Factor Model: How to do it?

[jasmin]

I would like to do a factor model in Eviews. Could someone advise me on how to do it in Eviews? (I use Eviews 7.1). Some people said that in Eviews it is not possible, while other said that it could be possible, so I would like to know how.

I have GDP_GROWTH variable, and I have five indicators (IND1, IND2, ... , IND5), which all affect GDP_GROWTH. I have all the data in standard quarterly time series in a workfile. I would like to turn it into a simple factor model based on these 5 indicators.

Please advise on concrete steps how to do it. Thank you.

[EViews Glenn]

The term "factor model" can be interpreted in a number of ways (though the fact that you have only a single dependent series confuses me). Can you be more specific about what you are trying to estimate?

[jasmin]

Thanks. I already found the option to create the factors out of the group.

First, my variables are quite heterogeneous, some like exchange rate range from 0.5 to 1.7, some are even negative, some variables have higher values from 1000 to 5000 and so on. What is still not clear to me is if I need to normalize the variables in the group before creating the factors, or if all that hard work is done by eveiws (and I just get the factors representing the variability explained by the variables in the group)?

Second, once I already got factor estimation, I understand I need to push “score” to get the factors values for each year, is that right? Still, this will not create the time series of the factors, and I need to click “name” the group of factor which creates time series of the factors in the workfile (also it creates the group which I delete), is that still right? After, I can use the new factor time series in the model specification to explain the single dependent series with these new factors; does it make sense what I am doing here?

[EViews Glenn]

I'm not quite sure what you are asking.

The Proc/Make Factors... from a group will create factor scores with properties as determined by the options you provide. By default, I believe that we compute the correlation matrix for the data, compute the decomposition of this matrix, and produce scores where the variances match the eigenvalues of the decomposition. You can choose various other options for the initial computation, and for the properties of the scores. You can take a look at the discussion in the manual, but the easiest way to double check the properties of your scores is simply to compute the covariance of the output series.

I hope this answers your question.

[jasmin]

Thanks, all clear.