11-07-2014 06:58 AM
I'm having a problem with Forecast Studio v13.2 when I select only to use my custom model list, but also want the selection process to consider inputs in the input data table.
I've got my custom model list for Forecast Studio with no inputs (arima family), but inputs in the data source that I also want to select. Their roles are specified as "Independent" and even "Force to use", but I still don't get my final models with any input. I'm sure that FS is not considering the inputs.
This doesn't happen when I leave the default FS list... Can anybody please tell me how to make FS use the inputs AND ONLY my custom model list??
Thank you in advance...
11-07-2014 03:44 PM
If you want to use independent series for your custom repository models, you will have to add these independent series to your models yourself.
Automatic input selection is only available if you use the diagnostic capabilities of SAS Forecast Studio - as it is using HPFDIAGNOSE under its hood. This Forecast Server procedure allow for automatic selection of independent series. More details can be found in documentation.
11-11-2014 06:25 AM
I'm telling you my situation; if it is not possible to use my customized list of models, then I'd need to use, from a large list of regressors, only the specified series regressors whith small p-values in each model, with no transfer function in any input. In other words, intoduce the significative inputs in the arima model like an event: Beta(t)*Input(t) to forecast y(t)
Can this be done in FS? Which options should be selected?
Thanks a lot in advance
11-11-2014 04:40 PM
Have you considered to use events instead? Not sure I'm following your situation, but maybe defining an event repository (including future values for your events) will be a better option. SAS Forecast Studio can perform events selection similar to independent series selection, so you could define a large range of potential events and let each model decide if including them will improve accuracy.
11-14-2014 05:55 AM
First thank you very much for your time and effort. I appreciate.
... I've come to the conclusion that inputs and events are essentially the same... in other words: inputs + use if significant (or try to use) in the variable settings panel, is essentially the same as defining an event repository + use if significant (or try to use)... and this works only if diagnosing in the arimax family (or related), but never with a custom list of models. The only advantage I see is that (I haven't tried this) by defining a regressor as a pulse, may be I'd avoid using the transfer function...
Do you think I'm right?
As you said, if I want to use all the power of the high performance forecasting system, I'll have to program in the hpf language...
Thanks a lot udo,
11-14-2014 08:45 AM
Yes - the only thing I'd like to add is that you can define a user specified model repository yourself using the GUI. But of course you will need to decide for yourself which independent series or events to include.
If you have a lot of potential candidates for independent series, you may want to preprocess your data trying to identify significant series yourself. One way to go about this is to do a similarity analysis of input combined with a cluster analysis. Purpose would be to identify similar patterns in your input series beforehand. This thread: https://communities.sas.com/message/172179#172179 might be of interest.