If you had put $100,000 into the S&P 500 in March 1999 and left it there for exactly ten years, you would have had $57,145 in February 2009. Not a bad decade. A decade in which you lost 43% of your money while doing everything you were told to do: diversified, low cost, long term, no panic selling.
That window is not a hypothetical. It is the worst of the 261 overlapping ten-year windows in S&P 500 month-end closing data since December 1994, and it is the number most retirement planning quietly assumes away.
Why one number is never enough
Ask what the stock market returns and you will get a single figure — 8%, 10%, whatever the source prefers. That figure is an average across a specific start date and a specific end date, and moving either one moves the answer enormously.
Rolling returns fix this by refusing to pick. Instead of measuring one ten-year stretch, you measure every ten-year stretch: January 1995 to January 2005, February 1995 to February 2005, and so on through the data. What comes back is not a number but a distribution — and the distribution is the honest answer.
| Holding period | Windows | Worst | Median | Best | Negative |
|---|---|---|---|---|---|
| 1 year | 369 | −44.43% | 11.57% | 47.46% | 21.4% |
| 3 years | 345 | −17.51% | 9.45% | 28.97% | 19.4% |
| 5 years | 321 | −8.18% | 8.82% | 25.58% | 25.9% |
| 10 years | 261 | −5.44% | 6.16% | 14.86% | 10.3% |
| 15 years | 201 | +1.99% | 6.33% | 13.40% | 0% |
Three things in that table are worth sitting with.
The spread narrows, but slowly. One-year outcomes range across 92 percentage points. Ten-year outcomes still range across 20. Time compresses the distribution; it does not collapse it.
Ten years is not safe. More than one in ten ten-year windows finished negative. If your plan needs the money at a fixed date a decade out, a 10% chance of a negative annualised return is a real planning constraint, not a footnote.
Fifteen years is where it changes. Zero negative windows, and a worst case of +1.99% a year. That is the closest thing to a floor in this dataset — and note it is a floor that barely beat cash, not a floor that made anyone rich.
The useful question is not “what does the market return?” but “what is the worst outcome I can survive, and how long must I be able to wait?”
The lost decade, in detail
The worst ten-year window ran from March 1999 to February 2009 and lost 5.44% a year. It is worth understanding why, because the mechanism was not one crash but two.
An investor entering in March 1999 bought near the top of the dot-com bubble. The subsequent collapse took roughly half the index. The recovery through 2003 to 2007 was genuine and substantial — and then the financial crisis removed it again, bottoming in early 2009. Ten years, two full round trips, and the investor ended with 57 cents on the dollar.
The Nasdaq-100 version is worse. Its worst ten-year window, February 2000 to January 2010, lost 8.57% a year: $100,000 became $40,803. That index needed more than fifteen years to reclaim its March 2000 level.
Both figures exclude dividends, which is a real limitation — a dividend-reinvesting S&P investor would have done meaningfully better than −5.44%, though still poorly. The direction of the point does not change.
What this should change
Not, for most people, whether to invest. The median ten-year window returned 6.16% a year before dividends, and the fifteen-year floor was positive across the entire dataset. Equities did their job for anyone who could wait.
What it should change is three narrower things.
Sequence risk deserves respect near retirement. A drawdown in the first years of drawing down a portfolio does far more damage than the same drawdown a decade earlier, because withdrawals lock in the loss. The SWP calculator shows how quickly a portfolio fails when withdrawals meet a bad decade.
Money with a deadline does not belong in equities. A house deposit needed in four years sits in a distribution where a quarter of historical windows were negative.
Be suspicious of any projection using a single growth rate. A retirement plan built on “8% a year” is not wrong so much as incomplete: it describes one path through a distribution that contains the lost decade.
Check it yourself
Every figure above comes from the same tool, running on real month-end closes. Change the index, change the holding period, and read the numbers for the horizon you actually care about — the worst window matters far more than the average one, and it is right there in the output.