Proc Similarity to cluster time series data

[neilxu]

If, for example, I have monthly sales number for different department (total 10 departments). Now I want to look at the trend of these sales based on departments. However 10 is too much so naturally clustering/grouping comes to my mind.

I saw the example of proc similarity to cluster time series and followed it to create the clusters. Now my question is I can use proc corr to group these departments, right? what is the benefit to do proc similarity?

Thanks

[ets_kps · Accepted solution]

Hello,

I assume you are following this example SAS/ETS User's Guide Example Programs.

As you can see, after SIMILARITY gives you the similarity matrix, then you can cluster in the same way you would use cross sectional clustering routines.  Use PROC CORR, CLUSTER, whatever you wish.

Similarity has a number of utilities but all are related to temporal ordering.  Typical methods of clustering ignore the ordering.  In the time series version of this clustering we are looking for variables(series) that we can treat as a group, rather than observations that we treat as a group.  The SIMILARITY procedure effectively transposes this information (with some other tweaks) so the clustering can be done on the variables.  If you were to use clustering directly (a perfectly sensible practice for some uses) then you would effectively be looking for intervals that behave similarly.  This might be perfectly reasonable for some sort of time series segmentation but that is not what we are showing in this example. 

Hope this helps-Ken

[neilxu]

Thank you so much.