How to interpret standard deviation

If you need the definition first, start with what standard deviation is. The steps below assume you already have a result and want to know what it says. They use one running example: the delivery times, in minutes, of 20 orders from a single restaurant last month. The data is loaded in the calculator so you can check every figure.

Step 1: pin down units, scale and version

The standard deviation is in the same units as the data, so write the unit down before anything else. For the deliveries the mean is 802 ÷ 20 = 40.1 minutes and the sample standard deviation is 7.61 minutes. Also confirm which version you have. The sample SD (divide by n − 1) is 7.61; the population SD (divide by n) is 7.42. For 20 orders standing in for the restaurant's usual performance, the sample version is the right one; the sample vs population page explains the choice.

Scale also means the possible range of the measurement. An SD of 12.4 on a test marked out of 100 is a wide spread; the same 12.4 on a 1,000-point scale is narrow.

Step 2: compare it with the mean

Dividing the SD by the mean gives the coefficient of variation, which removes the unit:

CV = s / x̄ = 7.61 / 40.1 = 0.190 = 19.0%

A typical delivery differs from the average by about a fifth of the average. That is easier to reason about than "7.61 minutes", and it lets you compare spread across things measured differently. Two caveats. There is no universal "good" CV: an analytical lab may reject a method above 2%, while 20% is ordinary for delivery times, sales or reaction times. And the CV only makes sense for data with a true zero; a CV of temperatures in °C changes if you switch to °F, so it means nothing.

Step 3: sanity-check it against the range

A quick way to catch a wrong input or a wrong formula is the range rule of thumb. For moderately sized, roughly bell-shaped data, the standard deviation is usually somewhere near a quarter of the range. For large samples (hundreds of values) the range stretches further and a sixth of the range is the closer guide.

range = 61 − 28 = 33 range / 4 = 8.25 range / 6 = 5.5

The computed 7.61 sits comfortably near 8.25, so nothing looks off. This is a rough check, not an estimate to report. If your SD is several times larger or smaller than range ÷ 4, look for a typo, a unit mix-up or a variance reported as an SD. The range calculator gives the minimum, maximum and range in one step.

Step 4: compare with a benchmark, a specification or last period

The most useful comparison is with something you already care about. The restaurant promises delivery within 50 minutes. With a mean of 40.1 and an SD of 7.61, the limit is (50 − 40.1) ÷ 7.61 = 1.30 standard deviations above the mean. If the times were normally distributed, about 9.7% of orders would run late; in the data, 2 of 20 (10%) did. The SD has turned into a late-delivery rate, which is what the owner actually wants to know.

The other common comparison is with a previous period. Suppose the month before had an SD of 5.9 minutes. Has the service become less consistent? Step 7 shows why 20 orders are not enough to say so with confidence.

Step 5: use z-scores to place individual values

The standard deviation is also a ruler for single observations. A z-score divides a value's distance from the mean by the SD:

z(61) = (61 − 40.1) / 7.61 = 2.75 z(28) = (28 − 40.1) / 7.61 = −1.59

The 28-minute order was quick but unremarkable. The 61-minute order is almost three standard deviations out, and worth a look at what happened. For bell-shaped data, the 68–95–99.7 rule says how often values that far out should appear; that page covers it, so it is not repeated here.

Step 6: separate what it says about individuals from what it says about the average

The 7.61 minutes describes how much individual deliveries vary. It does not describe how precisely you know the average. That is the standard error:

SE = s / √n = 7.61 / √20 = 1.70 minutes

With the t critical value for 19 degrees of freedom (2.093), the 95% confidence interval for the mean delivery time is 40.1 ± 3.56, or 36.5 to 43.7 minutes. So the typical order is somewhere around 40 minutes, but any one order could easily land 8 minutes either side of that. Mixing the two up is the most common interpretation error in published work: the standard error shrinks as you collect more data, the standard deviation does not.

Step 7: ask how stable the SD itself is

A standard deviation from a small sample is itself uncertain, more than most people assume. The confidence interval for the true σ comes from the chi-square distribution, and its width depends only on n:

For the deliveries, 7.61 × 0.76 and 7.61 × 1.46 give a 95% interval for σ of about 5.8 to 11.1 minutes. Last month's 5.9 sits inside it, so this data alone does not show that consistency has got worse. The interval assumes roughly normal data, and skew makes it less trustworthy; the confidence interval for standard deviation calculator does the exact calculation for your own n.

Step 8: check for skew and outliers

Squaring the deviations gives far-out values a lot of weight. Compare the SD with measures that ignore the extremes:

The mean sits above the median and the SD is well above what the middle half of the data implies: the data has a long right tail, as delivery times usually do. The 61 is above the upper IQR fence of 43.25 + 1.5 × 7.5 = 54.5, and removing that one order drops the SD from 7.61 to 5.96. Do not delete it to get a nicer number. Report that one late order accounts for a large share of the spread, and give the median and IQR alongside the mean and SD. (Quartile methods differ between tools; the values here use the inclusive method, and others give an IQR of 8.5.)

Step 9: write the interpretation

A good write-up states the numbers with units and n, gives them a meaning in plain terms, makes the comparison and flags anything that distorts the figure. For the deliveries:

Delivery times last month averaged 40.1 minutes (SD 7.6, n = 20; median 38.5, IQR 7.5). Most orders arrived within about 8 minutes of the average, and 2 of 20 missed the 50-minute promise. One 61-minute order accounts for much of the spread; without it the SD is 6.0. The SD is higher than last month's 5.9, but with 20 orders the difference is within sampling noise (95% CI for σ: 5.8–11.1).

Avoid "the data has a high standard deviation" with no reference point, avoid reporting the standard error as if it were the SD, and avoid more decimal places than the data supports. For the difference between a high and a low SD on the same mean, the two-dataset comparison shows it side by side.