Heteroskedasticity

A regression describes an average relationship between an outcome and explanatory variables, and the errors represent variation not explained by that relationship. Heteroskedasticity concerns the spread of those errors, not necessarily the direction or size of the average relationship.

A common visual sign is a widening or narrowing band of residuals, as when small customers show little variation around predicted spending while large customers show much more. The average sales relationship can still be useful in that case, but uncertainty differs across customer sizes.

The condition does not require a time series, and it can occur in cross-sectional data such as households, suppliers or stores measured at one time. The term is also spelled heteroscedasticity, its opposite is homoskedasticity where error variance is constant under the specified model, and spelling differences do not create separate statistical concepts.

With appropriate exogeneity assumptions, ordinary least squares coefficients can remain unbiased despite unequal error variance, but that does not make every reported standard error reliable. Conventional formulas assuming constant variance can give misleading confidence intervals and significance tests.

Managers should therefore distinguish an estimate from uncertainty around it, since a cost driver may retain the same estimated average effect while a corrected standard error changes the confidence in that effect. Heteroskedasticity-consistent standard errors address inference under suitable conditions, but they do not automatically correct a missing explanatory variable, a wrong functional form or reverse causation.

A label saying the standard errors are corrected should not replace a review of the whole model. Weighted least squares is another approach when the error-variance structure can be supported, giving observations with greater variance lower weight, and choosing arbitrary weights merely to improve a preferred result is not a valid solution.

Transforming an outcome can sometimes stabilise variation, for example a logarithm may be useful when errors grow proportionally with scale. It also changes interpretation, so a coefficient in a transformed model cannot always be read as a direct dollar effect.

NIST illustrates both transformations and weighted estimation for data with non-constant variation, showing that different methods can fit the observed data similarly, so scientific knowledge and additional evidence can be needed to choose between them. Financial return models sometimes describe conditional variance changing with previous information, and ARCH and GARCH are specific approaches to that problem.

General heteroskedasticity is broader and should not be treated as another name for those models. For a non-finance manager, ask whether forecasts have realistic uncertainty at different scales, since a single prediction interval may understate risk for large accounts and overstate it for small ones, and check residual patterns, assumptions and the method used for uncertainty before relying on the result.