I may be misinterpreting, but the documentation for
arviz.compare seems to suggest that it is computing the uncertainties on the weights for each model:
By default the uncertainty in the weights estimation is considered using Bayesian bootstrap.
but I’m not sure this result is being placed in the output
DataFrame. Is it possible to access the uncertainties on the weights?
@aloctavodia will probably know better. I think this only applies to bma-like methods and was not updated when we changed the default to stacking, and even or those mehods, the bootstrapping is only used to stabilize the values and only the average is returned. It could be possible to modify the function to return the bootrstrapped raw values but I’d have to look at the code carefully before trying to guess that.
Just want to mention that @OriolAbril comment is correct. The statement “By default the uncertainty in the weights estimation is considered using Bayesian bootstrap” only applies to the BB-pseudo-BMA method of computing weights. ArviZ does not offer an easy way to get the bootstrapped weights, but this is the line were they are averaged before returning them arviz/stats.py at main · arviz-devs/arviz · GitHub