Key takeaways
- In Robert Shiller's US real total return data from 1871 to 2026, the standard deviation of annualised returns fell from 19.14% over one year to 2.90% over 20 years.
- Over those same windows the gap between the best and the worst ending wealth widened, from 6.0x at one year to 13.4x at 20 years.
- A lump sum beat a staggered entry spread over 24 months in 69.9% of 1,081 start months, measured at a common ten-year horizon.
- Staggering cut the spread of those ten-year outcomes by 13.6%, but the worst single result got worse, from 0.601x to 0.555x.
- Bodie priced insurance against a shortfall at 7.98% of capital over one year and 41.61% over 30, at an assumed volatility of 20%.
Does a longer holding period actually make shares less risky?
You'll have been told that shares are risky over a year and safe over 30. It's the most repeated claim in personal finance, and it's built into products. Bodie's paper singles out the rules for target date funds, which he says "discourage stable value investments and encourage investment in stocks".
Half of it holds up. Robert Shiller's series of US real total returns runs monthly from January 1871 to August 2026. Measured on it, the standard deviation of annualised real returns falls from 19.14% across one-year windows, to 5.09% across ten-year windows, to 2.90% across 20-year windows. On that measure a long investment horizon is a powerful risk reducer.
Now look at the money rather than the rate. Across those one-year windows, the best ending wealth was 6.0 times the worst. Across 20-year windows it was 13.4 times. The rate converged. What you'd actually have at the end diverged, and it kept diverging as the horizon grew.
Both halves are true at once, from the same series. That's what time diversification is, and whether it deserves the word depends entirely on which of the two numbers you were sold.
Why time diversification shrinks one number and grows another
Zvi Bodie set the arithmetic out in a 2020 paper, Wishful Thinking About the Risk of Stocks in the Long Run, restating an argument he first published in 1995. One sentence carries the whole mechanism:
The standard deviation of the average rate of return declines with the length of the time horizon because it is an average. If σ is the standard deviation of the annual rate of return for 1 year, and T is the number of years to the time horizon, then the standard deviation of the average annual rate of return for T years will be σ/√T, assuming that returns have no serial correlation.
The other side of it sits in the next paragraph of the same paper: "The standard deviation of final wealth equals the initial wealth times σ√T."
Divide by the square root of the horizon, or multiply by it. Same σ, same T, opposite direction. Nothing in that depends on which market you pick or which century. It's a property of compounding a random series. It also means the two claims people argue about aren't in conflict at all. They're answers to different questions.
Bodie's own measure of risk is the price of insuring against a shortfall. That's a put option struck at the forward price, which pays out if the portfolio earns less than the risk-free rate. Priced with Black-Scholes at an assumed volatility of 20% a year, the insurance costs 7.98% of the investment over one year, 24.84% over ten years and 41.61% over 30. The chart above plots his table. If a long horizon removed risk, the line would slope the other way.
The record horizon by horizon, from 1871
Here is the same Shiller series cut into overlapping windows of five different lengths.
| Horizon | Windows | Standard deviation of annualised real return | Best ending wealth as a multiple of the worst | Windows that lost money in real terms |
|---|---|---|---|---|
| 1 year | 1,856 | 19.14% | 6.0x | 30.4% |
| 5 years | 1,808 | 7.77% | 8.6x | 18.6% |
| 10 years | 1,748 | 5.09% | 11.4x | 11.0% |
| 20 years | 1,628 | 2.90% | 13.4x | 0.1% |
| 30 years | 1,508 | 1.64% | 13.6x | 0.0% |
Read down the third column and a long horizon looks safe. Read down the fourth and it doesn't. Neither column is wrong.
The fifth column is the honest strong point for the long view, and it's worth stating plainly. No 30-year window in this sample lost money in real terms, and only 0.1% of 20-year windows did. The worst 20-year run started in June 1901 and returned -0.22% a year after inflation. The worst 30-year run started a year later, in June 1902, and still returned 1.89% a year.
The worst ten-year window is far more recent. It began in March 1999 and ended at 0.543 times its real starting value. A full decade in a developed market took away 45.7% of purchasing power, and that is what the fourth column is picking up.
Asset diversification pulls holdings apart. A calendar can't do that.
The SEC's description of why diversification works is exact: "Factors or market conditions that may cause one asset class to perform poorly may improve returns for another asset class." That's a claim about correlation, and it's measurable. Here's the correlation matrix of real annual returns for six US asset classes from 1928 to 2025, computed from Aswath Damodaran's dataset at NYU Stern.
| S&P 500 | 10-year Treasury | 3-month T-bill | Baa corporate | House prices | Gold | |
|---|---|---|---|---|---|---|
| S&P 500 | 1.00 | 0.09 | 0.07 | 0.43 | 0.19 | -0.07 |
| 10-year Treasury | 0.09 | 1.00 | 0.57 | 0.74 | -0.00 | -0.00 |
| 3-month T-bill | 0.07 | 0.57 | 1.00 | 0.48 | -0.04 | -0.03 |
| Baa corporate | 0.43 | 0.74 | 0.48 | 1.00 | 0.08 | 0.04 |
| House prices | 0.19 | -0.00 | -0.04 | 0.08 | 1.00 | -0.10 |
| Gold | -0.07 | -0.00 | -0.03 | 0.04 | -0.10 | 1.00 |
Almost every pair sits well below 1. Equities against 10-year Treasuries came in at 0.09, and equities against gold at -0.07. Those are the numbers that make a mixed portfolio steadier than its parts, and the effect arrives in the same year you own both. Full-period figures also hide what happens when it matters most, which is the separate question of diversification in a crisis.
A correlation matrix is what asset diversification buys you. Time diversification has nothing corresponding to it. The autocorrelation of annual real S&P 500 returns over that same 1928 to 2025 period is -0.035 at a one-year lag and -0.153 at two years. That's close enough to zero that one year tells you almost nothing about the next.
Here's the awkward part. That near-independence is precisely the assumption behind σ√T, and it's also the property that makes two assets diversify each other. The difference is what you do with the returns. Asset returns get added together within one period. Time periods get multiplied together. Adding independent things narrows the spread; multiplying them widens it.
Staggered entry against a lump sum, on the same capital
The practical version of this question is entry timing. The SEC defines the usual answer as follows: "Dollar-cost averaging means investing your money in equal portions, at regular intervals, regardless of the ups and downs in the market." So here it is, tested on the same capital over a common horizon.
The data is Kenneth French's monthly US market return series from July 1926 to July 2026, deflated using Shiller's consumer price index. That series is the "value-weight return of all CRSP firms incorporated in the US and listed on the NYSE, AMEX, or NASDAQ", in his own description of it. One unit of capital, two routes. The lump sum goes in during the first month. The staggered entry puts one twenty-fourth of the capital in each month for 24 months, with the uninvested balance earning French's risk-free rate. Both are valued on the same date, ten years after the start month. There are 1,081 start months.
| Real ending wealth after 10 years | Lump sum | Staggered over 24 months |
|---|---|---|
| Median | 2.114x | 2.014x |
| Standard deviation of outcomes | 0.941 | 0.813 |
| 5th percentile | 0.774x | 0.821x |
| Worst | 0.601x | 0.555x |
| Start months ending below their real starting value | 12.6% | 13.0% |
The lump sum finished ahead in 69.9% of those 1,081 start months. That much is the familiar result.
Staggering trimmed the middle and left the floor where it was
The interesting rows are the bottom three. Staggering cut the standard deviation of ten-year outcomes by 13.6%, from 0.941 to 0.813. It lifted the 5th percentile from 0.774x to 0.821x. Both effects are real, and they're what someone spreading an entry is hoping to buy.
Then the worst case, which went the wrong way. It fell from 0.601x to 0.555x, and both routes had their worst run starting in the same month: March 1999. The share of start months ending below their real starting value barely moved, from 12.6% to 13.0%.
Stretching the schedule doesn't rescue it. Over 36 months the standard deviation falls by 20.0% and the share of losing start months drops to 11.3%. The worst outcome is 0.598x, against the lump sum's 0.601x. Three years of instalments moved the floor by less than half a percentage point.
The mechanism is plain once you see it. Bad ten-year windows aren't caused by a bad month. They're caused by a bad decade. Staggering over 24 months moves roughly half the capital into the portfolio a year later than the lump sum does, on average, and a year is nothing against a ten-year regime. Whether lump sum vs cost averaging wins on average is a settled question, and the lump sum wins it. This is the other question. On this sample, staggering did nothing for the tail.
What did move the tail: the valuation you started from
Shiller's cyclically adjusted price/earnings ratio, or CAPE, divides the index by the average of the previous ten years of real earnings. Across 1,628 monthly observations from January 1881 to August 2016, its correlation with the following ten years of annualised real returns was -0.485. Sorting those observations into five equal groups by starting CAPE gives this.
| Starting CAPE | Median real return over the next 10 years | Worst |
|---|---|---|
| 4.8 to 11.1 | 10.64% a year | 1.25% a year |
| 11.1 to 14.6 | 7.17% a year | -4.15% a year |
| 14.6 to 17.5 | 6.47% a year | -4.63% a year |
| 17.5 to 21.3 | 5.59% a year | -3.97% a year |
| 21.3 to 44.2 | 4.58% a year | -5.92% a year |
The medians fall in order across the five groups. That's the variable the entry-timing argument is really about, and it's a decade-long condition rather than a month. Which is exactly why 24 months of instalments couldn't get around it. The August 2026 reading was 41.12, inside the top group. The highest reading in the whole series was 44.20, in December 1999, nine months after the worst ten-year window in the sample began.
The strongest objection to reading it this way
The σ√T result assumes returns carry no serial correlation. If real equity returns mean-revert over long horizons, the spread of ending wealth is narrower than the formula implies, and the insurance prices are too high. Bodie states the objection himself: "Some economists and other observers of the stock market have claimed that stock returns do not follow a random walk in the long run. Rather, they argue, the behavior of stock returns is best characterized as a mean-reverting process."
The evidence here leans slightly that way. The two-year autocorrelation of -0.153 in the 1928 to 2025 annual data is negative, which is what mean reversion looks like. It's also computed from 98 annual observations, which cannot separate it from noise. The CAPE correlation of -0.485 is a stronger version of the same idea, and high valuations have been followed by weak decades often enough to look like a mechanism rather than an accident.
The second objection is narrower and fair. Bodie's insurance costs come from a Black-Scholes calculation at an assumed 20% volatility, not from a market quote. That model assumes no mean reversion, so the table can't serve as independent evidence against mean reversion. Mike Dempsey, Robert Hudson, Kevin Littler and Kevin Keasey made that objection in the Financial Analysts Journal in 1996. Bodie's attempt to identify market risk with the price of a put, they wrote, "leads to a circular argument within the workings of arbitrage-free option pricing models". What the table does show is what the risk would cost if the random-walk assumption holds.
What this sample can't tell you
It's one country, and the one that won the century. The Bank of England's Millennium of Macroeconomic Data carries a spliced UK share price index. In real terms it fell 76.7% between 1968 and 1974, and didn't regain its 1968 level until 1989, 21 years later. That index excludes dividends, so it overstates the damage to a real investor by a wide margin. The shape of it is still a warning that the US windows don't contain. Nominal recovery and real recovery are also different dates, which is the subject of our work on bear market recovery time.
Bodie's paper adds Japan. The Nikkei 225 peaked in 1989 at 38,951, and 31 years later, in 2020, it was still 39% below that peak in nominal terms. A three-decade window with no recovery is not a hypothetical.
A second sample points the other way, and it belongs on the page. Kenneth French's developed markets series excluding the US runs monthly from July 1990 to July 2026. On it, the standard deviation of annualised real returns falls from 17.49% at one year to 0.78% at 20 years, the same collapse as before. But the spread of ending wealth narrows as well, from 3.2x at one year to 2.5x across 193 overlapping 20-year windows. That isn't a refutation of the arithmetic. It's what happens when 36 years of history can't contain a 1929 or a 1999. The widening in the US table needed the full run back to 1871 to appear, which is a fair measure of how much data this argument actually requires.
The window counts overstate what's actually there. Those 1,748 overlapping ten-year windows come from roughly 15 independent decades. The correlation of -0.485 and the quintile medians inherit the same limitation, and the standard errors on all of it are far wider than the sample sizes suggest. A backtest is not a forecast, and this one covers a single market over a single century.
Both routes also ignore tax, dealing costs and fund charges. A staggered entry pays more of all three, which makes its already thin advantage thinner. And none of this measures whether a person actually stays invested, which is the one variable the arithmetic cannot reach.
What would change the conclusion
Three things would.
The first is stronger evidence of mean reversion at long horizons. If real returns reliably pull back to a mean over 20-year spans, the spread of ending wealth narrows and the case for time diversification stops being purely arithmetic. A two-year autocorrelation of -0.153 in one country isn't that evidence. A reliably negative figure at longer lags, in several markets, would be.
The second is the traded price of long-dated shortfall insurance. The 41.61% at 30 years comes from a formula. If real long-dated index puts changed hands far below it, people with capital at risk would be disagreeing with the arithmetic, and that disagreement would be worth more than the model.
The third is the tail itself. If the worst multi-decade windows stopped clustering after high starting valuations, entry timing would stop mattering and staggering would stop being a question worth asking. That hasn't happened in this sample yet. Until it does, the honest description of a long horizon is that it makes your average annual return more predictable and your final balance less so.
The number worth following isn't next year's return. It's whether the top valuation group, where August 2026 sits at 41.12, keeps delivering the 4.58% median decade it has delivered so far.