Standard deviation

1. Find the mean: (4 + 8 + 6 + 5 + 3 + 7 + 9 + 2) ÷ 8 = 5.50.
2. Subtract the mean from each value, then square that difference — this turns every deviation positive so they don't cancel out when you average them.
3. Add up the squared differences: 42.00 in total.
4. Divide by 8 for population variance (5.25), or by 7 for sample variance (6.00).
5. Take the square root to get back to the original units — 2.29 for population, 2.45 for sample.

The middle column always sums to zero — that's baked into what a mean is, and it's a quick way to check your own arithmetic. If your differences don't sum to roughly zero, go back and recheck the mean.

Dividing by n−1 instead of n always produces a slightly bigger number — here it's the difference between 2.29 and 2.45. The reason: a sample's mean is calculated from the same data it's measuring spread against, so the data sits a touch closer to its own mean than it would to the true population mean. Dividing by the smaller n−1 (Bessel's correction) compensates for that. The gap matters most with small samples — with 8 points it's a real difference; with a few hundred it would barely register.

Use the population version only when your numbers really are the whole group — every student in the class, every unit made this shift. The moment the data is a subset standing in for something bigger, which covers most surveys, experiments, and quality checks, use the sample version instead.

A standard deviation of 2.29 against a mean of 5.50 says values here typically land about 2.29 away from average — a fair amount of spread relative to a mean under 10. The same 2.29 sitting under a mean of 5,000 would barely register. That's what the coefficient of variation is for: standard deviation as a percentage of the mean, so spread can be compared fairly across different scales. Here it's (2.29 ÷ 5.50) × 100 ≈ 41.7%.

The clearest way to see why this matters: two datasets can share an identical mean and still tell completely different stories. Take two classes of five students, both averaging exactly 70 on a test.

Class A's average of 70 tells you roughly what every student scored. Class B's average of 70 hides a room split between students who struggled and students who aced it — a very different teaching problem. The average alone can't distinguish the two; the standard deviation does, immediately.

When data is shaped roughly like a bell curve, standard deviation predicts how it clusters with unusual consistency: about 68% of values fall within one standard deviation of the mean, 95% within two, and 99.7% within three. Applied here, roughly 68% of values would typically fall between 3.21 and 7.79 if the underlying process were normal — worth treating as illustrative rather than solid with only 8 points. It doesn't apply at all to skewed or multi-peaked data.

A z-score turns this into a per-value measurement: (value − mean) ÷ standard deviation, telling you how many standard deviations a specific point sits from the average. The largest value here, 9, has a z-score of 1.53 — about 1.53 standard deviations above the mean. Under the empirical rule, anything beyond roughly ±2 starts looking unusual for normally distributed data, which is the logic most outlier-detection and fraud-screening rules are quietly built on.

• Using the population formula on data that's actually a sample — inflates confidence in a number that should carry more uncertainty.
• Comparing standard deviations across datasets with very different means without converting to coefficient of variation first.
• Applying the 68-95-99.7 rule to data that's visibly skewed or has more than one peak.
• Treating variance and standard deviation as interchangeable — variance is in squared units and doesn't mean much on its own.
• Reporting a mean with no measure of spread at all, which hides whether the underlying data was consistent or all over the place.
• Trusting a small sample's standard deviation as a precise, final number rather than an estimate that would shift with more data.

A fund's historical standard deviation of returns is one of the standard shorthand measures of how volatile it's been — two funds can post the same average annual return while one swings far harder above and below it along the way, and standard deviation is what separates them on paper. A factory tracks the standard deviation of a measurement over time on a control chart, because a rising figure often signals a developing problem before the average itself drifts off target — two production lines can both hit their target on average while one quietly produces far more units outside tolerance.

Standardized test scores get reported with a cohort standard deviation so a single score can be judged against how spread out everyone else's results were, not just against the average. Scientific papers report it alongside every mean, because a result is far less convincing without some sense of how consistent the underlying measurements actually were. And most machine learning pipelines standardize input features using exactly this calculation before training a model, so that a feature measured in thousands doesn't quietly dominate one measured in single digits.

With only 8 points, both 2.29 and 2.45 are honest calculations on the data given, but they're noisy estimates of whatever the "true" spread would look like with more data. Add or remove a single unusual value from a small dataset and the standard deviation can shift noticeably; the same swap barely moves the number once you're working with hundreds of points. There's no fixed threshold where a dataset suddenly becomes "big enough" — it depends on how much the answer needs to be relied on — but treat anything under about 10-15 points as a useful first read rather than a settled figure, and get more data before leaning on it for a real decision.

Standard deviation condenses a whole dataset's spread into one number, in the same units as the data itself. The mean tells you where the centre is; the standard deviation tells you how much individual values actually vary around it — report one without the other and you're only giving half the picture. Whether you're sizing up investment risk, watching a production line, or grading a test, it's the same four steps every time: mean, squared differences, average, square root.

Figures are computed live from the numbers you entered, or the illustrative dataset shown above. Both population and sample standard deviation are calculated since the right one depends on whether your data is the entire group of interest or a sample drawn from a larger one. The empirical rule percentages are approximations that hold exactly only for a truly normal distribution — treat them as a rule of thumb, especially with a small dataset.