topic Re: variable reduction in Time series in Statistical Procedures

variable selection

abcd123 — Thu, 17 Aug 2017 16:45:22 GMT

Hello, all

I am fitting a time series model(ARIMA) and I have about 100 variables to choose from. I have tried PROC VARCLUS to reduce this number to 40, but still it is too much. Is there a way I can further reduce this number?

I know PROC GLMSELECT could be one option; however, from what I know, PROC GLMSELECT only works for general linear model. It it suitable to use PROC GLMSELECT for time series model selection?

Thank you very much.

variable reduction in Time series

abcd123 — Wed, 16 Aug 2017 16:08:25 GMT

Hello, all

I know PROC GLMSELECT could be one option; however, from what I know, PROC GLMSELECT only works for general linear model. It it suitable to use PROC GLMSELECT for time series model selection?

Thank you very much.

Re: variable reduction in Time series

YueLi — Wed, 16 Aug 2017 20:35:18 GMT

You can still use GLMSELECT for variable selection in time series. However, since it is time series data, you will need to include some time series related variables in the model to better capture the time series aspects of the data, e.g. seasonality, lags of dependent variables, differencing, seasonal differencing, etc.

Re: variable reduction in Time series

abcd123 — Thu, 17 Aug 2017 18:11:27 GMT

Thanks. Are you suggesting adding those additional variables and treat the problem as a general linear model? (BTW, what does "seasonality" mean? Is it a dummy variable?)

Re: variable reduction in Time series

alexchien — Mon, 21 Aug 2017 13:13:42 GMT

Yes the suggestion was to formulate a GLM model with time series features such as lagged dependent variable and seasonal dummies (you can use the CLASS statement in GLM to create the dummies for you automatically). The model is not exactly the same as ARIMA but you can use this proxy model to select input variables, and then use the selected input variables in the subsequent time series models.

thanks