so even if i restrict the fitted period to the most recent period, I still get NaN and I would love to understand why 🙂

my fitted and hat df don't have nan values

ok, found the problem: it is a bit complex, but maybe good for you guys to understand. so we have the hierarchical in real production and run inference every week. now the forecast min date always jumps by one week of course. and I concatenate one more week of fitted values from my originally trained model. but this code had a bug it turns out and i didn't add the last week for the fitted values. interestingly the code didn't break, but just gave NaN on the proportional foreasts, which i guess makes sense as the fitted dates didn't match until the first forecast date. Some error message would be lovely here 😉 maybe I could even do the PR, let's see.

Indeed maybe a message on date alignment would be helpful, if you file a PR / draft I'll have a look!

@Olivier there is a PR here, let's see what you are saying: https://github.com/Nixtla/hierarchicalforecast/pull/335

Maybe one more comment: the feature branch name does not really resemble the issue anymore. It started off with nan values in my work projects when there was the date misalignment between fitted and hat values. But when I was trying to replicate this with your examples the problem don’t appear. But what I could show is that nan values in the initial dfs created nan values after the hierarchical methods. But only for average proportion and proportion averages. So this is included into the PR, but the branch name is slightly different. I hope that is still fine

I left some comments in the PR