[feature request] weekly.seasonality.prior.scale and yearly.seasonality.prior.scale about prophet

[wjhrdy]

If there is a reason that we would want to expand that list of prior scales in the future past weekly seasonality, yearly seasonality, changepoint and holiday it might be good to make it a list.

One scenario I can think of is if you have a large list of holiday components and you want to weight them differently you may want to list those weights in a more general prior_scales list.

One drawback of using this prior_scales list is that it's a less standard input and if you are tailoring this to be easy to use and automatic this added complexity of arguments may be prohibitively complex.

[seanjtaylor]

I wonder if it might make sense to make more generic, like prior_scales = list().

[peacej]

Definitely agree, would like to only decrease yearly.seasonality.prior.scale but not the weekly one.

[bletham]

Would anyone like to make a PR for this in the v0.2 branch? It would need to be on top of the add_regressor feature (8f1607c), which makes it more convenient to have prior scales with different values. This has been done for Python but is not yet in the R version (see #101).

• For holidays, I think we could add an optional column to the dataframe prior_scale (optional like lower_window and upper_window) which specifies the prior_scale for each holiday if given, and otherwise it defaults to holidays_prior_scale. There should be validation that a single prior scale is specified for each holiday. The make_holiday_features method could return the prior scales in a list, which could then immediately be used by make_all_seasonality_features as-is.

• Custom seasonalities (via add_seasonality) should be able to specify the prior scale, but then default to seasonality_prior_scale if not provided.

• There are three built-in seasonalities: yearly, weekly, and daily. Of the options discussed above, the cleanest seems to me to be having three corresponding arguments to Prophet(), yearly_seasonality_prior_scale, yearly_seasonality_prior_scale, and yearly_seasonality_prior_scale. I do worry about cluttering it with too many arguments, but also wouldn't want to sacrifice clarity and ease-of-use over two arguments. Any other thoughts on this?

[bletham]

This is done in Py with a620a6c and 8d27643 and can be ported to R once #278 is finished. I'll add documentation for it then.

[bletham]

This can now be done in v0.2 available in CRAN and pypi. Documentation is here:
https://facebookincubator.github.io/prophet/docs/seasonality_and_holiday_effects.html

If you have any issues with it feel free to reopen here, or to make a new issue.