EViews.com

This table can be produced when running the Automatic ARIMA table and allows you to specify what you would like as the criteria, e.g. AIC, SIC etc.

My question then is, why do the outputs in this table for AIC, SIC differ compared to if you estimate the equation separately?

How are the values in the Model selection criteria table being computed? The differnce is not huge, but there is a difference, I thought they would be identical, especially since I later made sure the samples were the same as well as the optimisaton when estimating.

They are not.

They are not.

Hmm, see screenshots attached please.

That's doesn't show anything. You'd need to provide the workfile.

That's doesn't show anything. You'd need to provide the workfile.

I need to do the model selection for the series named "rt_asx200"

Ah, in the (0,0)(0,0) case there is a difference. When we do the mle model selection we view the constant-only model has having two parameters (the intercept and sigma-squared). When estimating as an equation the constant only model is estimated via least squares and only has one parameter (the intercept). So the information criteria are slightly different.

Ah, in the (0,0)(0,0) case there is a difference. When we do the mle model selection we view the constant-only model has having two parameters (the intercept and sigma-squared). When estimating as an equation the constant only model is estimated via least squares and only has one parameter (the intercept). So the information criteria are slightly different.

I see. That makes sense for the (0,0)(0,0) case. However the difference still exists when comparing a (1,1)(0,0) via estimation and via the ARMA criteria table.

For a (1,1)(0,0) if you do the estimation separately and via quick>estimate equation>least squares. and then select from the options menu BFGS as the optimisation method and ML under ARMA then you will get the intercept, sigma-squared and the AR and MA terms.

The information criteria generated by this command will be the same as the Equation output in the Automatic ARIMA forecasting (if you tick the "Equation output" in the Options tab).

However in the same Automatic ARIMA forecasting, if you had also ticked "ARMA criteria table" under options (together with "Equation output" table), you will see the Information criteria in the ARMA criteria table are different from the Information criteria in the Equation output table, this is despite the fact that the (1,1)(0,0) case should now have been estimated the same using MLE, same optimisation method and considers the Intercept and sigma-squared as parameters.

Ah, turns out there is a slight bug in the IC calculation in the auto routine. When calculating the ICs it is currently taking the total number of observations in the current sample, without subtracting any NAs. Since you have a single NA,this calculation is off by one. It doesn't affect the results of course, since it is the same for all models, but does mean that the numbers don't quite match up. We'll fix it.

Thanks for pointing this out - sorry it took some effort on your behalf to drive it into my skull.

Ah, turns out there is a slight bug in the IC calculation in the auto routine. When calculating the ICs it is currently taking the total number of observations in the current sample, without subtracting any NAs. Since you have a single NA,this calculation is off by one. It doesn't affect the results of course, since it is the same for all models, but does mean that the numbers don't quite match up. We'll fix it.

Thanks for pointing this out - sorry it took some effort on your behalf to drive it into my skull.

No problem. Happy to help.