Éric Dupuis

I re-ran the same big data tests (the train-adjust, I should have done only the train part, but forgot to remove the adjust): |method| compute time [s]| |-|-| |improved-nbutils (as...

Yes, now it seems we can see your improvements with higher level functions, nice! The bottleneck I was mentioning (changing stack => flatten) doesn't seem so huge now if we...

Closing for now, will open an issue. `fastnanquantile` will be used optionally in #1513

Just to confirm: `loc` is indeed used to generalize the support of the distribution from [0,∞] to [`loc`, ∞] https://stackoverflow.com/questions/68948202/negative-observations-for-scipy-stats-gamma-fit The [paper](https://journals.ametsoc.org/view/journals/clim/23/7/2009jcli2909.1.xml?tab_body=pdf) introducing SPEI uses the three-parameter distribution of log-logistic...

I've cleaned the function a lot, it should be good for a read, hope this is digestible...

I have an implementation with `map_groups` in the branch `mbcn` which is much cleaner IMO. With 400 spatial points, 30 years data, a `group = sdba.Grouper("time.dayofyear", window=31)` was 'only' 2x...

Ah my bad, I removed `.reset_index(dims_or_levels=[win_dim])` so that incomplete or Gregorian calendar could also be used as input. Keeping this quantity is not crucial, it can be easily recomputed in...

Nice! Merci pour la review qui a dû demander du jus aussi! J'essaie de faire plus de PRs moins grosses...

Hi! It seems like a scipy error. I'll wait for details. It could be a bug on our side, but I suspect it could also be the data that is...

Hi @PraewKPT, thanks for sharing more info. * Could you also show the python code you used to produce this? A bigger error message for context would be good too....