Sktime: Vector Autoregression

[azev77]

Maybe I missed this, will there be options for vector autoregression?

[mloning]

In principle, yes. That would be an extension of the forecasting framework, but we haven't started working on it.

Would you be interested in contributing? We'd be happy to draft an API design proposal and discuss/review implementations.

[mloning]

I've added it to the dev roadmap #228

[azev77]

Unfortunately I do my work in Julia & MLJ (in the past R) & am not too familiar w/ TS in python.
I became interested in sktime bc yesterday I saw a GSoC project that will port sktime into MLJ.

I hope sktime will include:

• [ ] the usual ensemble benchmarks
• [ ] FB prophet
• [ ] LSTM time series models

[mloning]

Thanks for the suggestions! For an overview of what we have currently implemented for forecasting, see this issue #220. For deep learning we have a separate companion package: https://github.com/sktime/sktime-dl. Do you have a list of recommended LSTM models?

[azev77]

The only LSTMs I tried is through Keras in R.
I see you already made progress on the components of the ensemble.
Does SKTime compute the weighted ensemble in the same way as Rob:
https://robjhyndman.com/hyndsight/benchmark-combination/

[mloning]

At the moment we compute the simple mean as used in the M4 study, but I've added other ensembling methods to our forecasting wishlist #220

[aa25desh]

Hello @azev77 we are hosting MLJTime project here. It will be great to have you for the feedback:)

[mloning]

I'm closing this now, discussion can move to the MLJTime project: https://github.com/alan-turing-institute/MLJTime.jl/issues/1 🙂