What It Took to Collect the Average: Crashes, Recovery Time, and the Cost of Being Wrong

The companion article established the headline: US equities compounded at about 6.87% a year in real terms from 1900 to 2026, while Treasury bills lost purchasing power over 92 years and the price level rose 44-fold.

That figure is an average. Nobody experienced the average. What people experienced was a sequence — and the sequence contained periods long enough to consume a career.

This article measures the sequence. Same data, same script, same public sources. It covers every decline of 20% or more, how long recovery actually took, how the answer depends on which ruler you measure with, and what the arithmetic says about the cost of allocating part of a portfolio to something that turns out to be wrong.

It is a historical and arithmetical piece. It describes what happened and what follows from it by definition. It is not a forecast, and it is not advice about any security or allocation.

1. Two rulers, two different histories

There are at least four defensible ways to measure "how far down am I":

• nominal price — what the index quote does, which is what the news reports;
• nominal total return — price plus reinvested dividends;
• real price — the quote, adjusted for inflation;
• real total return — dividends reinvested and inflation removed, which is the only one that answers what happened to my purchasing power.

The first and the last disagree profoundly.

Look at the two lines around 1929 and again around 2000. They tell opposite stories, and both are correct.

1929: the famous number is the wrong number

The most-repeated statistic in market history is that the 1929 crash took 25 years to recover. Measured on nominal price, that is right: the peak was August–September 1929, the trough June 1932 at −84.76%, and the index did not regain its old level until September 1954 — 300 months.

Measured in real total return, the same episode looks like this: peak September 1929, trough June 1932 at −76.80%, recovered November 1936. Eighty-six months. Just over seven years, not twenty-five.

Nothing was fudged. The difference is that the 25-year number ignores two things that were unusually large in exactly that period: dividends were reinvested at collapsed prices, and consumer prices fell about 2% a year through the 1930s, so each surviving dollar bought more. Neither shows up in a price index.

2000: the same correction runs the other way

Now the reverse case, which is less famous and more uncomfortable.

On nominal price, the dot-com peak of August 2000 was recovered in May 2007 — 81 months, painful but ordinary. On real total return, an investor who bought at that August 2000 peak did not get back to even until May 2013: 153 months, twelve and three-quarter years. The longest unrecovered stretch in the entire 126-year record, longer than the Depression.

Why the inversion? Because 2000–2013 had the opposite conditions to the 1930s: dividend yields were near historic lows, so reinvestment added little, while cumulative inflation took about a quarter of a dollar's purchasing power (CPI-U rose 34.8% between those two months). The tailwinds that shortened 1929 were headwinds here.

The lesson is not that one ruler is right. It is that "how long does the market take to recover" has no single answer, and the number you get is decided by an accounting choice most people never make consciously.

2. The full record: every real drawdown of 20% or more

Measured on real total return, 1900 to 2026. "Distinct crashes" counts separate declines within one unrecovered stretch, identified by requiring a 20% rally off a low before the next leg counts as its own event.

Thirteen episodes in 126 years — roughly one every ten years. Five of them kept an investor below their previous peak for more than seven years. Two for twelve years or more — the 1973 episode ran to exactly twelve, the 2000 episode to twelve and three-quarters.

Long underwater periods are usually several crashes, not one long slide

The right-hand column matters more than it looks. The two longest stretches were not single events:

• 1973-01 → 1985-01 contained a −50.1% collapse bottoming in December 1974, a substantial recovery, and then a second decline to −37.7% in July 1982.
• 2000-08 → 2013-05 contained the dot-com bust bottoming at −44.9% in February 2003, a four-year rally that never quite reached the old high, and then the financial crisis taking it to −51.8% in March 2009.
• 1929-09 → 1936-11 contained three: the October–November 1929 crash to −33.8%, the far deeper grind to −76.8% by June 1932, and a further setback to −52.6% in March 1935.

This has a specific implication for how these periods were actually lived. An investor in 2007 had spent four years watching a recovery and could reasonably have believed the 2000 episode was over. It was not. The recovery that fails is a recurring feature of the record, not an anomaly — and it is invisible in any table that lists only peaks and troughs.

3. Time is the variable that actually changes the distribution

Every window of a given length since 1900, annualized real total return:

Read down the "worst" column. The full range at one year spans more than 200 percentage points. At thirty years it spans about nine, and every window is positive.

Three things are worth naming precisely, because the loose version of this table is used to argue something it does not support.

The median barely moves. 8.9% → 7.4% → 6.6% → 6.8% → 7.0%. Lengthening the horizon does not raise the expected return. It narrows the distribution. Those are different claims and only the second one is true here.

Ten years is not enough to be safe. 13.8% of ten-year windows finished negative in real terms — about one in seven. The worst of them, March 1999 to March 2009, lost 5.9% a year for a decade.

Twenty years has exactly one failure and it is nearly a tie. One window out of 1,280 finished below zero: June 1901 to June 1921, at −0.22% a year. That is a real fact, not a rounding artifact, and it is the strongest available argument against reading "20 years is safe" as a law rather than as a description of thirteen decades of a single surviving market.

4. The arithmetic of a side bet

Now the question this record actually poses. Most people who read the numbers above still want to try to do better than the average, at least with part of their money. That is a real preference, and the useful thing is not to argue with it but to price it.

The following is arithmetic. It contains no forecast and no assumption about anyone's skill. Given a core compounding at the historical real rate of 6.87% over 30 years, here is what a satellite allocation costs if it underperforms the core — or if it goes to zero.

Each cell shows the shortfall in final wealth against a 100% core, and in brackets the number of years of compounding that shortfall is equivalent to. The bracketed figure is the one worth sitting with: it converts an abstract percentage into the unit people actually care about.

Two properties of this table are exact, not estimated.

The total-loss column is the hard ceiling on the damage. If the satellite goes to zero, final wealth falls by exactly the satellite's share — no more. A side bet cannot cost more than its size. This is the reassuring half.

Underperformance short of disaster costs more than intuition suggests. A satellite that merely lags by 4 percentage points a year — not a catastrophe, roughly what costs and average timing can produce on their own — costs a 20% allocation about 2.2 years of compounding. Nothing dramatic ever happens. The loss is entirely invisible without the comparison.

This table produces no recommended number, and one is not offered. What it produces is the exchange rate: any allocation can be converted into the years of compounding it costs if it fails. The number a given reader finds acceptable is a question about that reader, and this article does not know anything about them.

What the published evidence says about the odds

Three findings from the standing literature, with the credible challenges to each — because the challenges are not fringe and citing the headline without them would be selective.

Most active US large-cap funds underperform the index over long horizons. S&P Dow Jones Indices' SPIVA U.S. Scorecard for year-end 2025 reports that 79% of active large-cap US equity funds underperformed the S&P 500 during 2025, and about 92% underperformed over the trailing 20 years.

That figure is contested on methodology. Research by K. J. Martijn Cremers (Notre Dame), Jon Fulkerson (Dayton) and Timothy Riley (Arkansas), sponsored by the Investment Adviser Association's Active Managers Council, argues SPIVA counts every closed fund as an automatic underperformer, weights all funds equally regardless of size, and benchmarks against an index nobody can actually buy. Adjusting for all three, they put 20-year underperformance at about 55% rather than 92%. Note what survives the correction: the majority still underperform. The critique moves the number from overwhelming to merely unfavourable.

Investors' own timing costs them something, and the size is disputed. Morningstar's Mind the Gap 2025 estimates that the average dollar in US funds earned 7.0% a year over the ten years to December 2024 against the funds' own 8.2% — a gap of 1.2 percentage points a year, roughly 15% of total return, and consistent across recent editions. A 2026 Financial Analysts Journal paper by Fulkerson, Jordan, Riley and Yan — titled, directly, Bad Timing Does Not Cost Investors 15% of Their Funds' Returns — argues the methodology overstates it.

The honest summary: the gap is real and persistently measured, its magnitude is genuinely under academic dispute, and even the low estimates are in the same order as the "lags by 2pp" column above.

On conventional allocation figures

The core-and-satellite framing — index the bulk, ring-fence a small explicitly speculative portion — is standard in the literature, and the satellite is conventionally described in single-digit to low-double-digit percentages of a portfolio. Specific figures are attributed to specific authors, including Malkiel himself, and readers who want a number from an author should take it from that author's own current edition rather than from a secondary summary. This article deliberately does not put a percentage in anyone's mouth, and does not offer one of its own.

5. What the record implies about the experience

The data above is about arithmetic. This section is about what the arithmetic means for what a person actually goes through, because that is where most of the damage documented in §4 originates.

The bad periods are long, not just deep. This is the finding most poorly represented in popular accounts. The 1973 and 2000 episodes each ran over twelve years to full real recovery. A crash is an event; being underwater is a condition, and the condition has historically lasted years. Anyone whose plan is built to survive a drop but not a decade is prepared for the wrong thing.

Recoveries that fail are normal. Five of the thirteen episodes contained multiple distinct crashes. In real time, the second leg arrives after the narrative has already turned optimistic — 2007 after four years of recovery, 1982 after eight. The record offers no way to distinguish a genuine recovery from a failed one while it is happening.

Consistency is not available at short horizons. 31% of individual years were negative in real terms, and 23% of five-year windows. An investor holding for a working lifetime should expect roughly one negative year in three as the base rate. That is the shape of the distribution that produced the 6.87%, not a deviation from it.

The most-quoted number leaves out most of the return. From the companion article: real total return compounded at 6.87% while price alone managed 2.71%. An investor tracking the headline index level is watching the smaller half of their own result, which systematically overstates how bad the bad periods were and how little the good ones delivered.

The gap between fund returns and investor returns is a measured phenomenon. Whether the true figure is Morningstar's 1.2 points a year or the lower academic estimate, both are positive, and both come from decisions made during exactly the periods catalogued in §2. The behaviour and the drawdown table are the same subject.

What follows from all of this is a question rather than an instruction, and it is the question the record actually asks: not "can I tolerate a 50% decline", but "can I tolerate being below my previous peak for twelve years while a plausible story explains why this time it will not recover". Two of the thirteen episodes demanded exactly that, and five demanded more than seven years of it. The first question is about a moment. The second is about a decade, and it is the one the data says to prepare for.

6. Methodology and known limits

Every figure comes from scripts/build_inflation_study.py, which writes data/inflation-study.json. Sources are BLS CPI-U, Robert Shiller's dataset, and FRED — the same three as the companion article, all republishable. No index point levels appear here; only percentage returns, drawdown percentages and growth multiples indexed to 1.00.

The limits that bear specifically on this article:

1. Drawdown depths are understated. Shiller's monthly price is an average of daily closes, not a month-end close. Averaging inside the month clips the extremes. The real 1929 trough on daily closes was deeper than the −76.8% shown; October 1987 barely registers here as a −26.7% event because the crash and much of the rebound fell inside a single month. Recovery lengths, which are this article's main subject, are far less sensitive to this — a month either way on a 153-month episode changes nothing.
2. Monthly resolution. Peaks and troughs are month-stamped. Precise daily dates would differ slightly.
3. The 20% threshold and the distinct-crash rule are conventions. A 15% or 25% threshold, or a different rally requirement for separating legs, would produce a different count of episodes. The rule used is stated so it can be reproduced or disagreed with.
4. Overlapping rolling windows are not independent observations. The 1,280 twenty-year windows come from 126 years of data, so "one window in 1,280 was negative" is not a 1-in-1,280 probability. It is a description of the historical path.
5. Everything is pre-tax and pre-cost. Realised outcomes were lower — notably in the 1970s, when tax was charged on nominal gains during double-digit inflation.
6. One country, and one that survived. The single largest limitation, restated because it applies with more force to a survival-and-recovery article than to any other kind. Markets that closed permanently have no recovery time to report.

References

Data

• BLS Consumer Price Index (CPI-U, CUUR0000SA0)
• Robert J. Shiller, online data
• FRED — 3-Month Treasury Bill Rate (TB3MS)

Active management and investor behaviour

• SPIVA U.S. Scorecard — S&P Dow Jones Indices; the year-end 2025 edition is the one cited
• S&P DJI U.S. Persistence Scorecard — whether past outperformance repeats
• Study Challenges Methodology in SPIVA Scorecard — summary of the Cremers, Fulkerson and Riley critique
• Morningstar, Mind the Gap 2025 — the investor-return gap
• Fulkerson, Jordan, Riley and Yan, Bad Timing Does Not Cost Investors 15% of Their Funds' Returns — Financial Analysts Journal, 2026; the academic challenge to that estimate

Long-run returns

• UBS Global Investment Returns Yearbook — Dimson, Marsh and Staunton; see the 2026 public summary edition
• Burton G. Malkiel, A Random Walk Down Wall Street, 50th anniversary edition (W. W. Norton, 2023)

Disclaimer

This article is published by Elnath Finance Academy for general informational and educational purposes. It is written for all readers and is identical for everyone. It is not investment advice, not a personalized recommendation, not a suggested allocation, and not an offer or solicitation to buy or sell any security. Elnath Finance Academy is not a registered investment adviser.

The allocation table in §4 is arithmetic, not guidance. It states what a given shortfall costs under stated assumptions. It does not identify a correct allocation for any reader, and no figure in it is presented as one.

All figures are historical measurements or deterministic calculations computed from the public sources named above. Historical returns describe the past. They are not a forecast and not a guarantee of any future result. Recovery times in particular are measurements of thirteen specific episodes in one market that recovered; they establish no entitlement to recovery in any future episode.

All returns are gross of fees, costs and taxes. Index, fund and company names are the property of their respective owners and are used here for identification and educational reference only. Elnath Finance Academy is not affiliated with, sponsored by, or endorsed by S&P Dow Jones Indices LLC, Morningstar, Inc., or any index or data provider named here.

Investing involves risk, including possible loss of principal. Consult a qualified professional where appropriate.