I am using Prophet on a timeseries with exogenous features.

I want to understand the significance (p value) of each of those features in my Prophet model, so that I can perform backward elimination of features, iterate and arrive at a robust model.

I understand that I cannot just do a t-test on the scaled coefficients (betas),to test if the coef is significantly different than 0. The reason is the normalisations taking place under the hood as explained here:https://github.com/facebook/prophet/issues/1264

In the above link also a solution is proposed:

From prediction df, whenever the interval [regressor_lower, regressor_upper] does not contain 0, that indicates a significant effect of that feature.

If i understand correctly that is the case when mcmc sampling is used.

I am not using mcmc sampling, but do have [regressor_lower, regressor_upper] in my prediction df.

Do I undertand correctly then that if the interval [regressor_lower, regressor_upper] of each row contains 0, then it is significantly affecting the predicted signal on that timepoint? And in that case how to generalise it to all timepoints, in order to be able to conclude if that feature is useful or not?)

I also found this on it, but unsure how to actualy get some significance/do the statistics from it: FBProphet: Understanding Regressor Impact on Multivariate Forecast

any help would be much appreciated!