search

InseadDataAnalytics / INSEADAnalytics

Other
122 stars 1.31k forks source link

ARIMA vs ETS vs TBATS #94

Open samjameswaller opened 6 years ago

samjameswaller commented 6 years ago

At a high level, in which scenario is it best to apply each of the above models? I have been reading some research that seems a bit conflicted on what to use when...

My understanding is that there is a trend and/or seasonality, go for exponential smoothing methods (ETS) because the ETS family explicitly model these components....if there's autocorrelation in the data go for the ARIMA methodology.

If anyone would like to add to this it would be much appreciated!

Antoine-Engerand commented 6 years ago

Hi Sam. Good question, and although there may be papers discussing about it, I imagine that all of them can pretty much provide a decent model of your training data. The question is more about the prediction: does it make sense to have an exponential number of users? to have a cap? It depends very much on the application and your personal judgment, which may be why it's not very easy to find about this issue.

I've found an short article comparing ARIMA and ETS: https://www.otexts.org/fpp/8/10

Anyone found a better source?