Coming from basically no time series back ground, this is likely a simple question, but what is the relationship between "being able to" use an additive decomposition of a series into seasonal, trend and remainder and a Box Cox transformation?

From Professor Hyndman's blog:

Because not all data could be decom­posed addi­tively, we first needed to apply an auto­mated Box-​​Cox trans­for­ma­tion.

I was wondering:

1) What makes an additive decomposition attractive relative to a multiplicative one (which I understand is basically the other choice).

2) What is the requirement for an additive decomp and what does Box Cox do to make this possible? I think of Box Cox for ANOVA and reducing heteroskedasticity. Is there a tie in with decomposition of a series?

  • $\begingroup$ Maybe you would like to add key-words relating to the forecast package (like "forecast", "Box-Cox" whatever is available). Prof. Hyndman is very active on CrossValidated. Or you comment in the blog, which he might read. To me, Box-Cox is not only to make data more homoskedastic but also to make it more Gaussian - which makes it easier to model. In the blog he uses stl which is meant for additive decomposition. In case you have a multiplicative model, the log transform makes it additive - and on generalization of log is BoxCox. $\endgroup$ – Ric Jan 8 '13 at 13:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.