[ENH] tsfeatures in sktime

[ltsaprounis]

tsfeatures is a popular R package that gives a lot of important time series features and really helps with EDA and modelling.
There are various implementations of this in R and python:

• original R tsfeatures
• python tsfeatures
• featsts - R tidyverts alternative

Although the current python package does the job well, having this functionality within sktime would be a very nice feature and would really help with the sktime forecasting workflow.

[mloning]

Thanks @ltsaprounis for the feature request. Note that we you can already use tsfresh (and possibly catch22 too, at least it's in development). An adapter to Python's tsfeatures would be nice if they add functionality beyond the existing one. We would welcome a PR!

[fkiraly]

this should be a series-to-primitives transformer, right?

[mloning]

At the lowest level, in theory, yes. But may be easier to write this as a panel-to-tabular transformer.

[fkiraly]

Guess whichever suits you best, @ltsaprounis? Looking forward to this.

[ltsaprounis]

Thanks both! I had a look at catch22 and tsfresh after @mloning's comment. I think tsfeatures has a slightly different functionality as it pulls features that are also useful for EDA e.g. spectral entropy or more classical features like the Guerrero lambda.

Panel-to-tabular sounds good! Input panel data and get a pandas DataFrame as the output, where each row is a time series and the columns are the tsfeatures features.

Would a wrapper on the existing tsfeatures python implementation work well here? If yes I can start on the PR!

[fkiraly]

update from #968 - no one is working on this, so it's free again

[AzulGarza]

Hi @fkiraly! Is anyone working on this issue? I could try to include it after including auto_arima :)