https://github.com/nixtla logo
i

Isaac

09/06/2023, 1:50 PM
I think the new
forecast_fitted_values
function in MLForecast is broken. It's producing the transformed predictions rather than the inverse transformed one. I've written a Github issue.
j

José Morales

09/06/2023, 3:07 PM
Thanks. We'll work on a fix soon
Hey. Looking at this I think it may be best to not apply the inverse transformations at all because some of the transformations can't recover the original values of the target for the training set. For example if you use Differences([24]) the last 24 values of each serie are stored and then added back when predicting, so for the training set this produces wrong values. WDYT? The fix would consist in having the target in the same scale as the model predictions (right now it's in the original scale)
i

Isaac

09/06/2023, 7:24 PM
I noticed the issues with Differences as well. I'm not sure what a good fix would be, but it seems pretty important to have. Otherwise, I'm not sure how you'd use mlforecast with hierarchicalforecast.
j

José Morales

09/12/2023, 5:37 PM
Hey. We just released 0.9.3 with a couple of bug fixes related to restoring the fitted values with target_transformations. Please let us know if you run into any more issues, it turned out to be more complex than I expected
👀 1
i

Isaac

09/12/2023, 6:55 PM
I'll take a look. Thanks for the update!
Do you know when you'll be pushing the new version of hierarchicalforecast? It was last updated in March, so I've been using the git version since then.
j

José Morales

09/12/2023, 7:31 PM
I have to fix some issue with the aggregation, after that I think we can make a release. I think next week
i

Isaac

09/12/2023, 7:52 PM
Sweet - it's great how fast and responsive your team is.