Heteroskedasticity is a condition in a statistical model in which the spread of the error terms is not the same for all observations. In a regression with constant error variance, called homoskedasticity, the scatter of points around the fitted line is similar everywhere. Under heteroskedasticity it fans out or narrows, for example when the size of forecast errors grows with income, firm size or the level of an asset price.
Ordinary least squares estimates stay unbiased, but the usual formulas for standard errors become wrong, so tests and confidence intervals can mislead. It is detected by plotting residuals or by formal tests such as the Breusch-Pagan and White tests. Common remedies are heteroskedasticity-robust standard errors, weighted least squares and transforming the variable, such as taking logarithms. In finance, variance that changes over time is the starting point of GARCH models, which treat it as something to model. The subject sits within econometrics and applies to linear regression generally.
Ref 1100001968