Statistics Glossary → Box-Cox transformation
Box-Cox transformationconcept
A family of power transformations, governed by a parameter λ, that aims to make the variance of a positive variable more constant and its distribution more symmetric.
The transformation is y(λ) = (y^λ − 1) / λ for λ ≠ 0 and y(0) = ln y. λ = 1 means no transformation (only a shift), λ = 0.5 corresponds to the square root, λ = 0 to the logarithm and λ = −1 to the reciprocal. λ is chosen by maximum likelihood or, for time series, by Guerrero's method, which looks for the λ that keeps the ratio of the standard deviation to the mean as constant as possible across subperiods of the series.
The transformation is defined only for positive values; with zeros or negative values, a constant must be added or an extension such as Yeo-Johnson used. In time series it turns seasonality whose amplitude grows with the level into constant-amplitude seasonality, making additive models suitable. A forecast made on the transformed scale and back-transformed gives approximately the median, not the mean; if the mean is needed, a bias adjustment should be applied. Results are harder to interpret on the transformed scale.
An everyday example
Monthly visitor numbers at a shopping centre rise over the years, and the December peaks grow too: in the first year the peak is 20 000 above the average, in the fifth year 60 000 above. Guerrero's method chooses λ ≈ 0.1, a transformation close to the logarithm, and on the transformed series the amplitude of the seasonal swings stays almost constant across the years.
In MerQur
'Transform / λ' in STL Seasonal Decomposition (no transform, log, Box-Cox Guerrero, Box-Cox manual λ); the 'Transformation (None / log / Square root (λ = 0.5) / Box-Cox) and Box-Cox λ' options in ARIMA Forecast.
Analyses where this concept appears
STL Seasonal DecompositionARIMA Forecast