a key time as a split variable in Microsoft Time Series?

Hello,

I was working with Microsoft Time Series model (MTS) with some data, when in the mining model viewer, decision tree tab, I realized that the key time variable that I define, it was acting like a split variable.

So, I ask you, this is possible?, because, for me, this should not happen….

After, I review the Data Mining Tutorial by Seth Paul, Jamie MacLennan, Zhaohui Tang and Scott Oveson, and I found, in the Forecasting part, that the key time variable (Time Index) it was acting like a split variable too, in for example, M200 pacific:Quantity and R250 Europe:Quantity.

So people, it’s possible that a key time variable act like a split variable in a MTS model?

Thanks

This is by design. Imagine you have a series that behaves in a certain manner, then at some time, the behavior changes radically - possibly due to some external event not captured in your data. Since the Time Series prediction always occurs in the future, splitting on time allows us to isolate and describe that past behavior without averaging out that behavior into your current predictions. If your series was mostly increasing or decreasing, this would manifest itself naturally in a split like (t-2)>70, however, if your series is mostly stationary, such a split wouldn't make sense.

The implementation is that we add a monotonically increasing series that is only used as a split candidate and not as a regression candidate.