The Basics of Standard Deviation: A Simple Guide
Updated: Oct 18
Introduction
Two portfolios each averaged 10% a year over a decade. The first never earned less than 8% or more than 12%. The second swung between +50% and -30%. On average, they are identical; as lived experiences, they have almost nothing in common. The standard deviation is a number that tells them apart.
In finance, standard deviation is usually called volatility, and it is the most widely used single measure of risk. It sits inside the Sharpe ratio, mean-variance portfolio optimisation, parametric VaR, option pricing models, and most risk limit frameworks. When a fund factsheet quotes "risk", a trader quotes "vol" or a regulator calibrates a capital charge, standard deviation is usually underneath.
The calculation takes five lines. The difficulty is everything around it: which returns to feed in, how many, how to scale across horizons, how to combine assets, and when the number stops meaning what people think it means. This article covers the calculation properly, then spends most of its effort on interpretation, implementation, and limits.
Practical Challenges
5 problems come up repeatedly, though standard deviation is easy to compute and easy to misuse.
Symmetry vs. Loss Aversion: Standard deviation penalises a return 10 points above the mean exactly as much as one 10 points below it. Investors do not feel them equally. Kahneman and Tversky's prospect theory, central to Kahneman's 2002 Nobel Prize, formalized loss aversion: losses hurt roughly twice as much as equal gains please. A measure that treats upside as risk can mislead whenever returns are skewed, as they are for options, credit, and many hedge fund strategies.
Non-Normal Returns: The familiar "68-95-99.7" interpretation assumes a normal distribution. Real returns have fat tails and, for equities, negative skew. Moves that a normal model treats as once-in-centuries happen every few years. Standard deviation describes the middle of the distribution well and the tails poorly, yet the tails are where risk management lives.
Volatility is Not Constant: Volatility clusters: calm periods stay calm, turbulent periods stay turbulent. A one-year historical standard deviation measured just before a crisis can understate the coming risk several times over, and one measured just after a crisis can overstate it for months.
Horizon Scaling: Annual risk is routinely estimated as daily volatility x √252. That rule assumes returns are independent and identically distributed. Autocorrelation, mean reversion and volatility clustering all break it, sometimes badly.
Estimation Choices: The same price history gives materially different volatilities depending on the return definition, the sampling frequency, the window length, the weighting scheme, and the treatment of stale prices. Two teams can report "the volatility" of the same asset and disagree by a third without either making an arithmetic mistake.
What's Covered
The article moves from concept, to calculation, to judgement.
Variability and risk – why finance treats dispersion as risk.
From variance to standard deviation – the formulas, and why we divide by n - 1.
Example – 10 daily returns, step by step.
Interpreting the number – ranges, the empirical rule, and comparing investments.
Scaling across horizons – annualisation and when the √T rule fails.
Portfolio standard deviation – correlation and the arithmetic of diversification.
Time-varying volatility – EWMA, GARCH and implied volatility.
Special cases / edge conditions – short samples, stale prices, outliers, regime breaks.
Standard deviation in risk assessment – parametric VaR, Sharpe ratios and z-scores.
Limitations and alternatives – where it fails and what to use alongside it.
Best practices and common errors.
Appendix – derivations of Bessel's correction, the √T rule and portfolio variance.
This article assumes you know how returns are defined; the companion piece on absolute, discrete and log returns covers that ground.
Variability and Why Finance Equates It With Risk
Variability (dispersion, spread) is the degree to which observations differ from their central value. For returns, it describes how far a typical period's outcome lands from the average period's outcome.
Finance equates variability with risk through a simple chain: risk is uncertainty about outcomes, and uncertainty shows up as dispersion of possible returns. If an asset's returns are tightly clustered, next period's return is predictable, and there is little risk. If they are widely spread, next period could be very good or very bad, and you cannot plan around either.
This equivalence underpins the risk-return trade-off. Investors demand higher expected returns for holding more variable assets, which is why equities have historically out-earned government bonds, and why a strategy promising high return with low volatility deserves scrutiny rather than applause.
Two caveats travel with the equivalence from the start:
Variability measures uncertainty, not loss. A volatile asset that mostly surprises on the upside scores as risky.
Variability is a property of a distribution, which we never observe. We estimate it from a sample of past returns, and the estimate carries its own error.
From Variance to Standard Deviation
Standard deviation is the square root of variance, and variance is the average squared deviation from the mean. 5 steps get you there.
Mean Return of n observations r1 … rn:
Deviations from the mean, ri - r̄. They always sum to exactly zero, which is why we cannot simply average them.
Square Each Deviation. This removes signs and weights large deviations more heavily: a deviation of 4 contributes 16, four times as much as a deviation of 2.
Variance, the average squared deviation. With a sample, divide by n - 1:
Standard Deviation, the square root, which returns the measure to the units of the data (percent, not percent squared):
Why n - 1 (Bessel's Correction): Deviations are measured from the sample mean, which is by construction the point that minimises the sum of squared deviations for that sample. Measured from the true (unknown) mean, the sum would be larger. Dividing by n therefore underestimates the population variance on average; dividing by n - 1 removes that bias exactly. The appendix shows the derivation.
The correction matters for short samples and vanishes for long ones: at n = 10 it raises variance by 11%, at n = 250 by 0.4%. Note that s itself is still slightly biased downward even with n - 1, because the square root is concave; for practical sample sizes this is negligible.
Population Formula: Divide by n when the data truly is the whole population, or when a convention requires it. Some risk systems use n, or assume a zero mean for daily returns and divide by n; both are defensible but should be stated.
Example: 10 Daily Returns
Ten daily returns produce a mean of 0.50% and a sample standard deviation of 3.03%.
Step 1: Mean: The returns sum to 5%, so r̄ = 5% / 10 = 0.50%.
Steps 2 and 3: Deviations and Squares:
Day | Return | Deviation from 0.50% | Squared Deviation (%²) |
1 | +2.0% | +1.5% | 2.25 |
2 | -3.0% | -3.5% | 12.25 |
3 | +1.0% | +0.5% | 0.25 |
4 | +5.0% | +4.5% | 20.25 |
5 | -1.0% | -1.5% | 2.25 |
6 | +3.0% | +2.5% | 6.25 |
7 | -2.0% | -2.5% | 6.25 |
8 | 0.0% | -0.5% | 0.25 |
9 | +4.0% | +3.5% | 12.25 |
10 | -4.0% | -4.5% | 20.25 |
Total | +5.0% | 0.0% | 82.50 |
The deviations sum to exactly zero. The squares show the weighting effect: days 4 and 10, each 4.5 points from the mean, contribute 40.5 of the 82.5 total, almost half the variance from two of ten days.
Step 4: Variance: s² = 82.50 / (10 - 1) = 9.17 %².
Step 5: Standard Deviation: s = √9.17 = 3.03%.
With the population formula (divide by 10), the result would be √8.25 = 2.87%. On 10 observations, the choice of denominator moves the answer by 5%, a reminder that short samples carry more estimation error than the extra decimals suggest.
4. Interpretation
A standard deviation of 3.03% around a mean of 0.50% says a typical day lands about 3 points either side of +0.5%, not that returns range from 2.53% to 3.53%. Ranges are always the mean plus or minus multiples of the standard deviation.
The Empirical Rule: If returns were normally distributed:
Band | Probability inside (normal) | Range in the example |
Mean ± 1σ | about 68% | -2.53% to +3.53% |
Mean ± 2σ | about 95% | -5.56% to +6.56% |
Mean ± 3σ | about 99.7% | -8.59% to +9.59% |
In the sample, 6 of the 10 days fall inside one standard deviation and all 10 inside two, roughly consistent with the rule. With 10 observations, that consistency proves very little; the rule's real test is in the tails, and real returns usually fail it there (section 10).
Comparing Investments: Standard deviation is most useful as a relative measure. Three funds illustrate the risk-return trade-off:
Fund | Mean Annual Return | Annual σ | 1σ Range | Chance of a Losing Year (if normal) |
Conservative Bond | 4% | 3% | +1% to +7% | about 9% |
Balanced | 8% | 10% | -2% to +18% | about 21% |
Aggressive Equity | 12% | 20% | -8% to +32% | about 27% |
The aggressive fund earns three times the bond fund's return for nearly seven times its volatility, and a losing year roughly one year in four rather than one in eleven. Which fund is "better" depends on the investor's horizon, liquidity needs and tolerance for drawdowns, not on either number alone. Risk-adjusted measures such as the Sharpe ratio (section 9) combine the two.
Units Matter. A daily volatility, a monthly volatility, and an annual volatility of the same asset are three different numbers. Always quote volatility with its horizon, and compare like with like.
Scaling Horizons
If one-period returns are independent and identically distributed, variance grows linearly with the number of periods T, so volatility grows with its square root:
From | To Annual | Multiply by |
Daily | Annual | √252 ≈ 15.87 |
Weekly | Annual | √52 ≈ 7.21 |
Monthly | Annual | √12 ≈ 3.46 |
Daily | 10-day | √10 ≈ 3.16 |
The example's daily 3.03% annualises to 3.03% x 15.87 ≈ 48.1%. Its mean, scaled the same way, would be 0.50% x 252 = 126% a year: a reminder that means estimated from short samples are far noisier than volatilities, and should rarely be annualised naively.
Why √T and Not T: Independent shocks partly cancel. Over 252 days, some up days offset some down days, so total dispersion grows more slowly than the number of days. Means do not cancel, so they scale linearly. This is also why, under these assumptions, the ratio of mean to volatility improves with horizon.
When the Rule Fails?
Autocorrelation: With serial correlation ρk at lag k, the T-period variance is:
Positive autocorrelation (trending, or smoothed prices in illiquid assets) means √T understates long-horizon risk; negative autocorrelation (mean reversion) means it overstates.
Volatility Clustering: If volatility is high today, it is likely to stay high, so a 10-day risk estimate from today's elevated daily vol differs from one using the long-run average.
Fat Tails and Path Dependence: √T scales the standard deviation correctly under independence, but it does not scale tail quantiles correctly when daily returns are fat-tailed or when positions are rebalanced or hedged over the horizon.
Regulatory frameworks have long allowed √T scaling of 1-day VaR to 10 days, and FRTB's expected shortfall also applies a square-root scaling across liquidity horizons. In both cases, it is a convention for consistency, not a claim that returns are truly i.i.d.
Portfolio Standard Deviation
Portfolio volatility is not the weighted average of asset volatilities; it is lower whenever assets are less than perfectly correlated. That gap is diversification.
For two assets with weights w1, w2, volatilities σ1, σ2 and correlation ρ:
For N assets, with weight vector w and covariance matrix Σ:
Correlation Drives the Result: A 50/50 portfolio of two assets, each with 20% volatility:
Correlation | Portfolio Volatility |
+1.0 | 20.0% |
+0.5 | 17.3% |
0.0 | 14.1% |
-0.5 | 10.0% |
-1.0 | 0.0% |
The Limit of Diversification: With N equally weighted assets of average variance σ² and average pairwise correlation ρ̄:
Idiosyncratic risk diversifies away; the common component does not. With σ = 20% and ρ̄ = 0.3, no number of assets takes portfolio volatility below √0.3 x 20% ≈ 11%.
Risk Contributions: Each position's contribution to portfolio volatility is its weight times its marginal contribution, and the contributions sum exactly to σp. This decomposition, called Euler allocation, is the basis of risk budgeting and of attributing portfolio VaR to desks and positions.
The practical catch. The correlations are estimated, unstable, and tend to rise in crises, exactly when diversification is needed. A portfolio volatility computed from calm-period correlations can materially understate stressed risk, which is one reason stress testing exists alongside volatility-based measures.
Time-Varying Volatility: EWMA, GARCH and Implied Volatility
A single standard deviation over a long lookback window treats a return from three years ago exactly like yesterday's. Because volatility clusters, practitioners use estimators that respond to recent data.
Equally Weighted Rolling Window: The sample standard deviation over the last N days. Simple and transparent, but it reacts slowly, and it produces "ghost" effects: a single extreme day keeps volatility high for exactly N days, then drops out abruptly.
EWMA (Exponentially Weighted Moving Average): Weights decline geometrically with age:
RiskMetrics popularized λ = 0.94 for daily data, which gives a half-life of about 11 days. EWMA reacts quickly to shocks and decays smoothly, without ghost effects. It has no long-run anchor, so forecasts at any horizon simply equal today's estimate. It also typically assumes a zero mean, which is reasonable for daily returns.
GARCH (1, 1): Adds mean reversion towards a long-run variance:
α measures reaction to new shocks, β measures persistence, and α + β measures how fast volatility reverts to its long-run level. EWMA is the special case ω = 0, α + β = 1. GARCH produces term structures of volatility forecasts and is the standard for volatility modelling; extensions such as GJR-GARCH and EGARCH capture the leverage effect, where volatility rises more after falls than after rallies.
Implied Volatility: Derived from option prices rather than past returns, it is the market's forward-looking estimate. The VIX measures the 30-day implied volatility of S&P 500 index options. It typically sits in the teens in calm markets and has spiked above 80 during the 2008 crisis and in March 2020. Implied volatility usually exceeds subsequent realised volatility on average; the difference, the variance risk premium, is the compensation option sellers earn for bearing crash risk.
Estimator | Responds to New Data | Long-run Anchor | Typical Use |
Rolling Equal-Weight | Slow | Implicitly the window | Reporting, regulatory history |
EWMA | Fast | None | Daily VaR, risk monitoring |
GARCH | Fast | Yes | Forecasting, term structures |
Implied | Forward-looking | Market-set | Pricing, hedging, sentiment |
Special Cases / Edge Conditions
Most bad volatility numbers come from the data, not the formula.
Short Samples: The standard error of a volatility estimate is roughly σ / √(2n) under normality. With 20 observations, a true 20% volatility is estimated within about ±3.2 points one time in three; with 250 observations, within about ±0.9 points. Fat tails widen these bands further. Quote short-sample volatilities with humility.
Stale and Smoothed Prices: Illiquid assets, private holdings and appraisal-based valuations update slowly. Their returns look smooth, autocorrelated and artificially low-volatility. Unsmoothing techniques or lower-frequency returns help; taking the reported number at face value does not.
Non-Synchronous Data: Assets that close at different times (Tokyo, London, New York) show depressed daily correlations because news reaches their prices on different days. Use synchronised prices, or weekly returns, when correlations matter.
Missing Data and Holidays: Filling a gap with the previous price records a 0% return that never happened and lowers volatility. Dropping the day and computing the return across the gap is usually better.
Outliers: Squaring gives extreme observations enormous weight, as the example showed. Check them before removing them: a genuine crisis day is the most informative point in the sample. Remove data errors, keep real events, and consider robust estimators (such as median absolute deviation) as a cross-check.
Regime Breaks: A currency peg ending, a central bank changing policy, or a company transforming its business can make history unrepresentative. A window that straddles the break describes neither regime.
Overlapping Returns: Computing 10-day volatility from overlapping 10-day windows uses data efficiently but creates strong autocorrelation between observations, which understates the uncertainty of the estimate and misleads standard statistical tests.
Returns Near Zero or Negative Levels: For rates, spreads, and other factors that can approach or cross zero, compute volatility on absolute changes in basis points, not percentage changes.
Standard deviation in risk assessment
Standard deviation is the input to several of the most widely used risk and performance numbers.
Parametric (Variance-Covariance) VaR: Assuming normal returns with zero mean over a short horizon, VaR at confidence α for a portfolio of value V is:
For a $10m portfolio with 1.2% daily volatility, 99% one-day VaR is 2.326 x 1.2% x $10m ≈ $279,000. The corresponding normal expected shortfall, the average loss beyond VaR, is σ x φ(z) / (1 - α) x V ≈ 2.665 x 1.2% x $10m ≈ $320,000. Both numbers inherit every weakness of the normality assumption; historical simulation and fat-tailed distributions exist precisely to address that.
Sharpe Ratio: Excess return per unit of volatility:
With a 3% risk-free rate, the three funds of section 4 score 0.33 (bond), 0.50 (balanced) and 0.45 (aggressive). The balanced fund delivers the most return per unit of risk, even though the aggressive fund has the highest return.
Tracking Error and Information Ratio: For benchmarked managers, the relevant risk is the standard deviation of active returns (portfolio minus benchmark), called tracking error. Active return divided by tracking error gives the information ratio, the benchmarked equivalent of Sharpe.
Z-scores and Move Sizing: Dividing a move by volatility expresses it in standard deviations. A -6% day on an asset with 1.5% daily volatility is a 4σ move: under normality, that should occur about once every 125 years of trading days. When "4σ" days appear every few years, the problem is the model, not bad luck.
Position Sizing and Limits: Volatility targeting scales positions inversely to volatility so each contributes a similar risk budget. Many desk limits are set in volatility-based terms, and volatility is often the denominator when risk is normalised across asset classes.
Limitations and Alternatives
Standard deviation summarises the centre of a distribution well and its tails, asymmetry and dynamics poorly. Five limitations matter most, each with a practical response.
It is symmetric. Upside surprises count as risk. For skewed payoffs (options, credit, short-volatility strategies), two assets with the same volatility can have very different loss profiles.
Alternatives: semi-deviation, the standard deviation of returns below the mean, and downside deviation, which measures shortfalls below a chosen target (minimum acceptable return, zero, or the risk-free rate). The Sortino ratio replaces volatility with downside deviation in the Sharpe formula.
It assumes the tails are thin. Real returns have excess kurtosis; extreme moves are far more frequent than a normal distribution implies, and equity returns are negatively skewed.
Alternatives: report skewness and kurtosis alongside volatility; use historical simulation or fat-tailed distributions (Student-t) for VaR; use expected shortfall, which averages losses beyond VaR and so measures how bad the tail is, not just where it starts.
It is backward-looking and unstable. Volatility changes with regime, and a historical estimate lags.
Alternatives: EWMA or GARCH estimates, implied volatility, and stressed-period calibration (stressed VaR, or FRTB's stressed expected shortfall).
It ignores path and sequence. Two return series with identical volatility can have very different drawdowns if losses cluster.
Alternatives: maximum drawdown, drawdown duration, and the Calmar ratio, which matter more to many investors than volatility does.
It misses non-linear and correlation risk. Option-heavy books have P&L that is not a linear function of underlying moves, and correlations break down in stress.
Alternatives: full-revaluation historical or Monte Carlo VaR, and scenario-based stress testing that shocks correlations and factors jointly rather than assuming they hold.
Limitation | What Goes Wrong | Complement With |
Symmetric | Upside counted as risk | Semi-deviation, downside deviation, Sortino |
Thin tails assumed | Tail losses understated | Skew, kurtosis, historical VaR, expected shortfall |
Backward-looking | Lags regime changes | EWMA, GARCH, implied vol, stressed calibration |
Path-blind | Drawdowns invisible | Maximum drawdown, Calmar ratio |
Linear and stable correlations | Options and crises misjudged | Full revaluation, stress testing |
Best Practices and Common Errors
Best Practices
State every volatility with its return definition, frequency, window, weighting, and horizon: "annualised, from 1-year daily log returns, EWMA λ = 0.94" is a number someone can reproduce.
Use log returns or daily simple returns for estimation; for daily data the difference is negligible, for monthly or annual data it is not.
Match the window to the purpose: short or exponentially weighted for current risk, long for strategic allocation, stressed for capital.
Annualise volatility with √T, but state the i.i.d. assumption and check autocorrelation for illiquid or smoothed assets.
Clean data before computing: adjust for corporate actions, synchronise across markets, treat gaps and stale prices explicitly.
Always pair volatility with at least one tail or drawdown measure.
Backtest: compare predicted volatility bands or VaR with realised outcomes and count exceptions.
Common Errors
Centring a range on the standard deviation instead of on the mean (3.03% ± 0.5% rather than 0.5% ± 3.03%).
Mixing daily and annual volatilities, or comparing volatilities from different windows.
Averaging asset volatilities to estimate portfolio volatility, ignoring correlation.
Annualising a short-sample mean as if it were as reliable as the volatility.
Using percentage-change volatility for rates or spreads near zero.
Reading a low volatility on an illiquid asset as low risk.
Treating a 4σ or 5σ day as a freak event instead of evidence of fat tails.
Appendix: Derivations
A1. Bessel's Correction: Let r1 … rn be independent with mean μ and variance σ². Adding and subtracting the sample mean:
The cross term vanishes because deviations from r̄ sum to zero. Taking expectations, the left side is nσ² and E[(r̄ - μ)²] = σ²/n, so:
Dividing the sum of squared deviations by n - 1 therefore gives an unbiased estimate of σ². One degree of freedom is used up estimating the mean.
A2. The Square-Root-of-Time Rule: The T-period return is the sum of one-period returns (exactly for log returns). Its variance is:
With identical variances σ² and zero covariances, this equals Tσ², so the standard deviation is σ√T. Non-zero covariances add the autocorrelation term shown in section 5.
A3. Portfolio Variance: The portfolio return is rp = Σ wi ri. Variance is bilinear, so:
Writing Cov(ri, rj) = ρij σi σj and expanding for two assets gives the formula in section 6.
A4. Euler Decomposition: σp is homogeneous of degree one in the weights (doubling every weight doubles σp). By Euler's theorem for homogeneous functions, σp equals the sum of each weight times its partial derivative, which is the risk-contribution identity in Section 6.
Comments