Decision Tree and Forecasting in VA

[andreas_zaras]

Hello,

I have a question conerning the decision tree and forecasting facilities in VA V6.2.

Let us take the decision tree first. As i read in the manual after the software creates the maximal tree it conducts a pruning process. As far as i know the pruning process needs data partitioning but the manual does not mention any options for data partitioning (training/validation/test). Is the software conducting data partitioning (and then pruning) in the background? if yes how does it select the percentages of training/validation/test sets? Is it fixed i.e. the same in any data set or it adjusts them depending on the size of the data set? Finally after the decsion tree model is created can you do scoring of new data based on it?

Concerning the forecasting cababilities in VA as far as i read it searches only in the exponential smoohting family to select a model. How is the best model selected? Does it use statistical tests such as unit root/ seasonal unit rootm, %logcheck facilities etc to diagnose the series and then fits the appropriate model or it does data partitioning fits the available models and selects the best? Or maybe something else that i am not aware of?f

Is there any documentation on the web abaout the above questions?

Thanks in advance,

Andreas

[DavidHenderson · Accepted solution]

There is no partitioning of training/validation/test data in the VA decision tree functionality.  All of the data are applied to build the tree.  There is also no tree scoring functionality in VA at this time.  VA Explorer allows you to explore the data-- not build models or score new data sets against models.

I can't provide details at this time, but there is a new VA product coming that provides the functionality you are asking about.  Look for more details around Global Forum time frame.  Also, SAS has other products that fill those needs such as SAS Enterprise Miner.

[andreas_zaras]

Thank you very much for your answer!

[DavidHenderson]

You are very welcome.

I also did some check for you on the forecasting functionality in VA Explorer, but I am afraid the answer isn't straight-forward.  It depends on what data variables are used and how many.  In the simple cases, one of the ESM's are used (as you surmised), but one of the ARIMAX family may also be used.

In VA Explorer, the model used is listed in the information pop-up for forecasting. Hit the “i” icon in the middle of the page to get the pop-up.

[udo_sas]

Hello Andreas -

Please excuse for delay - I was not aware of your post until now.

For forecasting the best performing model is picked based on the in-sample performance based on the root mean squared error fit statistic.

Thanks,

Udo

[andreas_zaras]

Thank you very much Udo!

[BSL]

Does anyone has the standard sample data for decision tree chart?

thnx

[AnnaBrown]

Hi BSL,

ted werner posted a sample data set for a decision tree, per your request: Sample Decision Tree Data.

FYI - I edited out your email address in your comment for security reasons.

Anna

[BSL]

thanks anna 🙂