Fourier seasonality transformations

[PabloRoque]

Currently we have YearlyFourier and MonthlyFourier as subclasses of FourierBase.

It would be nice to include WeeklyFourier (and perhaps DailyFourier) for datasets with finer time granularity.

Additionally, BaseMMM only makes use of yearly_seasonality. Maybe we could replace in favor of a seasonalities: List[FourierBase]

Also FourierBase.apply contains: periods = dayofyear / self.days_in_period. Perhaps we should make dayofyear a more general param

[wd60622]

Hearing few things here

• allowing passing fourier class(es) to MMM
• support weekly and daily fourier classes

2 is easier than 1 and probs should come first. I am all for. That why they are easy to subclass

[PabloRoque]

On allowing passing fourier class(es) to MMM.

We are currently passing yearly_seasonality : Optional[int] and then initializing as self.yearly_fourier = YearlyFourier(...).

We have 2 alternatives:

1. Easiest would be to add 2 new params: monthly_seasonality:Optional[int], weekly_seasonality:Optional[int] and initialize equivalently to YearlyFourier. Then handle these in build_model
2. Substitute yearly_seasonality param for a seasonalities:Optional[List[int]] | Optional[List[FourierBase]]. Then handle these in __init__ and build_model

Although 1 would add boilerplate code, I feel it might be easier for an end-user to understand how to use it.

What's your take?

[wd60622]

Good question. I think the easiest would be allow for a new parameter monthly_seasonality which would act like yearly_seasonality by taking an int or None.

I like the idea of taking a list, but what is the behavior when I pass more than one, say, YearlyFourier classes? Should that be allowed? Or [YearlyFourier(...), MonthlyFourier(...), MonthlyFourier(...)]

At the moment, we don't even have the ability to pass a custom Fourier class as the prior is defined in the model_config instead of with the class itself. So there is some complexity to add support there.

Think either can come first but mirroring the yearly_seasonality with monthly_seasonality parameter might be the easiest. We will have to be careful that the previous models are backward compat as the model_config is part of the comparison:

pymc-marketing/pymc_marketing/model_builder.py

Line 1008 in d3c7eda

hasher.update(str(self.model_config.values()).encode())