Standard Error (SE) Explained: Understanding Standard Deviation in Statistics

Search Fintrev

Search our guides, reviews and glossary.

Standard error (SE) is the standard deviation of a statistic’s sampling distribution. It measures how much an estimate calculated from a sample, such as a sample mean, would be expected to vary if you drew many samples of the same size from the same population. The most common version, the standard error of the mean, equals the standard deviation of the data divided by the square root of the sample size. A small standard error means the estimate is precise. A large one means another sample could easily produce a noticeably different answer.

What Standard Error Measures

Imagine taking a random sample of 40 stocks, computing their average return, then repeating the exercise with a fresh random sample of 40, and again, thousands of times. Each sample would give a slightly different average. Those averages form a distribution of their own, called the sampling distribution of the mean. Its standard deviation is the standard error.

In practice nobody draws thousands of samples. The value of the standard error is that it can be estimated from a single sample, using that sample’s variability and size. It tells you roughly how far your one estimate is likely to sit from the true population value, assuming the sample was collected properly.

The term applies to any sample statistic, not only the mean. Statisticians report standard errors for medians, proportions, differences between two means, correlation coefficients and regression coefficients. When people say “the standard error” without qualification, they usually mean the standard error of the mean, but a regression output, for example, lists a separate standard error for each estimated coefficient.

Standard Error Formula

The standard error of the mean is:

SE = σ / √n

where:

• σ = the population standard deviation
• n = the sample size

The population standard deviation is rarely known, so the usual working version substitutes the sample standard deviation:

SE = s / √n

where s is calculated as the square root of Σ(xᵢ - x̄)² / (n - 1). The divisor n - 1 rather than n corrects for the fact that deviations are measured from the sample mean, which is itself estimated from the same data.

How to Calculate Standard Error Step by Step

Take a hypothetical set of five observations: 2, 4, 4, 6 and 9.

1. Find the mean. (2 + 4 + 4 + 6 + 9) / 5 = 25 / 5 = 5.
2. Find each deviation from the mean. -3, -1, -1, 1, 4.
3. Square and sum the deviations. 9 + 1 + 1 + 1 + 16 = 28.
4. Divide by n - 1 and take the square root. 28 / 4 = 7, and √7 ≈ 2.646. This is the sample standard deviation.
5. Divide by the square root of n. √5 ≈ 2.236, so SE ≈ 2.646 / 2.236 ≈ 1.18.

The sample mean is 5 with a standard error of about 1.18, large relative to the mean because five observations can tell you little.

Units of Standard Error

Standard error is expressed in the same units as the data: percentage points for returns in percent, dollars for prices in dollars. This makes it directly comparable to the estimate: a mean of $50 with an SE of $2 is easy to read.

Relative Standard Error

To compare precision across estimates of different sizes, the standard error is sometimes divided by the estimate and shown as a percentage:

Relative SE = (SE / estimate) × 100

In the five-point example, 1.18 / 5 × 100 ≈ 23.6%. The relative standard error is undefined when the estimate is zero and becomes unstable when the estimate is close to zero.

Standard Error Formulas for Other Statistics

• Proportion: SE = √[p(1 - p) / n], where p is the sample proportion. A survey in which 40% of 400 respondents answer “yes” has SE = √(0.4 × 0.6 / 400) ≈ 0.0245, or about 2.4 percentage points.
• Difference between two independent means: SE = √(SE₁² + SE₂²). The standard errors of the two groups are squared, added and then square-rooted; they are not simply added together.
• Regression coefficients: each coefficient has its own standard error, produced by the regression calculation and reported by statistical software alongside the estimate.

Standard Error vs. Standard Deviation

The relationship between them is SE = SD / √n. A large sample of very volatile stocks can have a high standard deviation of returns and still produce a small standard error for the average return, because many observations pin the average down even when individual results vary widely.

Report the standard deviation to show how much individual cases differ, and the standard error to show how reliable an estimated average or coefficient is. A chart with error bars should state which one the bars represent, because SE bars are always narrower than SD bars for the same data.

How Sample Size Affects Standard Error

Standard error goes down as sample size goes up, but not in a straight line. Because n sits under a square root, the gains diminish:

• Doubling the sample divides the standard error by √2, cutting it by about 29%.
• Quadrupling the sample halves it.
• Cutting it to one-tenth requires a sample 100 times larger.

When the sample is a large share of a finite population of size N, the finite population correction √[(N - n) / (N - 1)] can be applied to reduce the standard error further. When the sample is a small fraction of the population, the correction is close to 1 and usually ignored.

Using Standard Error in Confidence Intervals

A confidence interval gives a range of plausible values for the population parameter. The general form is:

Estimate ± (critical value × SE)

For large samples, the critical value for a 95% confidence interval is about 1.96, taken from the normal distribution. For smaller samples where s replaces σ, the critical value comes from the t-distribution and depends on the degrees of freedom (n - 1 for a single mean). It is somewhat larger than 1.96, which widens the interval to reflect the extra uncertainty.

The width of the interval is driven by the standard error, so two studies estimating the same quantity can report very different intervals if their samples differ in size or variability.

The interpretation is about the method, not one interval: if the same procedure were repeated on many samples, about 95% of the resulting intervals would contain the true value. An estimate reported as “± 1 SE” covers only about 68% under the normal approximation, so it should not be described as a 95% range.

Using Standard Error in Hypothesis Testing

In z-tests and t-tests, the test statistic measures how many standard errors the sample result lies from the value assumed under the null hypothesis:

t = (sample estimate - hypothesized value) / SE

A large absolute value means the observed result would be unusual if the null hypothesis were true. As a rough guide in large samples, a result more than about two standard errors from the hypothesized value is statistically significant at the 5% level. The test statistic is converted into a p-value, the probability of seeing a result at least that extreme if the null hypothesis holds.

There is no standard error value that is “significant” on its own. Significance depends on the ratio of the estimate to its standard error. An SE of 0.5 is small next to an estimate of 3 (t = 6) and large next to an estimate of 0.4 (t = 0.8).

Example of Standard Error

In a hypothetical example, an analyst has 25 monthly returns for a strategy. The average monthly return is 1.0% and the sample standard deviation is 4.0%.

Standard error: 4.0% / √25 = 4.0% / 5 = 0.8%.

95% confidence interval: with 24 degrees of freedom, the t critical value is about 2.064. The margin is 2.064 × 0.8% ≈ 1.65%, so the interval runs from about -0.65% to 2.65% per month.

Test against zero: t = (1.0% - 0) / 0.8% = 1.25, well under two standard errors. The average return is not statistically distinguishable from zero, and the 1.0% could be a product of chance.

Now assume the same average and standard deviation over 100 months. The standard error becomes 4.0% / 10 = 0.4%, and t = 1.0 / 0.4 = 2.5. With four times the data, the same average return becomes statistically distinguishable from zero. Nothing about the strategy’s volatility changed; only the precision of the estimate did.

The same logic applies to cross-sectional studies. An analyst regressing subsequent returns on the P/E ratio for a random sample of S&P 500 companies gets a slope coefficient and a standard error for that slope. Whether the relationship is meaningful depends on how large the coefficient is relative to its standard error, and a larger sample typically tightens the estimate.

Interpreting High and Low Standard Error

A low standard error means the estimate would change little from sample to sample, so it is likely close to the population value, provided the sampling was sound.

A high standard error means the estimate is imprecise, usually because of a small sample, highly variable data or outliers that inflate the standard deviation. It does not mean the estimate is biased; the data simply cannot narrow it down much.

Standard Error in Finance and Investing

Financial analysis relies on estimates from historical data, and each of those estimates carries a standard error:

• Average returns. A historical average return is a sample mean, and its standard error shows how far the true expected return could be from what the record happened to deliver.
• Regression coefficients. A stock’s beta, a factor loading or the effect of a valuation ratio on returns is estimated by regression, and each coefficient’s standard error determines whether it is statistically different from zero.
• Backtests. A strategy that looks profitable over a short history may have an average return indistinguishable from zero, as in the example above.

A confidence interval around a historical average describes uncertainty about the average under the assumptions of the model. It does not predict where any single future return will fall, and it assumes the process generating returns stays the same.

Limitations of Standard Error

• It assumes a random, representative sample. A biased sample can produce a reassuringly small standard error around a wrong estimate. The formula measures random sampling error, not bias.
• It assumes independent observations. Financial time series often show autocorrelation and volatility clustering, so the simple s / √n formula tends to misstate the true uncertainty and adjusted methods are used instead.
• Small samples make it unreliable. With few observations, s itself is a noisy estimate of the population standard deviation. The t-distribution compensates, which is why t-tests are preferred over z-tests when data are limited.
• Distribution assumptions matter. Normal or t-based intervals rely on an approximately normal sampling distribution. The central limit theorem makes this reasonable for sample means in large samples, but highly skewed data, heavy tails or extreme outliers can make standard errors misleading, especially in smaller samples.
• Statistical precision is not practical importance. A huge sample can produce a tiny standard error and a significant result for an effect too small to matter.

Terms Often Confused With Standard Error

• Margin of error: the critical value multiplied by the standard error, i.e. half the width of a confidence interval. At 95% in a large sample it is about 1.96 SE.
• Standard error of the estimate (of a regression): the typical distance between observed values and the values a regression predicts. It describes the model’s fit, not the precision of one coefficient.
• Standard error of measurement: a psychometrics term for the expected spread of an individual’s scores around their true score on a test, based on the test’s reliability. It concerns repeated measurement of one subject, not sampling from a population.
• Typical error: a term used in some sports and measurement science for the within-subject standard deviation of repeated measurements. Despite the similar name, it is not the standard error of a mean.

FAQs

The Bottom Line

Analysts who rely on historical averages, betas or backtests need standard error to judge whether a result reflects more than chance. A 1.0% average monthly return can be statistically indistinguishable from zero over 25 months yet significant over 100. A small standard error only shows precision, so a biased sample, autocorrelated returns or a trivially small effect can still make an estimate misleading.