < 25th starts (the prize), then crash-window sf (the stress test). Worst decline is a reference column. US stocks and cash, 1926–2026.
"Risk-adjusted value" is explained with the example above: it is the sure amount a cautious investor would accept instead of the uncertain result. The full-exit version (all of the stock allocation to cash when the rule fires) is shown for reference only; the practical solution is the default.
Better than simply owning less stock. Owning less stock always feels calmer: a plain 60/40 mix stayed in the comfort zone of the time. So the real test is whether a rule beats a plain mix holding the same average amount of stock. The practical solution averaged in stocks. Against a plain mix with that much stock, the practical solution scored higher on both counts: points more Stick-to-Plan Factor, and a risk-adjusted value higher by of cash's 10-year result. A plain mix trims stocks all the time; the practical solution trims them only when valuations are extreme and the trend has broken, which is when the deep, long declines happen.
Why 2-year periods. Investors and advisors review portfolios about once a year (Benartzi and Thaler, 1995). Most people tolerate one bad year; few stay with a plan through two bad reviews in a row. At 1-year periods the practical solution looks no better than buy-and-hold ( vs ); the advantage appears at 2 years and grows at 5 years ( vs ).
The crash test. In the four big declines (1929–32, 1973–74, 2000–02, 2008–09), halving stocks still means falling more than 20%, so the practical solution stayed in the comfort zone in only of crash periods (buy-and-hold: ). The full-exit version managed . If the goal is to stay calm in a crash, halving may be too little; that trade-off is the main reason the full-exit version is shown.
The other side of the ledger. Seen from the practical solution's seat, buy-and-hold's pain is crashes: its worst 2-year stretch was points behind (to ). Seen from buy-and-hold's seat, the practical solution's pain is fast recoveries: its worst 2-year stretch was points behind (to ), missing the 1932–34 rebound. Both regrets are real, and they come at different times.
How to read it. Above zero, buy-and-hold was ahead over the previous 2 years (the rule's pain: a fast recovery it partly missed). Below zero, the rule was ahead (buy-and-hold's pain: a crash). The worst moments for each sit at the ends of the spikes.
Drawdown over time. How far each strategy sat below its previous high, every month since 1926. Depth and duration are what make investors give up.
Rolling returns, every window. Average, worst and best annualized returns over every rolling 1- to 10-year window, Portfolio-Visualizer style. The worst 10-year window for the practical solution was a year, against for buy-and-hold.
A variant being tested. A "practical fat pitch" (never below half in stocks; fully invested only when prices are reasonable and the trend is up) scored better on the same data. But it was picked from about 30 rules tested on that data, so it is being tested on five markets it has never seen (Australia, Germany, the Netherlands, Sweden and Switzerland). Until then, the practical solution stays the default.
Definitions. sf = share of rolling 24-month windows with drawdown from the running high no worse than −20% and a non-negative window return. It is a descriptive comfort proxy, not a measured probability of adherence; at 12-month windows the practical solution and buy-and-hold tie. CE uses constant relative risk aversion γ = 4 on each 10-year terminal wealth relative to T-bills, from the 165 starts with ECY below its real-time 25th percentile (overlapping, a handful of episodes). "Matched static mix" compares against a constant stock/T-bill mix with the same average equity weight.
Staying with the plan. In the decade after the Dec 1999 peak, buy-and-hold spent of its 2-year periods in the comfort zone; the practical solution .
The reward for the risk. Buy-and-hold ended the decade with of what cash would have given; the practical solution with . In dollars, $100,000 became , or with the practical solution, and the worst decline along the way was instead of .
US stocks, total return, Dec 1999 to Nov 2009. The rules check the 10-month price trend at each month-end and act the next month, with a 0.1% trading cost. 100/0 holds cash when out of stocks; 60/40 holds bonds.
Not just one lucky month. Across all start months in the highest-CAPE 5% (1928–30, 1961–66, 1995–2004), halving finished ahead of holding equities of the time (60/40: ), by a median per $100,000. Holding only cash all decade finished ahead just of the time, which is why the practical solution reduces stocks rather than abandoning them. Limits: those months come from only three episodes; the expensive starts of 2017–2021 can't be scored yet and are strongly positive so far; and trend rules lag in rising markets (2017–2026: vs a year for the full-exit version). 1999 is the full-exit version's best case.
Why every period matters. A conclusion drawn from one start date can reverse with a different end date: that is end-date bias, and it is behind many confident claims about investing, so every month from to was tested.
The core finding, every period. From the highest-CAPE 5% of starts, the practical solution ended ahead of buy-and-hold in of ten-year periods, by a typical per $100,000; in 8 of 10 periods the result was between and , and the worst was . At a random point in the cycle, by contrast, it helped in only of periods, and the typical difference was . Its value is concentrated exactly where today's market is: compare the two views below.
Across all periods, the practical solution mostly did nothing, cheaply. The CAPE never reached its 90th percentile in of periods, and the practical solution ended within 1% of holding equities in . The worst case (a start: a strong bull market with nine trend breaks) was ; the best (a start) was ahead. Without the valuation filter, halving on every trend break trailed holding equities in of periods, which is why the filter matters.
From expensive starts, the practical solution ended ahead in of periods, and behind in . The worst (a start) was . These periods come from only three episodes, so treat the rate as an illustration, not a promise.
Staying with the plan, every period. From the same expensive starts, the practical solution kept investors in the comfort zone more often in of periods. That is of , and less often in , by a typical ( of 2-year periods, against ). Across all periods it mostly changed nothing: the same in of , better in , worse in .
The reward for the risk, every period. From the same expensive starts, buy-and-hold's risk-adjusted value was versus cash, and the practical solution's . Across all starts: vs . These differ from the findings above ( and ) only because the groups differ: here, the highest-CAPE 5% of starts; there, starts where stocks were dear relative to bond yields.
Each bar is one start month: green = the practical solution ended ahead of holding equities, red = behind, grey = the same, within 1% (mostly because the CAPE never reached its 90th percentile, so the practical solution never acted). Height = the difference in ending wealth. Overlapping 10-year periods share most of their data, so periods are only about independent decades.
Valuation arms the rule; only the trend acts. Extreme valuations alone never trigger a sale, and a falling price alone never does either. The first version re-checked valuation every month, so falling prices switched the first version off in the middle of a crash (1930, 2008). The fix, adopted after two Red Team reviews, stays halved until the trend recovers: the worst decline fell to (vs ). Details: A Practical Behavioural Solution for High Markets.
The full-exit version, for reference. The same signals, but all of the stock allocation moves to cash. The full-exit version cut the 2000–02 loss from to and stayed calm in far more crash periods, but it is a bigger step for most investors and advisors to take and hold. Halving on every trend break, with no valuation filter, cost almost nothing ( vs a year) and cut the worst decline from to , but trailed holding equities in most periods.
Thin history, US only, not a forecast. 100 years of data add up to about independent decades, and every "expensive start" result rests on 3 to 5 episodes. A test on five markets the rules have never seen is under way. None of this says when a decline will come.
Today's prices may be a new normal, or not. The typical valuation has drifted up: median CAPE was about before 1990 and in 1990–2016. Ranking still works (cheap and dear still lined up with the next decade's returnsrank correlation with the next decade's real return ), but a since-1881 benchmark now calls the market "expensive" almost all the time: of history sits below today's CAPE, against of the last 30 years. The causes of the drift (accounting changes, lower rates, a different mix of companies) are not verified here.
Taxes. Switching inside an RRSP or TFSA costs no tax. In a taxable account, every sale can realize gains; the halving version has not yet been tested after tax.
The hypothesis began as a set of claims. Each is checked against the evidence.
Each dot is one start month (every third month shown), 1881–2016. Across: starting CAPE. Up: the next 10 years' return per year, after inflation. The line is the fitted trend; the dashed marker is today.
Higher prices, lower returns. Each doubling of CAPE lowered the expected 10-year return after inflation by points a year (in-sample R² ). At today's CAPE of the two model forms give and a year. The right edge of this chart is almost empty: only 1929 and 1997–2001 started above CAPE 30 and have finished their 10 years.
Real-time vs hindsight. Real-time ranks each month's CAPE against history known at that month, the only version an investor could have used. Hindsight ranks against the full 1881–2026 record. Bars show the share of starts where stocks ended 10 years behind cash. The label over each bar is the number of independent decades behind that bar (neff = months ÷ 120).
Bayes' rule applied to 1,081 US start months. "Expensive" = the highest 10% of CAPE readings, in real time.
Knowing the score changes the odds. Expensive starts were of the starts where stocks lost to cash, but only of those where stocks won: a ratio of about . Updating the odds with that ratio ( × = ) gives a chance of losing to cash of about from an expensive start. That is real information, like being two goals up in the third period. It still points to a minority outcome, and the "game" has been played only 1–3 times at today's score.
Being right half the time is very different from being right 95% of the time. The familiar odds hold up, and each one drops sharply from expensive starts. US total market, month-end returns, 1926–2016 starts. (The core-evidence section uses Shiller's S&P Composite, so the same group can differ by a few points between sections.)
95% confidence drops to a coin flip. Stocks rose in of rolling 12-month periods since 1926 ( of calendar years), and 10-year periods were positive of the time before inflation ( since 1881). After inflation, the 10-year figure is . From the highest-CAPE 5% of starts, those become , and . The expensive column rests on about independent decade(s).
The same tests with cash earning the T-bill rate plus or minus a spread. Savings-account and GIC interest is taxed like T-bill interest, so the after-tax results carry over.
Small edges move easily. The expensive-start answer is close to a coin flip, so half a point of cash yield moves it visibly (from to ). The conclusion doesn't change: from extreme prices, stocks' edge over cash is small and uncertain.
The reward disappears; the risk stays. From the highest-CAPE 5% of starts, the reward per unit of downside risk fell from to , while the ups and downs along the way barely changed ( vs a year). The typical worst decline inside the 10 years deepened from to . (The Advanced Detail shows the full risk tables and the textbook stock-allocation calculation.)
Sharpe = excess return over T-bills ÷ volatility. Sortino = the same, but it counts only months below the T-bill return as risk. Both are computed from the monthly returns inside each forward 10-year window, then the median is taken across start months. CVaR (conditional value at risk) = average of the worst 10% of 10-year real returns.
The reward disappears; the risk stays. The median Sortino ratio falls from for all starts to from top-5% starts, while median volatility barely moves ( vs ). The typical worst drawdown inside the window deepens from to .
The textbook answer (Merton): stock share = expected excess return ÷ (risk aversion × variance). Volatility is held at its 1926–2026 level (), and each row changes only the expected excess return. Risk aversion 2 is aggressive, 4 moderate, and 6 conservative. A financially independent investor who doesn't need more wealth usually sits at the conservative end.
Estimation risk matters. For a moderately risk-averse investor, the textbook stock allocation falls from about on history to at today's valuation. Barberis (2000) shows that predictability does change the optimal allocation, but an investor who ignores the uncertainty in the forecast "may overallocate to stocks by a sizeable amount." The calibrated row is the honest middle: the raw CAPE forecast, shrunk by how well it has actually predicted out of sample. That calibration is itself flattered by the upward drift in CAPE since 1990, so the truth likely lies between the calibrated and raw rows. This is a one-period, parametric illustration, not an estimate of anyone's personal optimum.
2026 Ontario combined marginal rates applied to TSX and Canadian T-bill history (– starts). In a taxable account, interest is taxed every year at the full rate. Eligible dividends get the dividend tax credit. Price gains are deferred for the whole 10 years and then taxed at the 50% inclusion rate. Inside an RRSP or TFSA none of this applies, and the pre-tax results hold.
The exposure most Canadians hold in their US funds: the US total market converted to CAD, vs Canadian T-bills, $200k bracket, – starts. US dividends are foreign income, taxed at the full rate (the 15% withholding is creditable), so the only tax edge is deferral of price gains. From the highest-CAPE 5% of US starts, US stocks beat taxable CAD cash only of the time ( pre-tax). The tax edge helps at every other valuation level; it does not rescue the expensive case.
Deferral is the overlooked edge. Even in the high-rate 1970s–80s, when T-bills beat the TSX pre-tax in about a third of starts, after-tax cash almost never won, because tax on interest consumes the real yield. Today a Canadian T-bill at nets about to a year after tax and 2.1% inflation. In a taxable account, deferred gains let the TSX beat taxable cash of the time. Assumptions: today's rates applied to past decades, and an end-of-period net capital loss credited at the gains rate. US-stock dividends are taxed as foreign income (not modelled).
A caution on method. A relative valuation rule (dividend yield ranked against Japan's own history since 1950) flagged "extreme" as early as , 30 years and +4,000% before the peak, because yields trended down for decades. Absolute thresholds, or valuation combined with trend, avoid that trap. Data: FRED Nikkei 225 month-end (a price-weighted index), dividend yield and cash rates from the Jordà–Schularick–Taylor database, so returns are approximate.
The behavioural lesson. The real damage came from buying heavily near the top and selling in despair near the bottom. A written rule, set in advance, is the practical defence against both.
The same model as above, re-run each month using only data available at the time, compared with the 10-year return after inflation that followed. The "historical average" column is what a planner assuming normal returns would have used. In Dec 1999 the model forecast a year; the result was .
Unbroken runs of start months where stocks ended 10 years behind cash. The 1964–66 losing starts were in the highest-CAPE 5%; the 1967–73 losses began at more ordinary prices, where high inflation and high cash yields did the damage. That is the case for also checking the rate-adjusted signal.
These starts don't have 10 years of results yet, so the tables above leave them out. Every one began with the CAPE in its highest 10%, and every one is far ahead of cash so far.
A claim that circulates widely, including in AI-generated answers. The table tests it: the worst month-end decline in US total return within 3 years of each episode's first month at CAPE ≥ 35.
The claim overstates both the size and the timing. None of the finished episodes fell 40–60% within 3 years. The declines were . The larger 2000–02 fall () came more than 3 years after CAPE first crossed 35 in 1998.
After allowing for interest rates, today's reading is , lower than all but the cheapest of history, so dear but not extreme. In 1999–2000 it fell to . On this measure today sits in the "10th–25th" row, not the bottom rows where stocks always lost.ECY today is , at the real-time percentile, so dear but not extreme. In 1999–2000 it fell to . On this rate-adjusted measure today sits in the "10th–25th" row, not the bottom rows where stocks always lost.
US total market (month-end) vs 1-month T-bills. Signals use only data known at month-end, with CAPE lagged one extra month. Each switch costs 0.1%. No taxes. "Beats lump sum" compares 10-year ending wealth across every start month; the "expensive starts" column is limited to the starts in the most expensive 10%.
Valuation plus trend, not valuation alone. Valuation alone cost return ( vs ), the pattern Asness, Ilmanen and Maloney call "sin a little." Trend alone cut the worst decline from to at the cost of about points a year. Valuation + trend (the full-exit version here) led at , but trailed lump sum slightly in 1976–2026 ( vs ); its edge is mostly 1929–32. Six rules were tested here, so treat the winner with multiple-testing suspicion. Averaging in over 12 months trailed lump sum in of the highest-CAPE 10% of starts and of the highest-CAPE 5%, similar to the 64% PWL Capital reports for the highest-CAPE 5% of startsat the 95th percentile.
Canadian stocks trailed Canadian T-bills over 10 years in of starts (–). The losing starts clustered in 1964–73 and 1980–87, when T-bills yielded 6–20%. The TSX's 1999–2000 CAPE of about 42 still beat cash over the following decade, helped by the commodity boom. Today's Canadian T-bill yield is against inflation, a real cash yield of on trailing inflation, or about against FP Canada's 2.1% inflation guideline.
Signal: US CAPE (real-time). Return: US total market converted to CAD, vs Canadian T-bills, – starts. The highest-CAPE 5% (1995–2004 starts) lost to cash about two times in three, because the Canadian dollar rose against the US dollar through 2002–2011 on top of the tech unwind.
Price indexes understate what investors keep. A price-only reading of the TSX misses roughly a third of the median 10-year return. Taxes widen the gap between stocks and cash for a Canadian in a taxable account: interest is fully taxable each year, while capital gains are deferred and half-included, and eligible dividends earn the dividend tax credit.
| Source | What it found | Bearing on the hypothesis |
|---|---|---|
| Asness, Ilmanen & Maloney (2017), Market Timing: Sin a Little, JOIM | US 1900–2015: a CAPE tilt (50–150% stocks) beat buy-and-hold before 1958 but not after (5.4% vs 5.5% excess, 1958–2015), with no Sharpe gain. Adding 12-month momentum lifted the Sharpe ratio from 0.38 to 0.43 and cut drawdowns. | Valuation alone is weak because "cheap" and "dear" drift for decades. Add momentum and keep tilts small. |
| Palazzo (2026), The CAPE That Cried Wolf, SSRN 6900766 | CAPE's post-1992 break is largely accounting: R&D expensing and special items. The adjusted CAPE-H restores out-of-sample predictability. As of Dec 2025 CAPE and CAPE-H both sat near their 90th percentiles, and both implied about a 60% chance of a 5-year correction. | The 2011–2020 false alarm doesn't discredit today's reading. Both measures now agree. |
| Estrada (2026), Multiples for Valuation: Go High, Go Low, Ignore the Middle, SSRN 6152048 | US 1871–2025: CAPE-yield correlation with 10-year real returns is 0.66 in the top and bottom quartiles and 0.09 in the middle. Out of sample, 0.70 vs 0.22. | Directly supports the hypothesis: valuation carries information at the extremes, which is where the market is today. |
| Felix / PWL Capital, lump sum vs DCA (Rational Reminder #418) | Six markets, 10-year windows: lump sum beat 12-month DCA about 65% of the time; DCA cost about 0.38%/yr. At the 95th percentile of real-time US CAPE, DCA still trailed 64% of the time (71% unconditional). With hindsight percentiles, lump sum still won 54%. | Valuation was tested, and averaging in is not the lever. It says little about stocks vs cash. |
| Vanguard (2023), Cost averaging: invest now or temporarily hold your cash? | Lump sum beat cost averaging about two-thirds of the time, 1976–2022. The paper does not condition on valuation. | This is the source of the "2/3" rule of thumb. It is an unconditional base rate. |
| Luskin (2017), FPA Journal, DCA using the CAPE ratio (per the FPA abstract page; paper not read) | S&P 500, 1950–2015, 15-year DCA vs lump sum: DCA won about a third of periods overall and every period that started above CAPE 31. | Supportive, but every "CAPE above 31" start comes from one episode (1997–2000), and DCA over 15 years is a very different bet. |
| Barclays QIS (2017), The Many Colours of CAPE | At a top-decile CAPE (31 at the time), subsequent 10-year real returns averaged 0.9%, with a best case of 5.8% and a worst of −6.1%. | Same conclusion as the core evidence here: the centre is near zero and the range is wide. |
| Globe and Mail, 2026 (coverage of constituent-matched CAPE and of market peaks) | A constituent-matched CAPE (index members' own earnings history) forecasts better than the standard one; Palazzo cites Ma et al. (2026) and Commins et al. (2025) for the same fix. Rosenberg: CAPE 39.5 exceeds every peak except 1999–2000. | The same accounting-mismatch critique as CAPE-H, reaching the popular press. |
| Barberis (2000), Investing for the Long Run when Returns Are Predictable, Journal of Finance | Even after parameter uncertainty, predictability justifies allocating substantially more to stocks at long horizons. But the evidence is statistically weak, so an investor who ignores estimation risk "may overallocate to stocks by a sizeable amount." | Why the allocation table shows a calibrated forecast next to the raw one, and does not treat the conditional estimate as known. |
| Pfau (2015), Long-Term Investors and Valuation-Based Asset Allocation, SSRN 2544636 (per a search-result summary; not verified at source) | CAPE-based allocation centred on 50% stocks delivered returns comparable to 100% stocks with substantially less risk, across many risk measures, for conservative long-term investors. | For conservative investors, valuation-based allocation improves risk-adjusted outcomes. |
| Kitces & Pfau (2015), Retirement Risk, Rising Equity Glide Paths, and Valuation-Based Asset Allocation, Journal of Financial Planning | When retirement starts in an overvalued market, an accelerated rising-equity glide path offers downside protection. In most scenarios valuation-based allocations supported higher sustainable withdrawal rates than a fixed allocation, and T-bills beat longer bonds as the safe asset from high valuations. | Directly addresses different objectives: retirees and sequence risk, not maximum terminal wealth. |
| Benartzi & Thaler (1995), Myopic Loss Aversion and the Equity Premium Puzzle, QJE | The equity premium fits investors who evaluate their portfolios about once a year. | Why staying with the plan is measured over 2-year periods: two annual reviews in a row. |
| Morningstar (2025), Mind the Gap; Fulkerson, Jordan, Riley & Yan (2026), Financial Analysts Journal | Morningstar: the average dollar in US funds earned 7.0% a year vs 8.2% for the funds (10 years to 2024), a 1.2-point "behaviour gap". The FAJ rebuttal finds that poor timing costs only about 0.10% a year once the gap's mechanics are separated out. | Bad timing is costly in stories and smaller in measured averages. The damage is concentrated in specific investors and specific crashes, which is why a pre-committed rule matters more than a statistic. |
| Campbell & Shiller (1998; 2001 update, NBER w8221) | At the start of 2000, valuation-ratio regressions suggested "real stock returns below zero, over the next ten years"; their dividend-yield model implied a 55% real price decline. | Direction right, magnitude too extreme, and early: the warnings started in 1996. |
| Welch (2000), Views of Financial Economists on the Equity Premium, Journal of Business | 226 finance academics surveyed in 1997–98 forecast a 7% a year arithmetic equity premium over 10 and 30 years. | The consensus assumption just before the lost decade. |
| Boudoukh, Richardson & Whitelaw (2005), The Myth of Long-Horizon Predictability | With persistent predictors, long-horizon R² largely repackages short-horizon predictability, inflated by overlapping observations. | This is why every group here reports its independent decades and episodes, not month counts. |
| Goyal & Welch (2008); Campbell & Thompson (2008), as cited by Asness et al. | Valuation predictors mostly fail to beat the historical mean out of sample. Sensible restrictions recover modest value. | Matches our out-of-sample results: positive before 1990, strongly negative since. |
| Clare, Seaton, Smith & Thomas (2017), FAJ, trend following and CAPE | US 1872–2014: trend following cuts sequence risk. CAPE at the start of retirement helps set a sustainable withdrawal rate. | Supports valuation + trend as a risk tool, especially for retirees. |
| Shen (2002), Kansas City Fed RWP 02-01 | Switching out of stocks when the E/P-minus-short-rate spread was extremely low beat buy-and-hold, 1970–2000, after costs. | Rates matter. That is consistent with our Canada result and the rate-adjusted table. |
| Zakamulin (2014; 2016) | Reported moving-average timing gains are largely look-ahead bias and missing costs. Done properly, they are marginal. | Why this backtest trades the month after the signal and charges a cost per switch. |
| Maggiulli (2022), Of Dollars and Data | Since 1926, buying more after declines only improved forward returns after 40%+ drawdowns. | Relevant to buy-after-a-decline trigger thresholds. |
| AQR (2025), buy-the-dip study (from a secondary summary; not read at source) | Holding cash for dips underperformed in more than 60% of 196 implementations, with 18.7% lower wealth than averaging in. | An unconditional result. This report re-tests the question on the most expensive starts. |
| FP Canada / IFP (2026), Projection Assumption Guidelines | Canadian stocks 6.3%, US 6.4%, cash 2.4%, inflation 2.1%. Shiller earnings yield is one of five equal inputs, and 0.5% is deducted. | Planners' default assumption is only lightly valuation-adjusted. |
How the measures are distributed. Each measure reads like the Benefit panels: the best, the line 1 in 10 results beat, the typical (median), the line 1 in 10 fell below, and the worst. CAPE and the Excess CAPE Yield are monthly since 1881; the 2-year measures are for buy-and-hold since 1926.