8

I'm reviewing Prophet's documentation on the add_regressors method and came across something called a prior.scale, which is explained as Float scale for the normal prior. If not provided, holidays.prior.scale will be used..

I'm looking for information on what this is/ how it effects predictions, and how to assess/tune the values that it is set to.

CorerMaximus
  • 653
  • 5
  • 15

1 Answers1

2

Usually if you see some type of scale parameter associated with a prior, it's talking essentially about the standard deviation or spread of the prior. In this case, when you're adding a regressor, I'm guessing the prior mean is that the coefficient is equal to 0. The scale parameter is essentially a check on how large of an effect you expect this regressor to have. If you think it should have a relatively large effect, increase it and vice versa. For what it's worth, here's the description for holidays.prior.scale:

Parameter modulating the strength of the holiday components model, unless overridden in the holidays input.

ClancyStats
  • 1,219
  • 6
  • 12
  • 1
    So `prior.scale` controls how much of an effect the regressor has on predictions, correct? If so, I've seen repos setting it from 0.10 to 10- I understand that this needs to be tweaked through trial and error- but what range of values can really be assigned to `prior.scale`, and is there any way to extract out what `prior.scale` was used to a given parameter in a model object? – CorerMaximus Oct 25 '19 at 20:02
  • @CorerMaximus did you ever get a response here? I'd like to know, too :) – Christoph Terwitte Dec 21 '21 at 21:16