CODEWITHSUNDEEP
pythonstatisticstime-seriesarima
Иван
Иван

Reputation: 97

Best information criterion for ARIMA model?

I use ARIMA model with python:

from statsmodels.tsa.statespace.sarimax import SARIMAX

model = SARIMAX(x, order=(p, d, q),
                            enforce_stationarity=False,
                            enforce_invertibility=False).fit(disp=False)

I want to compare some models with different paramenters with each other and choose a model with fewer regressors (choose more easier model).

Which Information Criteria should I use. I read about AIC and BIC. And I read, that BIC better than AIC to choose more easier ARMA model, but which best with ARIMA?

Maybe I must use other information criteria like HQIC?

I would be grateful for useful links.

Upvotes: 0

Views: 1402

Answers (1)

Cettt
Cettt

Reputation: 11981

I don't think there is a general answer to your question. In R there is an auto.arima function which is written by Rob Hyndman: he uses AICc. You can read all about it in his online book (chapter 8.7).

Note that classical information criteria (AIC, BIC, etc) do not allow to compare ARIMA models with different parameter d or D (since the number of useable observations depends on d and D). Here is a list of things to keep in mind when working with information criteria.

Therefore ultimately, the final choice of the model can (in my experience) not be based on one simple figure. Rather the final choice should be supported by different diagnostic plots and information criteria.

Upvotes: 1

Related Questions