Standard Deviation Guide 2026: Formula, Meaning and Examples
Table of Contents
What Standard Deviation Measures
It measures spread: how far data typically strays from the mean.
Example: Test scores 70, 80, 90. Mean = 80. Deviations: -10, 0, 10. Squared: 100, 0, 100. Variance = 200/3 = 66.7. SD = 8.16.
Scores 78, 80, 82 have the same mean but SD = 1.63 - much more consistent.
Identical means can hide opposite realities. Datasets 78, 80, 82 and 60, 80, 100 both average 80, but their standard deviations are 1.63 and 16.33. A bolt factory averaging 50 mm with SD 0.5 mm ships good parts; the same mean with SD 3 mm ships scrap. The mean aims the process; the standard deviation is the quality number.
Standard deviation speaks the data's own units - 8.16 points, 3 inches, $1,200 - because the final square root undoes the squaring inside the formula. Variance, at 66.7 points-squared, is mathematically purer but humanly useless. Quote variance to statisticians; quote SD to everyone who has to act on the number.
The 68-95-99.7 Rule
For bell-shaped (normal) data:
- 68% within 1 SD of the mean
- 95% within 2 SD
- 99.7% within 3 SD
Heights: Mean 69", SD 3". 68% of men are 66-72"; 95% are 63-75". A 76" man is 2.33 SD above - about the 99th percentile.
Z-scores convert any measurement into a percentile with one division. SAT scores with mean 1050 and SD 200: a score of 1250 gives z = (1250 - 1050)/200 = 1.0, landing near the 84th percentile (50% + 34%). A z of 2.0 reaches about the 97.7th percentile. Test publishers, growth charts, and lab reference ranges all run on this conversion - "two standard deviations above average" is a percentile claim.
The rule also sets manufacturing tolerances. Bolts with mean length 50 mm and SD 0.5 mm: 95% fall within 49 to 51 mm. Specifying +/-2 SD as the acceptable band means only about 5% need rework by design. Quality engineers work backward from the allowed defect rate to the required SD - the empirical rule is the bridge between statistics and the factory floor.
Population vs Sample
| Population SD | Sample SD | |
|---|---|---|
| Divides by | N | N - 1 |
| Use when | You have ALL data | Estimating from a sample |
The N-1 (Bessel's correction) compensates for samples underestimating spread. With large N, the difference is negligible.
The population/sample gap is visible in small datasets. For 4, 8, 15, 16, 23 (mean 13.2): dividing squared deviations by 5 gives SD 6.62; dividing by 4 gives 7.40. Same data, different answer - the N-1 correction inflates the estimate because samples systematically understate spread. With thousands of observations the gap vanishes, but with five it changes the result by 12%.
Choose by what you hold: all 500 employees' salaries means population SD; a poll of 1,000 voters estimating 200 million means sample SD. Spreadsheet and calculator functions default to sample (STDEV.S), so whole-population analyses must deliberately switch to the population variant. The most common SD error is not math - it is using the default.
Volatility: Finance Application
Two funds, both 8% average return:
- Fund A: SD 6% (steady)
- Fund B: SD 18% (wild swings)
Fund B's 95% range in a given year: 8% +/- 36% = -28% to +44%. Same average, terrifyingly different ride. SD is the standard measure of investment risk - the Sharpe ratio divides excess return by SD.
Compute a fund's SD from its returns to see past the advertised average. Returns of +10%, -5%, +15% average 6.67%, but deviations of 3.33, -11.67, and 8.33 give SD near 8.5%. A "6.67% average return" with 8.5% SD is a roller coaster, not a savings account - the mean describes the destination, the SD describes the ride. Always demand both numbers before investing.
Volatility scales with the square root of time. A 3% monthly SD annualizes to 3% x sqrt(12), about 10.4% - not 36%. This square-root-of-T rule is why options prices, Value-at-Risk models, and Sharpe ratios all use annualized SD: risk grows with time, but slower than linearly. Comparing a monthly SD to an annual SD without annualizing is a category error.
Frequently Asked Questions
The typical distance data points fall from the average. Low SD = consistent; high SD = spread out.
Variance is SD squared (in squared units); SD is the square root, back in original units - more interpretable.
Depends on context - there's no universal 'good.' Compare to the mean (coefficient of variation = SD/mean) for scale-free comparison.
SD itself stabilizes with more data, but the standard error (SD/sqrt(N)) shrinks - bigger samples give more precise mean estimates.
No - it's a square root of a sum of squares, always zero or positive. Zero means all values are identical.