I am trying to find the best model for a log return time series. In SPSS the expert modeller is stating that an ARMA(0,0,0) is the best while in R the best fit is (5,1,0).
Why don't they agree on the same model?
Asked
Active
Viewed 399 times
1
-
1Why should they agree? There's not a single way to pick ARIMA orders. Everyone does it their own way. – Aksakal Mar 08 '18 at 16:14
-
1ARIMA(5,1,0) is a very strange model for log return. Perhaps you made an error when using R? – Rob Hyndman Mar 08 '18 at 20:18
-
@Aksakal I genuinely thought that both use information criteria. – Anna Mar 09 '18 at 06:27
-
@RobHyndman why is a strange model? I am looking at the log returns of Bitcoin. – Anna Mar 09 '18 at 06:27
-
If the log return is differenced it’s unusual. Log price is differenced – Aksakal Mar 09 '18 at 22:27
-
1Log returns would normally be mean-stationary. – Rob Hyndman Mar 11 '18 at 10:10
1 Answers
3
SPSS and auto.arima() probably use very different criteria for model selection. auto.arima() searches heuristically over the space of possible models, attempting to minimize an information criterion. I don't know how SPSS decides on a model (but see this), and I strongly suspect that the algorithm is different - it may minimize one-step ahead forecast errors instead of information criteria, or a different IC than auto.arima(), or it may even have the same target function, but iterate through the models differently and end up in a different local optimum.
Stephan Kolassa
- 123,354