prophet holiday effect I use prophet to do time series forecasting.  During Christmas season, the data rise from 12/22, reach peak on 12/23 and 12/24 (positive effect), then drop a lot on 12/25 and 12/26 (negative effect).
I want the model to capture both the positive effect and negative effects around Christmas time, should I define the Christmas holiday as one holiday(12/25, lower_window=-3,upper_window=1)? Or define two 'holidays's, one before_Christmas (12/24, lower_window=-2,upper_window=0) and one Christmas (12/25, lower_window=0,upper_window=1)?
Thanks!
 A: AUTOBOX , which I helped to develop , uses a deep-learning algorithm/heuristic (a.k.a.as Hyperparameter Optimization Algorithtm) to identify as many as 6-10 individual holiday effects customized/tailored around EACH specific holiday to sort out and extract the "statistically significant structure" while conditioning for long-weekend effects , week-of-the-month-effects, day-of-the-month effects et al while being robust to anomalies.
I suggest that you implement this kind of approach or something similar that also is conditioned for sarima structure , day-of-the-week-structure and changes in these components over time while dealing with multiple level/step shifts and possible multiple break-points in trend. 
Attempting to find break-points in trend without accomodating level/step shifts is like washing one side of your face.
You asked "how to set up holiday effects" thus my answer is …. "Given  that you are limited to Prophet , I would heavy-load the possible pre and post holiday impacts."
