AI Glossary / Statistics and probability / Foundation
Standard deviation
Definition
Standard deviation is a measure of spread obtained by taking the square root of variance. It summarizes how dispersed values are around their mean and is expressed in the same units as the original measurements. A larger standard deviation generally indicates more spread on the same measurement scale. It is not the average absolute distance from the mean, and it is different from the standard error, which describes the sampling variability of an estimate.
What it means
Two datasets can have the same average while differing greatly in consistency. Standard deviation adds information about that spread. Because variance uses squared deviations, values far from the mean can have a strong influence. A single standard deviation cannot describe every aspect of a distribution, such as skewness, multiple clusters or tail behavior. It is therefore useful alongside a display and other summaries, especially when unusually large or small values are important to the decision.
How it works
For a finite population with equally weighted values, subtract the population mean from every value, square the deviations, average them using the population count and take the square root. The usual sample standard deviation instead divides the squared-deviation sum by n minus one before taking the square root. Under independent identically distributed sampling with finite variance, that divisor makes the sample variance unbiased for population variance; it does not make its square root exactly unbiased. State the intended calculation and any weights rather than relying on an ambiguous software label.
Three practical examples
Example 1: A service team compares the consistency of completion times as well as their average. Example 2: An AI team summarizes variation in a metric across independently repeated evaluations, while identifying exactly what was repeated. Example 3: A production analyst describes the spread of component measurements. The same numerical standard deviation can have different practical meaning on different scales, and variation across individual items is not the same object as variation across batch averages.
Think of it this way
Think of a group of people standing around a meeting point representing the mean. A compact group has little spread; a widely scattered group has more. Standard deviation uses squared distances in its calculation, so a person very far away can matter disproportionately. The analogy helps with spread, but it should not be read as saying that standard deviation equals the ordinary average distance from the meeting point.
Worked example
Consider measurements 2, 4, 4 and 6 minutes. Their mean is four minutes. Deviations from the mean are -2, 0, 0 and 2, and their squared values sum to eight. If these four observations are the entire equally weighted population of interest, the variance is 8/4 = 2 square minutes and the standard deviation is the square root of two, approximately 1.414 minutes. If they are used for the usual sample calculation, the variance is 8/3 and the sample standard deviation is approximately 1.633 minutes. Adding ten minutes to every value changes the mean but leaves either spread unchanged. Multiplying every value by three multiplies the standard deviation by three and the variance by nine. Neither calculation establishes a normal distribution or a prediction interval for a future observation.
Illustrative case study
In this fictional case study, Willow Delivery compares two processes with similar average completion times. One has a larger standard deviation and a visibly long upper tail. Analysts report the mean, sample standard deviation and relevant percentiles, rather than summarizing performance with a single number. They check whether exceptionally long times reflect genuine operating conditions or recording errors before making changes. Managers use the full distributional picture to assess consistency and customer experience.
Common misconceptions and mistakes
Common mistakes include confusing standard deviation with standard error, assuming that approximately 68 percent of values always lie within one standard deviation, and ignoring whether the divisor is n or n minus one. The familiar 68-95-99.7 pattern is a property of a normal distribution, not a universal rule. Another error is comparing standard deviations across incompatible units. Extreme values can strongly affect the result, and some theoretical distributions do not have a finite variance or standard deviation at all.
Business use
Businesses use standard deviation to describe consistency, compare variability and support further statistical calculations. It should be attached to a clearly defined population or sample and a named measurement unit. If data are skewed or contain heterogeneous subgroups, examine additional summaries and distributions before acting. A lower standard deviation is not always better: it could reflect restricted coverage or the removal of difficult cases. For AI evaluations, distinguish variability across examples, random seeds, datasets and repeated studies, because each answers a different question.
People also ask
Can standard deviation be negative?
No.
When is it zero?
When all values are identical, for a defined finite-data calculation.
Does it require normal data to be computed?
No, although some interpretations and inferential methods assume normality.
Does adding a constant change it?
No.
Does a larger sample necessarily have a smaller standard deviation?
No; that common expectation concerns the standard error of a mean under appropriate sampling assumptions.
Check your understanding
What is the squared-deviation sum in the example?
Answer: Eight.
What is the usual sample standard deviation?
Answer: Approximately 1.633 minutes.
Why is the population calculation different?
Answer: It divides by four rather than three.
What happens if values are multiplied by three?
Answer: Standard deviation triples.
Does the calculation prove normality?
Answer: No.
Related terms
- standard deviation restores the original units
Learn next
Learn next why estimating a mean affects the sample-variance divisor and how standard deviation changes under rescaling. Compare datasets with the same mean and standard deviation but different shapes. Then distinguish descriptions of individual variability from uncertainty about an estimated average and from uncertainty in future predictions.
More in Statistics and probability
- Squared ranks test - The squared ranks test, often called the Conover squared ranks test, is a rank-based procedure for comparing dispersion across independent groups.
- Stable model - A stable model is a formally defined interpretation of a logic program with default negation.
- Staged tree (mathematics) - A staged tree is a probabilistic event tree whose nonterminal situations are grouped into stages when they share the same conditional distribution over...
- Standard error - The standard error is the standard deviation of an estimator's sampling distribution.
- Standard error of differences between means - The standard error of differences between means is the standard deviation of the sampling distribution of a difference between estimated means.
- Standard lottery - A standard lottery is a reference gamble used in expected-utility assessment, typically offering a defined best outcome with probability p and a defined...
Sources
Sources: NIST/SEMATECH Engineering Statistics Handbook, Measures of Scale, https://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm; OpenStax, Measures of the Spread of the Data, https://openstax.org/books/introductory-business-statistics-2e/pages/2-7-measures-of-the-spread-of-the-data. These primary instructional references describe the formulas and interpretation. The four-value calculation is original and independently checked; the delivery case is fictional.
Previous: Stance detection Next: Standard error