Key takeaways
- US stocks in the 1970s and the 1980s had almost the same annualised standard deviation — 13.41% against 13.17% — and nothing else in common: a 50.06% real drawdown against 26.68%, and 83 consecutive months below a prior peak against 24. The 1970s run was still going when the decade ended; measured to its own end it lasted 143 months.
- Standard deviation counts good surprises and bad ones identically. August 1932, a single month that gained 52.43%, accounts for 11.32% of all the squared deviation in the 1,205 monthly returns from February 1926 to June 2026 — more than three times the contribution of November 1929, which lost 26.19%.
- Since January 2020, 10-year Treasuries have been half as volatile as stocks in real total-return terms, 7.14% against 14.03%, and have had the deeper hole: a 37.3% drawdown against 24.5%, and still 34.2% below their April 2020 peak after 74 months.
- Maximum drawdown gets worse the longer you measure, which makes naive comparison between funds misleading. On the same index since 1871, the median maximum drawdown across all rolling 5-year windows is 22.8%; across 40-year windows it is 50.1%. Median standard deviation moves only from 11.79% to 12.48%.
- Sampling frequency moves the answer on its own: over the same ten years of the S&P 500, daily returns annualised by the square root of 252 give 18.12% and month-end returns annualised by the square root of 12 give 15.46% — and the daily series says the worst episode was March 2020 at 33.92%, while the monthly series says 2022 at 24.77%.
Standard deviation measures how much it moved; drawdown measures how far down it went
Fund factsheets report volatility, usually as the annualised standard deviation of monthly returns, and the question is whether that number describes the risk a person actually experiences. It measures something real. It's not the thing you feel.
Here's the cleanest demonstration in the data. Take Robert Shiller's monthly series for the S&P Composite, on a real total-return basis, and measure the 1970s and the 1980s the same way. The 1970s came out at an annualised standard deviation of 13.41%. The 1980s came out at 13.17%. On the number most funds report, those two decades were nearly identical, a quarter of a percentage point apart. The 1980s even had the worse single month: −12.38% against −12.06%.
Now measure the same two decades by how far the portfolio fell from its own high-water mark. The 1970s fell 50.06% from the January 1973 peak to December 1974. The 1980s fell 26.68%, in the four months from August to December 1987. And the stretch of consecutive months spent below a prior peak — time underwater — was 83 months in the 1970s and 24 in the 1980s. That's a difference of nearly five years of waiting, between two decades that a volatility screen would have ranked as twins.
| Metric | 1970–79 | 1980–89 | What it answers | Where it breaks |
|---|---|---|---|---|
| Annualised standard deviation (monthly, ×√12) | 13.41% | 13.17% | How widely returns scattered around their average | Counts gains and losses identically; unchanged if you shuffle the order |
| Downside deviation (monthly, threshold 0, ×√12) | 9.89% | 8.53% | How widely the losing months scattered | Fixes the sign; still blind to whether losses cluster |
| Maximum drawdown | −50.06% | −26.68% | The worst fall from a high-water mark | One realised path; deepens with the length of the record |
| Longest stretch below a prior peak | 83 months | 24 months | How long the wait was | Can be censored by the window; rarely published |
| Share of months below a prior peak | 89.2% | 65.0% | How often the position was worth less than before | Says nothing about depth |
| Annualised real return | −1.34% | +11.48% | What the decade actually paid | Not a risk measure at all |
The table sets the four metrics side by side on those two decades. The conventions behind them matter enough to state before going further.
How these numbers were built, because the convention changes the answer
Every figure above comes from monthly data. Standard deviation is computed on simple monthly returns and annualised by multiplying by the square root of 12; downside deviation is annualised the same way. Drawdowns and time underwater are measured on the same monthly series, so the comparison is like-for-like — mixing a daily drawdown with a monthly volatility is the most common silent error in this whole subject, and the last section of this piece quantifies what it costs.
The series is Shiller's real total-return index: dividends reinvested, CPI-adjusted, monthly, running from January 1871 to June 2026. Two properties matter. Shiller's Yale documentation states that "stock price data are monthly averages of daily closing prices through January 2000, the last month available as this book goes to press". That qualification dates the book, not the method; the sentence appears without it in the current documentation at shillerdata.com, where the file is published. So peaks are shaved and troughs filled in, and every drawdown here is shallower than the daily figure for the same episode. And using real rather than nominal returns deepens drawdowns in inflationary decades. Checked in nominal terms, the 1970s-versus-1980s result survives: standard deviations of 13.28% and 13.02%, drawdowns of 39.16% and 26.04%, and 41 months underwater against 20.
One caveat on the deflator. Shiller's file marks its June 2026 CPI as estimated, and every end-of-sample real figure rests on it. Every other month of that column matches FRED's CPIAUCNS series — CPI-U, not seasonally adjusted — to three decimals. For June 2026 the BLS has since published 333.952 against Shiller's 336.174, which would put the bond position 33.7% below its April 2020 peak rather than 34.2%.
Over the full 1871–2026 sample, that index has an annualised standard deviation of 14.12%, an annualised real return of 7.10%, and a maximum drawdown of 76.80%, from September 1929 to June 1932.
Half of what standard deviation counts is good news
The mechanism behind the mismatch isn't subtle. Standard deviation squares the distance of each return from the average, which means a month that is 20 points above average and a month 20 points below contribute exactly the same amount. Nothing in the arithmetic knows which one hurt.
The scale of that is easy to underrate. Across the 1,205 monthly returns from February 1926 to June 2026, months that came in above the average supplied 47.0% of the total squared deviation. And the single largest contributor in a century of data is not a crash. It's August 1932, a gain of 52.43%, which by itself accounts for 11.32% of the entire sum — against 3.05% for November 1929's 26.19% loss and 1.60% for March 2020's 18.74% loss.
William Sharpe, who put standard deviation in the denominator of the ratio that carries his name, was explicit about the limitation in his 1994 Journal of Portfolio Management paper: the framework "assumes that the mean and standard deviation of the distribution of one-period return are sufficient statistics", and "comparisons based on the first two moments of a distribution do not take into account possible differences among portfolios in other moments". Skewness is a third moment. Standard deviation can't see it.
Downside deviation repairs the symmetry and leaves everything else broken
The obvious fix is to count only the bad half. Downside deviation squares only returns below a threshold — commonly zero — and the Sortino ratio divides return by that instead of by standard deviation. Harry Markowitz proposed the idea himself. In his 1990 Nobel lecture he wrote that in his 1959 book he had put forward "semi-variance" as a measure of risk, and that it "seems more plausible than variance as a measure of risk, since it is concerned only with adverse deviations".
He also gave the reason it never displaced variance, in the same paragraph: "as far as I know, to date no one has determined whether there is a substantial class of utility functions for which mean-semi-variance succeeds while mean-variance fails". Thirty-six years later that's still roughly where the argument sits.
Run downside deviation on the two decades and you can see how much of the gap it closes. The 1970s come out at 9.89% and the 1980s at 8.53% — 1.36 points apart, where standard deviation was 0.24 points apart. It moved in the right direction. It didn't come close to describing a 23-point difference in drawdown, because it still treats each month on its own. A run of twelve consecutive −2% months and twelve −2% months scattered across a decade produce the same downside deviation and completely different holes.
Time underwater is the metric closest to the experience, and almost nobody publishes it
Drawdown answers "how far down". It doesn't answer the question people actually sit with, which is "how long am I going to be looking at this". Time underwater does: the number of months the index spent below a level it had already reached.
On the full Shiller real total-return series, 75.6% of the 1,866 months since 1871 were spent below a prior peak. Three-quarters of the time, in the best-documented equity market in the world, the honest statement was "worth less than it was worth once before". The longest single stretch ran 152 months. The other 24.4% of months are the ones that set a new high.
The chart plots the deepest drawdown inside each decade, with the peak reset at the start of each one. The 1930s show −72.64%, the 2010s −12.36%. Standard deviation across those same decades ran from 9.43% to 14.70%, with the 1930s at 30.44% as the sole outlier. Set the 1930s aside and the nine remaining decades span 5.3 points of standard deviation and 39.4 points of drawdown, measured on the same returns.
Whether this metric appears anywhere a retail investor looks is a separate matter. In the United States, the fund prospectus rules in SEC Form N-1A require a bar chart of annual returns for the past ten calendar years and, following it, the fund's "highest and lowest return for a quarter" over that period. The words "standard deviation", "volatility" and "drawdown" don't appear in the form at all. The mandatory disclosure is closer to a crude drawdown measure than to a volatility measure, and neither shows how long the fund spent below water.
A bond position with half the volatility and a bigger hole
The clearest recent illustration involves the asset people hold to reduce risk. Measured on Shiller's real total-return series for 10-year Treasuries from January 2020 to June 2026, bonds ran an annualised standard deviation of 7.14%. Stocks over the identical window ran 14.03%. On the reported number, bonds were half as risky.
The drawdowns invert it. Stocks fell 24.50% from November 2021 to October 2022 and recovered. Bonds fell 37.34% in real terms from April 2020 to October 2023, and in June 2026 were still 34.2% below that April 2020 level — 74 months underwater, with no recovery yet in the data. The mechanism is visible in the yield: the 10-year Treasury constant-maturity rate hit 0.52% on 4 August 2020 and reached 4.98% on 19 October 2023.
Nothing about the volatility figure was wrong. Month to month, bonds did move about half as much as stocks. They simply moved in the same direction for three and a half years, and standard deviation has no way of registering that, because it's computed on returns considered one at a time and is completely unchanged if you shuffle their order. Drawdown is nothing but order. Our companion piece on how long bear markets take to recover puts numbers on the equity version of that wait.
The case for standard deviation is stronger than its critics usually allow
The serious defence is not that standard deviation captures the experience. It's that maximum drawdown is a badly behaved statistic, and that this matters more than the elegance of the critique.
Standard deviation is estimable from a short sample. Sixty monthly returns give a usable estimate, and the estimate is reasonably stable as the sample lengthens. Maximum drawdown is a single realised number from a single path — one observation, not a distribution — and it can only get worse as you add data, never better. The size of that effect, on the Shiller series since 1871: the median maximum drawdown across all rolling 5-year windows is 22.8%, across 10-year windows 31.7%, across 20-year windows 47.1%, and across 40-year windows 50.1%. Median annualised standard deviation across the same four sets moves from 11.79% to 12.30% to 12.33% to 12.48%.
Put that in fund terms. Take one index, one end date of June 2026, and vary only the length of the record. The trailing 3-year maximum drawdown is 11.7%. The trailing 5-year is 24.5%. The trailing 20-year is 49.9%, and the trailing 30-year is 51.8%. Standard deviation over those same four windows: 10.77%, 12.09%, 12.81%, 12.89%. A fund with a 30-year history shows more than four times the drawdown of an identical fund launched in 2023, on roughly the same volatility. Ranking funds by maximum drawdown without matching their track-record lengths ranks them mostly by age.
And Markowitz chose variance partly because it was tractable, not because it was perfect. His Nobel lecture is candid: variance "came to mind as a measure of risk", helped by the fact that "the variance of the portfolio, that is the variance of a weighted sum, involved all covariance terms". That property is what makes portfolio construction solvable. Drawdown has no equivalent — the drawdown of a portfolio isn't any function of the drawdowns of its parts.
Change the sampling frequency and the worst episode changes
One more reason to distrust a bare volatility number: it depends on how often you sample, and the convention is rarely stated. Using FRED's daily S&P 500 series over the ten years from 1 August 2016 to 30 June 2026 — 2,492 trading days, price only — daily returns annualised by the square root of 252 give 18.12%. Month-end returns from the identical series, annualised by the square root of 12, give 15.46%. Same index, same decade, same price data, a 17% relative difference produced by nothing but sampling frequency.
Sharpe flagged the assumption behind the square-root rule in 1994: the scaling holds when returns "have zero serial correlation", and "underlying differential returns may be serially correlated". A monthly figure that comes out below the daily one is what negative serial correlation over that horizon produces, and that is the pattern in this sample.
Drawdown is worse than merely scaled by this. On daily data, the deepest fall in that decade was 33.92%, between 19 February and 23 March 2020. On month-end data, it was 24.77%, between December 2021 and September 2022. The two frequencies do not disagree about the depth of one episode; they nominate different episodes as the worst thing that happened. The daily series also spent 86% of its days below a prior peak, with a longest stretch of 512 trading days, against 54% of months and 23 months on the month-end series.
What this evidence cannot tell you
This is one country's history, one path, one realisation. The US equity market is the survivor of the sample, and its 7.10% real annualised return since 1871 is not a neutral draw. Shiller's monthly averaging understates every drawdown, and a daily series would be harsher throughout. The Damodaran dataset at NYU Stern, built annually rather than monthly, puts the standard deviation of S&P 500 total returns at 19.40% for 1928–2025 against 15.42% from the monthly Shiller series over the same years — the same asset, four points apart, purely on measurement frequency.
A past drawdown is not a bound. The 76.80% real fall from 1929 is the worst in 155 years of this series, which says nothing about whether a worse one is possible. Time underwater has the same problem in reverse: the current bond spell is 74 months and counting, and its final length is unknown. FRED's SP500 series carries a rolling ten-year window, so its 1 August 2016 start date drops out of the file within weeks. And every metric here is calculated on an index. A portfolio with contributions or withdrawals has a different drawdown from the index it holds, because the amount at risk changes as the path unfolds — the arithmetic of that is in our piece on rebalancing during a crash.
What would change the conclusion
If returns were independent and roughly symmetric, standard deviation would carry almost all the information, and drawdown would be a derived quantity rather than an additional fact. The gap between the two decades exists because returns cluster — bad months arrive next to other bad months. A market where that stopped being true would be a market where the volatility number was sufficient.
If drawdown statistics were reported with matched windows, the strongest objection to them would mostly dissolve. A 10-year maximum drawdown compared against another 10-year maximum drawdown is a fair comparison; the misleading part is the mixing of horizons, not the metric.
If your holding period is genuinely short, the objection carries less weight. Over a few years the two measures carry more of the same information: the median 5-year maximum drawdown of 22.8% is close to twice the median 5-year standard deviation of 11.79%. It's over multi-decade horizons that they come apart, where the 40-year figures are 50.1% and 12.48%.
The number worth watching is the one almost no factsheet contains: how many months a holding has been below its own high, and how many the reader is willing to sit through. LedgerTouch computes drawdown and time underwater on a portfolio's actual path rather than an index's. Whether 83 months underwater is tolerable or intolerable is not a question any statistic answers.