Investing at all-time highs: what the record actually shows

“Wait for the dip” is the most expensive sentence in retail investing, and the least examined. It treats a record level as a warning, when a rising market produces records as a matter of course. The historical record allows a colder reading: distributions of returns after all-time highs, side by side with distributions after any other day, both tails included. This page computes them over 155 years, then reads the result through valuation, cost of waiting and the macro regime.

The page assumes the basics are in place; start investing step by step covers the vocabulary and the groundwork, and nothing here re-teaches it. What this page adds is a single discipline: replacing the feeling that “it is too high” with the distribution of what actually followed.

1. The question behind the question: is the market’s level information?

The anxiety has a precise form. An investor with cash sees the index at a record and concludes that the entry point is bad, because what goes up must come down. The conclusion smuggles in three separate claims: that the level predicts the direction, that the record makes a fall more likely, and that waiting has no cost. Each claim is testable, and the tests disagree with the intuition to different degrees. Hesitation at records also keeps company with the biggest beginner mistakes, most of which are timing decisions dressed up as prudence.

Taken one by one, the three claims fare differently. The first, that the level predicts direction, fails cleanly: the one-year distributions below are nearly identical with and without a record. The second, that a record raises the odds of a fall, fails at short horizons and partially survives at ten years, though the working variable turns out to be valuation, not the record. The third, that waiting is free, fails on documented arithmetic: the cost of standing aside is measurable and compounds. What remains of the intuition, once tested, is a single legitimate residue, the long-horizon valuation question, which gets its own section below.

Before the numbers, the grid. Five things can be read at an entry point, and they do not inform the same horizons.

The record itself occupies the weakest line of the grid. All-time highs are what a rising market produces, not what it warns of.

2. What happened after past records

Method first, because the result is only as honest as the definition. The series is the S&P 500 and its predecessors in Robert Shiller’s monthly data, monthly average prices with dividends reinvested, from January 1871 to June 2026; a record month is a month whose average price exceeds every month before it. The underlying series are public: the S&P 500 index series for the level and S&P 500 historical returns data for the annual record. On this definition, 333 of 1,866 months were records.

Two consequences of the monthly convention are worth stating before any number is read. Monthly averages smooth the intramonth noise, so this count is not comparable to the daily record-close tallies quoted in market commentary: a year like 2024 contains 57 daily records (S&P DJI) but far fewer record months, and the 17.8% share of record months below is a different statistic from the roughly 7% of record trading days since 1950. And because the analysis is in total return, a “record” refers to the level an investor’s capital had never reached, dividends included, which is the definition that actually matters to a holder rather than to a headline writer.

One, three, five and ten years later

Over one year, the two branches are nearly indistinguishable: a median total return of 12.6% after record months against 11.3% after all months, with 75% of episodes positive against 73% (Eco3min calculations, same series). Over three years the medians coincide to the decimal, at 32.9%. The tails are the honest part of the picture: one year out, the worst decile after a record still lost 9.1% or more, and records stood weeks before the 1929, 2000 and 2007 peaks; every S&P 500 crash since 1950 catalogues what those episodes cost. A record does not immunize anything. It just fails to predict anything either, at these horizons.

The recoveries deserve their own line, because they carry the horizon lesson better than any average. Measured in total return on the same series, the September 1929 peak was regained in January 1945, fifteen years later; the August 2000 peak in October 2006, six years; the October 2007 peak in August 2012, under five (Eco3min calculations). In price alone, ignoring dividends, the 1929 peak stood for a quarter century, until 1954. Two readings follow. The worst entry months of the last 155 years were record months, and even those were eventually repaired for a holder who could wait and reinvest; but “eventually” has meant fifteen years at the extreme, which is why the grid ranks the ability to stay invested above the level itself.

The picture changes where intuition least expects it: far out. At five years the medians are still close, 61.4% after records against 59.5% overall, but the lower quartile is weaker after records, 10.8% against 21.0%, and the share of positive episodes drops to 81% from 90%. At ten years the gap is frank: a median of 112% after records against 131% overall, positive in 90% of cases against 97%. The record is not the culprit; the valuation that often accompanies late-cycle records is, and that is the subject of the next section. External corroboration exists for the short-horizon half of this result: J.P. Morgan Asset Management, on daily data from 1988 to 2024, finds little difference in one-, three- and five-year cumulative returns between investing at a record close and investing on any other day (analysis published 2025).

All-time highs are not rare events

Nothing about a record is exceptional in a trending market. Since 1871, 17.8% of all months have been record months on this series; on daily closes, J.P. Morgan counts a new all-time high on roughly 7% of trading days since 1950 (cited 2026). The dense years are famous: 77 record closes in 1995, 70 in 2021, 65 in 1964, 62 in 2017, 57 in 2024 (S&P DJI counts). The deserts are the counterpart the statistic hides: on monthly data, not a single record month between the early 1930s and 1950, six in the entire 1970s, nine in the 2000s. An investor’s lifetime contains both kinds of decade, which is precisely why the horizon line of the grid outranks the level line.

Why records cluster

The clustering is mechanics, not prophecy. A record can only occur in a market already near its peak, and markets near their peaks are usually in uptrends that persist for stretches before ending without notice. Each record therefore raises the conditional odds of another record close behind it, which is exactly what the 1995 or 2024 sequences show, and exactly what stopped, without warning, in March 2000. Reading the cluster as a promise that the trend continues is the same error in the other direction, a pattern the extrapolation bias dissects: the series that produced the records says nothing about its own continuation. Both readings, the fearful and the greedy, over-interpret the same data point.

The calendar of clusters is not random either, and it previews the regime section. The dense decades, the 1980s, 1990s and 2010s, were disinflationary stretches of falling rates and expanding multiples; the deserts sit on the inflationary 1970s and on the post-bubble 2000s, when valuation was being worked off rather than built up. Records are a symptom of the environment that produces them, which is one more reason to read the environment rather than the symptom.

3. What CAPE says, and does not say, about the short term

Valuation is where the level regains information, at the right horizon. The cyclically adjusted price-earnings ratio compares the real price to a ten-year average of real earnings; its full monthly history since 1881 is downloadable in the CAPE ratio dataset. The empirical pattern is an asymmetry: CAPE correlates meaningfully with subsequent ten-year returns and barely at all with the next twelve months. High valuations have historically compressed the following decade’s outcomes without saying anything useful about the following year, which is exactly the shape of the record-month results above: no gap at one year, a frank gap at ten.

The current reading deserves to be stated plainly rather than dramatized. At 39.5 in July 2026, CAPE sits in the top percentile of its 145-year history, against a long-run median of 16.6 (Eco3min calculations on the dataset). Historically, decade returns starting from the top of the valuation distribution have clustered below the all-period median; that is a statement about distributions, not a forecast, and the one-year branch of the same data remains as uninformative as ever. The two horizons carry two different messages, and collapsing them into one is where most record-level commentary goes wrong, in both directions.

The asymmetry has a mechanism, and it also has a warning label. Multiples revert slowly, over years, while twelve-month returns are dominated by news that no valuation ratio contains, which is why the signal only accumulates at long range. The warning label is persistence: CAPE has spent 68% of the months since January 2017 above 30 (same dataset), a level once considered extreme, and the stretch from 2017 to 2026 paid handsomely regardless. A high CAPE is not an exit signal and has never functioned as one; it is a base rate on the following decade, and base rates lose to nothing except the habit of ignoring them.

4. The historical cost of waiting

The waiting strategy has a documented price, because exits require re-entries and the market’s best sessions bunch around its worst. In J.P. Morgan Asset Management’s twenty-year window to February 2025, a fully invested S&P 500 position compounded at 10.6% a year; missing only the ten best days cut that to 6.4%, and missing the sixty best turned the twenty-year outcome negative (J.P. Morgan AM; CNBC, April 2025). Seven of those ten best days occurred within fifteen days of the ten worst, which is the statistical way of saying that whoever steps out to avoid the storm usually misses the rebound that pays for it.

The honest objection is the mirror image: an investor who missed the ten worst days would have done spectacularly better, so the statistic proves nothing by itself. Correct, and the overlap is the answer. Best and worst days share the same volatile weeks; a rule that reliably avoids one set without sacrificing the other has to call turning points within days, repeatedly, in both directions. Nothing in the record suggests that skill exists at scale, and the behavioral evidence points the other way: exits happen after the damage, re-entries after the recovery. The asymmetry is not in the arithmetic, which is symmetric, but in what people actually manage to execute.

Cash, meanwhile, is not a neutral waiting room. Its visible yield is nominal, and over the periods a would-be timer typically waits, the purchasing-power arithmetic runs against it; the real vs nominal gap is the lens through which any “I will wait in cash” plan should be priced. The trap compounds with the market itself: each new record raises the level at which the waiter must eventually re-enter, so a strategy designed to enter cheaper has historically tended, in trending stretches like 1995 or 2024, to end up re-entering higher after a delay, or never re-entering at all.

None of this makes any particular moment a good or bad entry; it prices the alternative, which is what the waiting intuition systematically forgets to do.

5. What if I spread it out?

Spreading an investment over months is a different question from timing it, and this page deliberately stays out of it. Timing asks whether the market’s level should change the decision; spreading asks how to schedule a decision already made, and its trade-offs are documented territory: lump sum or spreading it out compares the two approaches on the historical record, both tails included.

The one connection worth keeping is behavioral: for an investor whose real risk is abandoning the plan after an early loss, the scheduling question becomes a staying-invested question, and dollar-cost averaging mechanics and limits lays out what the technique does and does not deliver. Nothing in this page’s distributions changes either analysis.

6. After a record: the distribution of returns

The simulator below shows the two branches of the historical record side by side: total returns 1, 3, 5 or 10 years after a record month, against the same horizons after any month, on the Shiller monthly series from 1871 to 2026. The boxes mark the median and interquartile range, the whiskers the 10th and 90th percentiles; the readouts state the share of positive episodes. Read it as a shape, not a verdict: where the two boxes overlap almost entirely, as at one year, the record carried no usable signal; where they part, as at ten years, the divergence is the valuation effect from section 3 showing up in the raw data. Everything displayed is retrospective, a description of what happened, never a projection of what will.

7. Read through the macro regime

The regime is the one layer where the environment, as opposed to the level, carries usable information. The distributions above pool 155 years of very different worlds; conditioning on the regime narrows them. Inflationary regimes have historically compressed equity multiples and hardened the valuation constraint, a mechanism documented in the inflationary regime atlas; disinflationary stretches did the opposite, which is part of why the 1980s and 1990s dominate the record-count table. What each configuration did to the major asset classes is compiled in how assets performed under each regime.

The current stretch belongs to the dense side of the historical table: the index crossed 7,000 points for the first time in April 2026, another entry in a record sequence running since 2024. Placing the environment behind it is a measurement question: where the regime stands today is read from the Eco3min classifier, rules-based and published. The regime does not tell an investor whether this record is a ceiling; nothing does. It tells them which historical sub-sample their environment most resembles, which is a humbler and more useful thing.

8. FAQ

What happened after past S&P 500 all-time highs?

On Shiller monthly data from 1871 to 2026, the median total return twelve months after a record month was 12.6%, against 11.3% after any month, with three-year medians identical at 32.9% (Eco3min calculations). At ten years the relationship inverts and records underperform the all-month median, reflecting the valuations that often accompany them. Records also preceded the 1929, 2000 and 2007 peaks: the distribution has two tails.

How often does the index set new highs in a bull regime?

In dense years, constantly: 77 daily record closes in 1995, 70 in 2021, 57 in 2024 (S&P DJI counts). Since 1950, roughly 7% of all trading days have been record closes (J.P. Morgan, cited 2026). The counterpart is the droughts: on monthly data, no records at all from the early 1930s to 1950, and single digits across the 1970s and the 2000s.

What does CAPE say about short-term returns?

Historically, almost nothing. The ratio’s correlation with the next twelve months is weak; its information concentrates at the ten-year horizon, where high starting valuations have been followed by below-median decades on average. The July 2026 reading of 39.5 sits in the top percentile of the series against a median of 16.6, a statement about the shape of the ten-year distribution, not about next year.

What has waiting in cash historically cost?

Two costs, documented separately. Missing the market’s best sessions: over the twenty years to February 2025, skipping the ten best days cut annualized returns from 10.6% to 6.4% (J.P. Morgan AM). And the cash itself: its yield is nominal, and in real terms extended waiting has tended to erode purchasing power. Neither figure says the market cannot fall after a record; both price the alternative to being invested.

How is timing different from dollar-cost averaging?

Timing conditions the decision on the market’s level; averaging schedules a decision already made. The first asks “is now a bad moment”, the second “in how many steps”. The historical comparison between spreading and investing at once is a separate, documented question, and how effective DCA really is treats it directly. For sizing the periodic amount itself, how much to invest monthly covers the budgeting side. Conflating the two questions is how a scheduling tool gets mistaken for a timing signal.

9. The level is a headline; the horizon is the decision

What the 155-year record supports is narrower and more useful than either camp’s slogan. At one to three years, a record month has been statistically ordinary; at ten years, the valuations that ride along with records have mattered, and pretending otherwise would be the mirror image of the “wait for the dip” error. The layers that actually rank an entry decision sit elsewhere: the horizon, the regime, and the allocation the entry serves. Market entry read through the cycle connects the decision to the environment, strategy before the trade puts it in its architectural place, and the 2026 investment landscape maps the vehicles it can land in. An investor who has settled those three layers will find that the record on the front page has quietly stopped being a question; what remains is a schedule and an allocation, neither of which reads the tape. All-time highs are what a rising market produces, not what it warns of.

Last updated: 8 July 2026

Last updated — 8 July 2026