Adding BSTS (Bayesian structural time series models) model via tensorflow_probability

Is your feature request related to a problem? Please describe. I and @koralturkk would like to add a wrapper for forecasting with BSTS Examples are given here.

Describe the solution you’d like

Idea 1) Wrapping each possible module via an argument of type list containing one or several dict. For example:

tfp.sts.DynamicLinearRegression(
    design_matrix, drift_scale_prior=None, initial_weights_prior=None,
    observed_time_series=None, name=None
)

…would become

add_DynamicLinearRegression = [{
"design_matrix" = design_matrix,
"drift_scale_prior"=None, "initial_weights_prior"=None,
    "observed_time_series"=None, "name"=None}]

So this is also the way I was wrapping in fbprophet with add_seasonality() e.g.

Idea 2) Allowing that the user creates the modules outside sktime so that it can be passed as a tfp.sts class:

import tensorflow_probability as tfp
dlr = tfp.sts.DynamicLinearRegression(
    design_matrix, drift_scale_prior=None, initial_weights_prior=None,
    observed_time_series=None, name=None
)

We would need **kwargs for it to pass dlr in sktime.BSTS(dlr)

Looking forward for your feedback @mloning

[fkiraly]

Hm, regarding design: I think aiming for both wrapped individual components and wrapped composites makes sense here! If I had to choose, I’d say to aim for wrapped composites, since that is more flexible and robust with regards to extensions and changes of BSTS.

The “really nice thing”, design-wise, would be adding a compiler-like layer that recognizes when adjacent sktime components are from BSTS and replaces them by a wrapped (and more efficient) BSTS composite. But perhaps out of scope due to complexity…

Regarding dependencies: as @mloning says, one of these days, we need to fix how we manage the neural network package dependencies. Currently, we have no maintainer dedicated to that family of dependencies - it is quite some work to keep these spinning, so help would be very much welcome. The problem is greater than just BSTS, it is essentially “anything that depends on tf”, so it would be maintaining “time series models using deep learning packages”.

[mloning]

Ok, we can give it a try to integrate it into sktime!

But we may have to make a difficult decision to not include it in the end, if we need to touch too much of the CI/CD pipeline, and move it to a separate repo. I worry about the extra maintenance burden for the CI/CD pipelines when upgrading versions later on. We have had similar issues with other packages already (e.g. the ongoing work to integrate the signature-based module).