Can outcomes have different means and still be homoscedastic?

Yes; the example has means 20 and 50 with equal conditional variances.

Does a flat residual plot prove constant variance?

No. It is diagnostic evidence with limited resolution and sample size.

Are homoscedasticity and stationarity identical?

No. Stationarity involves additional distributional or moment conditions, depending on its definition.

AI Glossary / Time series and signal processing / Intermediate

Homoscedasticity

Definition

Homoscedasticity means that an error or outcome distribution has the same variance across the conditions being considered. In regression, it commonly means that the conditional error variance is constant across predictor values. In time-series work, it may refer to constant error variance over time. The conditioning information and model must be specified to make the claim precise.

What it means

A model can predict different average outcomes while retaining the same uncertainty around those averages. For example, larger orders may take longer to process, yet deviations from the predicted processing time might have a similar spread for small and large orders.
Homoscedasticity concerns that spread, not equality of the means or equality of individual errors. Its opposite, heteroscedasticity, describes variance that changes with the conditions under consideration.

How it works

For a regression equation Y = m(X) + error, a standard homoscedasticity condition is Var(error given X) = sigma squared, where the same finite value applies across X. This is separate from the condition that the conditional error mean is zero.
Analysts inspect residuals against fitted values, predictors, and time, looking for systematic changes in spread. Formal tests can supplement these diagnostics but depend on assumptions and have limited sensitivity in finite samples. In time series, distinguish unconditional variance from variance conditional on past information: a series can have constant unconditional variance while its conditional volatility changes over time.

Three practical examples

A delivery-time model predicts different means by distance while assuming comparable error variance across the relevant distance range.
A demand forecaster checks whether uncertainty grows when expected volume rises. A widening residual pattern suggests that constant absolute error variance may be inappropriate.
A manufacturing analyst compares measurement variability across operating settings, distinguishing changes in average output from changes in precision.

Think of it this way

Think of archery targets placed at different horizontal positions. The centers can move while the spread of arrows around each center remains the same. Homoscedasticity concerns that comparable spread. It does not say every arrow hits the center, that the centers coincide, or that successive shots are independent.

Worked example

Construct an original discrete error distribution that takes values −3, −1, 1, and 3 with equal probability. Its mean is zero. Its variance is [9 + 1 + 1 + 9]/4 = 5.
Suppose a model predicts mean outcome 20 in one condition and 50 in another, with this same error distribution in each. Possible outcomes are [17, 19, 21, 23] in the first condition and [47, 49, 51, 53] in the second. Both conditional outcome variances are 5 even though their means differ by 30. This is a stipulated homoscedastic model.
Now suppose errors in the second condition are instead −6, −2, 2, and 6, again equally likely. Their variance is [36 + 4 + 4 + 36]/4 = 20, four times the first condition's variance. Their standard deviation is twice as large. That revised model is heteroscedastic.
The divisor four is correct here because the four values define the entire assumed probability distribution. If four observed values were used to estimate an unknown variance with the usual sample formula, the divisor would instead be three, yielding 20/3 for the first set. A small sample with similar spreads would not prove that a real population is homoscedastic.

Illustrative case study

Illustrative fictional case: Aspen Dispatch notices that forecast errors for large daily workloads are more variable than errors for quiet days. The team first checks whether missing predictors explain systematic mean errors, then examines the remaining spread.
It compares a model with workload-dependent uncertainty against a constant-variance alternative and evaluates interval coverage across workload ranges. Managers receive wider ranges where the evidence supports greater uncertainty. This fictional case shows why one average error scale may be inadequate for operational decisions.

Common misconceptions and mistakes

Homoscedasticity does not imply normality, independence, zero bias, or a correct mean model. The worked distribution is discrete rather than normal. Errors can also have equal variances while being correlated across time.
Do not conclude that heteroscedasticity automatically biases ordinary least-squares coefficients. Under suitable zero-conditional-mean and other conditions, coefficients can remain unbiased, while conventional homoscedastic standard errors become inappropriate and efficiency may suffer. Conversely, robust standard errors do not repair an incorrect mean model or automatically supply accurate individual prediction intervals.

Business use

The assumption affects how businesses interpret uncertainty in regression, forecasting, and comparisons across operating conditions. Incorrectly treating variable noise as constant can make some intervals too narrow and others unnecessarily wide.
Assess the spread where decisions are made, not only across the full dataset. Depending on the problem, consider variance modeling, appropriate transformations, justified weighted least squares, or robust inference. For temporal data, serial dependence requires separate attention; heteroscedasticity-only adjustments do not automatically address autocorrelation.

Check your understanding

What variance does the first stipulated error distribution have?

Answer: Five.

If errors are multiplied by two, what happens to variance?

Answer: It is multiplied by four, while standard deviation doubles.

Why does the example distinguish a distribution from a sample?

Answer: Variance calculation and variance estimation use different interpretations and may use different divisors.

Would constant error variance rule out temporal correlation?

Answer: No. These are separate properties.

Related terms

Ordinary Least Squares
Weighted least squares
Autocorrelation
Stationarity
heteroscedasticity
conditional variance
residual diagnostics
robust standard errors
volatility

Learn next

plot residuals against fitted values and time for a simple model, distinguishing changes in mean from changes in spread. Then examine how variance modeling and robust coefficient inference answer different questions, especially when individual forecasts are required.

Sources

Sources: NIST, Homoscedasticity Plot: https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/homoplot.htm . NIST, detecting non-constant variation: https://itl.nist.gov/div898/handbook/pmd/section4/pmd442.htm . statsmodels, official statistical diagnostics and robust covariance overview: https://www.statsmodels.org/stable/stats.html . Primary indexed guidance consulted. The finite-distribution calculations and fictional case are original.

Previous: Homophones Next: Hopfield Network