-
Notifications
You must be signed in to change notification settings - Fork 315
Closed
Labels
TVPsRelated to time varying parametersRelated to time varying parametersenhancementNew feature or requestNew feature or requestrequest discussion
Milestone
Description
Following #1246, these components should be integrated into the MMM
class.
This might also support the previous boolean flags for time_varying_intercept
and time_varying_media
but also include passing of either of the classes, HSGP
or HSGPPeriodic
from pymc_marketing.prior import Prior
from pymc_marketing.mmm import MMM, HSGPPeriodic, GeometricAdstock, LogisticSaturation
media_hsgp = HSGPPeriodic(
scale=Prior("HalfNormal", sigma=1),
ls=30, # days
period=365, # days
...,
dims="date",
)
mmm = MMM(
date_column=...,
channel_columns=...,
adstock=GeometricAdstock(l_max=10),
saturation=LogisticSaturation(),
time_varying_intercept=True, # Previous behavior or updated with HSGP behind the scenes
time_varying_media=media_hsgp,
...,
)
The complicated aspect is that the current implementation of GPs will perform a softplus
transformation on that variable. Would that be explicit now? Or handled for all custom classes within the MMM
class?
Thoughts on this? @juanitorduz @cetagostini
Metadata
Metadata
Assignees
Labels
TVPsRelated to time varying parametersRelated to time varying parametersenhancementNew feature or requestNew feature or requestrequest discussion