Standard Deviation: Formula, Meaning & Use Cases

Updated 2026.08.01~3 min read

Standard deviation measures spread: ~68% of data falls within ±1σ, 95% within ±2σ (normal distribution). Sample version divides by n−1, population by n. Includes step-by-step calculation, CV for comparing volatility, and real uses from quality control to fund risk.

Key Takeaways

Meaning: the typical distance of data points from the mean — small σ = consistent, large σ = volatile

68-95-99.7 rule: ±1σ holds ~68%, ±2σ ~95%, ±3σ ~99.7% of normal data

Sample vs population: sample SD divides by n−1 (Bessel's correction); population by n

CV (coefficient of variation): σ ÷ mean — compares volatility across different scales (fund A vs fund B)

Core answer: Standard deviation measures spread around the mean: it's the square root of average squared deviation. Data [2, 4, 4, 4, 5, 5, 7, 9] (mean 5): population SD = √(32/8) = 2.0; sample SD = √(32/7) = 2.14. The difference: population divides by N, sample divides by N−1 (Bessel's correction, fixing underestimation from sampling). In a normal distribution, 68% of data sits within ±1σ, 95% within ±2σ, 99.7% within ±3σ.

The step-by-step (population)

Mean: (2+4+4+4+5+5+7+9)/8 = 5
Deviations squared: 9, 1, 1, 1, 0, 0, 4, 16 → sum 32
Variance = 32/8 = 4 → SD = √4 = 2

For a SAMPLE of the same data: variance = 32/7 = 4.57 → SD = 2.14. The N−1 correction matters for small samples and vanishes for large ones.

The 68-95-99.7 rule

Heights: mean 170 cm, SD 7 cm → 68% of people are 163–177 cm, 95% are 156–184 cm. A person at 191 cm is 3σ out — 0.15% territory. This is how control charts, test curves, and risk models all reason.

Worked examples

Example 1 — Comparing consistency. Two shooters both average 9 rings: A's SD 0.3, B's SD 1.5. A is a metronome, B swings between 6 and 12. Same mean, different players — SD is the difference between "average" and "reliable".

Example 2 — Exam standardization. Class A mean 75, SD 5; Class B mean 75, SD 15. A 90 in Class A is +3σ (elite); in Class B it's +1σ (good but common). Comparing raw scores across classes without SD is meaningless — hence z-scores: z = (x − mean)/SD.

Example 3 — Finance volatility. Fund A: 8%/yr mean, SD 10%; Fund B: 8% mean, SD 25%. In a normal-ish world, A's worst 5% year ≈ 8 − 1.65×10 = −8.5%; B's ≈ −33%. Same return, wildly different ride — Sharpe ratio (return ÷ SD) exists to price exactly this.

Example 4 — Quality control. A fill machine targets 500 g, SD 2 g. Spec 494–506 g = ±3σ → 99.73% in spec, ~2,700 defects per million. If SD creeps to 3 g, defects jump to 45,500 per million — that's why factories watch σ like a heartbeat.

Common mistakes and myths

Variance vs SD units — variance is in squared units (cm²) and uninterpretable; SD returns to original units. Report SD, compute in variance.
N vs N−1 casually — Excel: STDEV.P (population) vs STDEV.S (sample). Using .S on a full population understates spread.
Assuming normality — the 68-95 rule fails for skewed data (income, city sizes); use percentiles/IQR there.
SD = average error — it's √(mean of squares), which weights big deviations MORE than a plain mean of |deviations|; that's a feature (penalizes outliers), not a bug.
Comparing SD across scales — SD 5 means nothing without context; coefficient of variation (SD/mean) normalizes: CV 2% vs 20% tells the consistency story across different-sized things.

Frequently Asked Questions

What does standard deviation actually tell me?

The typical spread around the mean. Two classes both average 75 on a test: σ=3 means almost everyone scored 69–81; σ=15 means scores ranged wildly from 45 to 100. Same average, completely different situations. SD is the "consistency" number behind any average.

How do I calculate standard deviation step by step?

(1) Find the mean. (2) Subtract the mean from each value and square the differences. (3) Average the squares — divide by n for population or n−1 for a sample. (4) Take the square root. Example [2, 4, 4, 4, 5, 5, 7, 9]: mean = 5, squared differences sum to 32, population SD = √(32÷8) = 2.

Why does the sample formula divide by n−1?

A sample tends to underestimate the population's true spread because the sample mean is fitted to the sample itself (the deviations are minimized around it). Dividing by n−1 (Bessel's correction) inflates the estimate slightly to make it unbiased. With large n the difference is negligible; with n=10 it matters (÷9 vs ÷10). Use n−1 whenever your data is a sample of a bigger population.

What is the 68-95-99.7 rule?

For a normal distribution: ~68% of values lie within 1σ of the mean, ~95% within 2σ, ~99.7% within 3σ. Heights with mean 170cm and σ=6cm: ~95% of people are 158–182cm. A value beyond 3σ is genuinely rare (0.3%) — the basis of quality control and anomaly detection.

How do I compare volatility between things of different sizes?

Use the coefficient of variation: CV = SD ÷ mean. Fund A: mean return 8%, σ 12% → CV 1.5. Fund B: mean 15%, σ 18% → CV 1.2. B has higher absolute volatility but LOWER volatility per unit of return — the better risk-adjusted profile. Raw SD comparison misleads when means differ.

Is a bigger standard deviation always worse?

Depends on context. In manufacturing, large σ means defective products — bad. In investing, large σ means higher risk but possibly higher return — neither good nor bad alone. In sports analytics, a player with large σ is "streaky." SD measures variability; whether variability hurts you depends on what you need.

Feature	Detail
Keyboard	0-9, + - * / % ( ), Enter, Backspace, Esc
%	50% = 0.5; 200+10% = 220
Chained ops	After =, press an operator to continue with the result
Scientific mode	Switch to the Scientific Calculator for sin/log/√

Feature	Detail
Keyboard	0-9, + - * / % ( ), Enter, Backspace, Esc
%	50% = 0.5; 200+10% = 220
Chained ops	After =, press an operator to continue with the result
Scientific mode	Switch to the Scientific Calculator for sin/log/√