AIC

[AsyaNech]

Hi everybody,

I have 2 questions concerning AIC 1/ for some reason, AIC function does not work for tbat model. Below are the code and console output tumblr_tbats <- tbats(tumblr_ts) tumblr_tbats AIC(tumblr_tbats)

2/ which difference in AIC is significant? What are typical ranges for AIC? AIC of different models we used vary from 1244 to 1315. Whether this range is significant? Or all these AIC are bad?

Thank you in advance!

[VarunKShetty]

Hi @AsyaNech, 1)Perhaps the AIC "function" is not defined to be applied to the class of bats. But you can still see the AIC value of your model in the last line of the output -- below sigma, when you call tumblr_tbats. 2) This is a good and deep question with no "correct" answer. See this link to get started on appreciating the thought process around it: https://en.wikipedia.org/wiki/Akaike_information_criterion#How_to_apply_AIC_in_practice Cross Validated -- a subsidiary of stack exchange -- is another good place to look for answers to such questions: https://stats.stackexchange.com/

-VK

[marionroger]

Hi @AsyaNech From my understanding we can compare AIC of models which are from the same type, eg compare AIC from the different ETS models but we cannot really compare directly AIC between ETS and TBATS model. @VarunKShetty would you mind validating this statement? :) don't want to induce mistakes!!

[VarunKShetty]

Anton's comments on this matter: "Generally Yes" is the correct answer. Very different models can be compared as long as they are fit using the same data. For instance, ETS and TBATS would be OK for compare, but ARIMA would only be OK if it is (p,0,q), i.e., without differencing.

Please see here: https://robjhyndman.com/hyndsight/aic/ (item #4 speaks directly to this question)

[marionroger]

Thanks @VarunKShetty!

[AsyaNech]

Thanks @VarunKShetty, @marionroger!