What it means
A regression model estimates a mean relationship, and observed outcomes vary around it to produce residuals that estimate model errors. Homoskedasticity means the error spread stays broadly constant under the specified model.
The term is also spelled homoscedasticity, and its opposite, heteroskedasticity, allows the error variance to change across conditions. The condition does not mean the outcome itself has the same value everywhere.
Sales can rise with store size while the unexplained variation around the prediction remains similar, because the mean pattern and the error variance are different features. NIST illustrates a scatter plot with a linear relationship and similar variation around the predicted line across values of the explanatory variable, whereas a cloud of points widening sharply at one end suggests a different variance pattern.
The assumption can apply to cross-sectional or time-series models. In a cost study it might concern residuals across factories, while in a financial series it might concern variance across periods under the chosen model.
Residual plots are a useful first check, since plotting residuals against fitted values or explanatory variables can reveal changing spread, although a plot is evidence for review and not proof that every possible variance pattern has been excluded. Statistical tests can also be used, with their own assumptions and limitations.
A failure to reject unequal variance does not prove perfect constancy, and small samples can have limited power to detect a problem. A high R-squared does not establish homoskedasticity either, because a model can explain much of the outcome's mean variation while residual spread still changes with scale.
Constant variance does not establish independence or normality, since errors can have a common variance while being correlated over time or following a non-normal distribution. Model diagnostics should not collapse several distinct assumptions into one check.
When unequal variance is present, alternative inference or estimation methods, such as heteroskedasticity-consistent standard errors, transformations and supported weighting, may be appropriate, and the choice depends on the model and data rather than a desire to make one significance result stronger. A model can appear homoskedastic after an appropriate transformation, but the change also affects interpretation of coefficients and forecasts, so an analyst should explain the transformed scale and any conversion back to original units.
For a non-finance manager, ask whether uncertainty is realistic across different sizes or periods. A single prediction range can be useful only when its assumptions fit the data, and the goal is reliable interpretation and planning, not forcing every dataset to look like a textbook diagram.
In practice
Real-world examples.
Example
A cost model's residual spread stays similar across small and large shipments. That supports a constant-variance description for the analysed range, subject to further checks.
Example
A regression has a high R-squared but residuals widen with predicted sales. The good fit does not establish homoskedasticity.
Example
A time-series model has constant error variance but correlated errors. The analyst reviews correlation separately because constant variance does not imply independence.
Formula
Calculation
Homoskedasticity assumes Var(error | X) = sigma squared, a constant under the model. If residual standard deviation is approximately $20 across several groups, the corresponding variance is approximately $400 squared. A group with a $60 standard deviation instead has a variance of 3,600, suggesting unequal spread rather than a small difference in average outcome.
The practical effect shows up in prediction ranges. Using roughly two standard deviations, a $100 forecast with a $20 standard deviation has a range of about $60 to $140, calculated as $100 plus or minus $40. For the group with a $60 standard deviation, the same forecast would have a range of $100 plus or minus $120, or about -$20 to $220. A single range applied to both groups would be too wide for one and too narrow for the other.Case study
Seen in the real world.
Fictional case study: Birch Analytics built a store-expense regression and claimed its high fit score proved all assumptions were satisfied. The planning team used the same uncertainty range for every store. An analyst checked residual plots, variance patterns and time dependence separately. The spread was reasonably stable across store sizes, but residual correlation remained a separate concern.
Birch documented the supported constant-variance assumption without overstating it. The team improved its uncertainty calculations and stopped using one favourable diagnostic as proof of the whole model. In a follow-up review of its regional warehouse model, the same analyst found residuals widening with volume. Birch applied a transformation and reported its forecasts back in original dollars, with a plain-English note explaining the change so planners did not misread the ranges.
Watch out
Common mistakes.
- Equating constant variance with identical outcomes. It concerns error spread around the modelled relationship.
- Using R-squared as a variance test. Check residual behaviour separately.
- Assuming homoskedastic errors are automatically independent or normal. These are different conditions.
Questions
People also ask.
Does it mean there are no errors?
No. Errors can exist while their variance remains constant.
Is it the opposite of heteroskedasticity?
Yes. The distinction concerns constant versus varying error variance.
Can a model have constant variance and correlated errors?
Yes. Variance and dependence are separate properties.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%Related