Currently, xs.ensemble_stats tries to handle multiple functions in xclim.ensembles that do not necessarily have compatible input requirements. Some examples:
mean_std_max_min and ensemble_percentiles compute their statistics along realization. Thus, time should already have been aggregated (e.g. through a 30-year mean). --> Works well with horizons.
robustness_factors with ref=None also only works across realization, and considers the input to be pre-computed deltas where time has been aggregated. --> Works well with horizons.
robustness_factors with ref=xr.DataArrayneeds to have a time axis and will average time itself. I think this is because some tests need daily or yearly data. --> Completely breaks the concept of horizons.
Other functions in xclim.ensembles can also be called in theory, but diverge so much from what we currently have implemented that they probably don't work at all.
Steps To Reproduce
No response
Additional context
xs.ensemble_stats performs no checks and never warns the user if the list of tests is incompatible with the data provided.
At minimum, we should add that. Better yet, we should try to find a way to support that, or split xs.ensemble_stats in multiple functions.
Contribution
[X] I would be willing/able to open a Pull Request to address this bug.
Setup Information
Description
Currently,
xs.ensemble_statstries to handle multiple functions inxclim.ensemblesthat do not necessarily have compatible input requirements. Some examples:mean_std_max_minandensemble_percentilescompute their statistics alongrealization. Thus,timeshould already have been aggregated (e.g. through a 30-year mean). --> Works well with horizons.robustness_factorswithref=Nonealso only works acrossrealization, and considers the input to be pre-computed deltas wheretimehas been aggregated. --> Works well with horizons.robustness_factorswithref=xr.DataArrayneeds to have atimeaxis and will averagetimeitself. I think this is because some tests need daily or yearly data. --> Completely breaks the concept of horizons.xclim.ensemblescan also be called in theory, but diverge so much from what we currently have implemented that they probably don't work at all.Steps To Reproduce
No response
Additional context
xs.ensemble_statsperforms no checks and never warns the user if the list of tests is incompatible with the data provided.At minimum, we should add that. Better yet, we should try to find a way to support that, or split
xs.ensemble_statsin multiple functions.Contribution