What Do Equity Return Distributions Look Like?

Key takeaways

• Of 1,191 rolling one year windows in the US stock market since 1926, 29.5% lost money after inflation. Of the 843 thirty year windows, none did.
• The median outcome hardly moves with time: 10.4% a year in real terms over one year, 7.8% over ten years, 7.2% over thirty.
• The bad end does the travelling. The fifth percentile real return runs from -24.8% a year at one year to 4.8% a year at thirty.
• Vanguard's global record agrees. Of the 120 calendar years from 1901 to 2020, 32 lost money, against an average of 7.3% a year.
• Monthly returns are not bell shaped. Nine of 1,203 months moved beyond four standard deviations, where a normal curve expects 0.08 of one.

Three one year holdings in ten have lost money, and the worst lost 62%

You want the odds, not an opinion. If you hold a broad stock index for a year, or for ten years, or for thirty, how often has it lost money? What has the middle outcome been? And how bad has the worst one been?

Here's the answer, from the longest record anyone can download and check. Ken French's data library publishes the value weighted monthly return on US incorporated common stocks listed on the NYSE, AMEX and NASDAQ, running from July 1926 to September 2026. Deflate that by the US consumer price index and you get 1,191 overlapping one year windows. Of those, 29.5% ended with less purchasing power than they started with. The median window gained 10.4%. The worst, the twelve months to June 1932, lost 62.0%. The best, the twelve months to June 1933, gained 174.6%.

Run the same exercise over thirty years and the shape changes completely. The record holds 843 overlapping thirty year windows. Not one lost money in real terms. The weakest of them, the thirty years to November 1991, still compounded at 4.4% a year after inflation.

Those two rows are the ends of the range. Everything between them is the subject of this page.

The odds table, one year to thirty years

Every row below is a slice of the same equity return distribution, cut at a different holding period. The index, the century and the inflation series stay fixed. These are rolling returns measured from every month end rather than calendar years, and they're annualised and stated after US inflation.

The chart above plots the third column. The probability of loss falls from 29.5% at one year to 12.4% at ten, then 2.9% at fifteen, then nothing at twenty and thirty. Now read the fifth column, and notice how little happens there.

One step in that table does something the others don't. Going from three years to five barely improves the odds: 21.3% of three year windows lost money and 20.3% of five year windows did. The worst case improves sharply over the same stretch, from -38.4% a year to -15.4% a year. So five years has been long enough to shrink the size of a bad outcome and too short to make one unlikely, which is an awkward place for a plan to sit.

The upside narrows too, and that half gets less attention. The best one year window gained 174.6%. The best thirty year window compounded at 11.5% a year. A long holding period hasn't only removed the chance of the worst outcome, it has removed the chance of the spectacular one, and both effects come out of the same arithmetic.

The middle barely moves while the ends come in

The median real return starts at 10.4% a year and finishes at 7.2% a year. Twenty nine extra years of holding move the typical outcome by about three percentage points a year. That isn't nothing, but it's small next to what happens either side of it.

The fifth percentile is where the action is. At one year it sits at -24.8% a year. At thirty years it sits at 4.8% a year. One window in twenty came in below each of those figures, and the same measure flips from a serious loss to a solid real gain purely by lengthening the holding period.

The worst case loss follows the same path. At ten years the floor is -5.0% a year, set by the decade to February 2009. At fifteen years it's -1.9% a year, set by the fifteen years to July 1982. At twenty years the floor turns positive, at 0.2% a year, set by the two decades to August 1949. That one deserves a moment. An investor who bought at the end of August 1929 and held until August 1949 earned 0.2% a year above inflation for twenty years of patience.

So a long horizon has historically compressed the spread of annualised outcomes without lifting the middle of them. Whether that counts as risk going away depends on which of the two numbers you were sold, a question taken apart in are stocks less risky in the long run.

Equity return distributions are not bell shaped

A lot of portfolio maths assumes returns follow a normal curve. The record doesn't.

Across the 1,203 monthly observations from July 1926 to September 2026, the standard deviation of monthly returns is 5.3%. Excess kurtosis, which measures how much weight sits in the tails relative to a normal curve, comes out at 7.36. A normal distribution scores zero on that measure. Nine of those 1,203 months moved more than four standard deviations from the mean. A normal curve with the same mean and the same spread expects 0.08 of one month to do that in a century.

The extremes are real months, not statistical artefacts. September 1931 took 28.7% off the US market in a single month. April 1933 added 38.9%.

The one year picture is just as lopsided. Vanguard's record of global equities, built from a world equity total return index to 1969 and the MSCI World index after it, covers the 120 calendar years from 1901 to 2020. Of those, 32 lost money. Another 32 landed between 0% and 10%. Only one year, 1931, fell by more than 40%, and only one, 1933, rose by more than 50%. The average across the whole run is 7.3% a year, which puts the average inside a band that occurred in roughly a quarter of years. The typical year isn't the average year, and that gap is what a volatility vs drawdown comparison is really about.

A different index, a different currency, the same shape

If these equity return distributions were an artefact of one dataset, a second record would disagree with them. It doesn't.

In Vanguard's global series the best single year for a 100% equity portfolio returned 68.9% and the worst lost 41.0%. Stretch to five years and the worst annualised outcome is -9.1% a year. At ten years it's -1.4% a year. At twenty years it turns positive, at 2.9% a year, against an average of 7.5%. Different index, different currency, different century of data, and the worst case still crosses zero somewhere between ten and twenty years.

The index that record uses after 1969 is still running. At 30 September 2026 the MSCI World index held 1,249 constituents across 23 developed markets, and had returned 9.02% a year in dollars on a gross basis since the end of 1987. Over that same span its deepest fall was 57.46%, between 31 October 2007 and 9 March 2009. A diversified index of the world's largest listed companies can still lose more than half its value.

For a UK investor there's a closer reference point. The Vanguard LifeStrategy 100% Equity Fund has returned 11.09% a year in sterling since it launched on 23 June 2011, after an ongoing charge of 0.20%. Of the ten twelve month periods to 30 September shown on its factsheet, two were negative: a fall of 0.17% in the year to 2020 and 4.47% in the year to 2022. Two losing years in ten sits in the same region as the 24.4% nominal one year loss frequency in the record above.

The strongest objection: 843 windows are not 843 observations

Here's the criticism that does real damage to the bottom row, and it deserves to be taken seriously.

Overlapping windows aren't independent. The thirty years starting in March 1950 and the thirty years starting in April 1950 share all but one of their months. A record running from July 1926 to September 2026 contains about 3.3 non-overlapping thirty year periods, not 843. The clean zero in the loss column therefore rests on roughly three independent observations, and three observations can't support a probability estimate.

It's worse than that, because the three overlap in what they contain. Every thirty year window beginning before 1930 carries the same crash. Every window beginning in the 1990s carries the same two. The record is one country's century, and it's the century that country won. Survivorship runs through all of it.

What survives the objection is the direction, not the precision. Loss frequency falls with holding period in the record where it can be measured, and the worst case improves at every step in both of them. What doesn't survive is reading 0.0% as a guarantee, or treating 843 as a sample size. The arithmetic of loss recovery is the reason the left tail matters more than its frequency suggests.

What the odds table cannot tell you

Inflation is doing more work here than most tables admit. The US consumer price index rose roughly 19 times between June 1926 and September 2026. Measured in nominal money the same windows look friendlier: 24.4% of one year windows lost money rather than 29.5%, the median one year return is 13.9% rather than 10.4%, and the weakest thirty year window returns 7.5% a year rather than 4.4%. Both sets are true. They answer different questions.

The table is also an index, not an investment. It carries no charges, no dealing spread, no tax and no cash drag. The LifeStrategy fund's 0.20% ongoing charge is the sort of deduction that sits between an index line and an account balance. The figures are in dollars too, so a sterling investor's experience has included currency moves that appear in no column here.

The table also measures lump sums. It follows money invested on one date and then left alone, which isn't how most portfolios are actually built. Regular contributions spread the entry point across many of these windows at once, so a real investor's experience is a blend of rows rather than any single one of them.

And the sample contains no market that shut. Every window in it reopened. This is a backtest of the markets that kept trading, which is not a forecast, and the FCA's past performance rule requires a promotion to carry a prominent warning that "past performance is not a reliable indicator of future results". That rule exists because figures like these are easy to read as a promise.

What would change the conclusion

One thirty year window ending in a real loss would change it, and nothing in the data rules that out. The zero in the bottom row is three independent observations wearing the clothes of 843. A long enough stretch of poor real returns would put a negative number there, and the row would stop looking like a law.

The quieter thing to watch is the deflator rather than the index. The fifteen years to July 1982 returned 5.4% a year in nominal money, which reads as unremarkable rather than alarming. The same window lost 1.9% a year in real terms. Inflation moves the real columns while the nominal ones keep looking respectable, and it arrives slowly enough that nobody dates the damage. That's the failure mode this table is least able to warn you about.