Expose statsmodels random_errors argument in ExponentialSmoothing.predict

[bernhardschaefer]

Is your feature request related to a current problem? Please describe.
For num_samples > 1, ExponentialSmoothing.predict() calls HoltWintersResults.simulate with the call self.model.simulate(n, repetitions=num_samples, random_state=rng).
The simulate function has an additional argument random_errors to specify how the random errors should be obtained., e.g. by using another distribution function that should be fitted to the residuals or by directly sampling from the random errors.
This argument is currently not exposed.

Describe proposed solution
Add random_errors argument to ExponentialSmoothing.predict() and pass it to self.model.simulate

Describe potential alternatives
Creating a subclass of ExponentialSmoothing with a custom predict which is more of a workaround.

Additional context

[dennisbader]

Hi @bernhardschaefer and thanks for raising this issue. We had a similar request a while ago to allow passing kwargs to the fit() method of the underlying base model. See PR #2460. Unfortunately we haven't merged that PR yet because of some concerns that I described here.

A similar concer would also apply to the predict parameter that you describe. I would prefer defining this in the model constructor rather than the predict() method itself to keep a unified API and avoid some downstream issues. I'm aware that this would mean that the model must be re-fit to try out different random_errors scenarios, but I believe it's better to go this way.

So I propose to add the random_errors (and maybe errors) to the model constructor. What do you think?

[bernhardschaefer]

Hi @dennisbader sorry for the late reply. Amazing to see that @jakubchlapek already created a PR that you just merged, really appreciate it. :-)