Usage Case Forecasts

[Steviey]

I'm wondering if it makes sense to use motif data of stumpy, to annotate a time series and use this as explanatory variables for forecasts. My guess is, there would be a problem with the index-length of a profile, because of the lag of items for the last window. Is there any code example out there (R or Python)?

[seanlaw]

@Steviey Thanks for the question and welcome to the STUMPY community. I haven't really come across anything significant related to forecasting with matrix profiles and STUMPY is probably not the best tool for this (specialized tools like Facebook Prophet might be more well suited). Having said that, time series chains is a very crude approach but your mileage may vary. Again, I think there are other better options and, in R, I believe that Rob Hyndman's forecast package is the gold standard.

[Steviey]

Hello Sean, Thank you for your welcome. I have all the mentioned R-forcasting packages already in use. What I want to do, is time series annotation with 'domain knowledge' of stumpy. The main forecasting package would be still a specialized one- like prophet, forecast etc. They then would use the annotated time series, to do there work. So the main question would be, how to annotate the time series with stumpy accordingly. Simply some more columns to the time series(es). This way- we also could use ML-algos for regressions.

[seanlaw]

I haven't thought about this a ton but there are two potentially related concepts that might be useful (just thinking out loud here):

• Using Weakly Labeled Time Series to Predict Outcomes
• Shapelet Discovery

I will have to think about this more.