This message was deleted.
# mlforecasts
Slackbot
03/05/2024, 8:17 PMThis message was deleted.
j
José Morales
03/05/2024, 8:31 PMHey. Can you provide the code that you're running? There was an issue with the built in lag transformations, are you using those?
j
Jing Qiang Goh
03/05/2024, 11:26 PMHi @José Morales May you share the issue of lag transformation? Perhaps I missed it but did not find discussions in the issue of repository
j
José Morales
03/05/2024, 11:27 PMThe transformations that kept state like ExpandingMean weren't being backed up correctly during predict, so the second model would get the feature updated with the first model's prediction and so on
👌 1
j
jan rathfelder
03/06/2024, 8:15 AMthe first thing i realized, which might be a problem on my side. i have features in my training_df (densities around events like christmas), but when i run predict, i dont give that information (i guess i need to specify my exogenous features). it is a bit weird that predict still runs. otherwise this is my full for training:
jan rathfelder
03/06/2024, 8:19 AMok, i fllowed up on your hint and expandingmean and expotentiallyweightedmean creates the bug. roollingmean creates a stable result: _lag_transforms_={
1: [ExpandingMean()], # creates a bug
7: [ExponentiallyWeightedMean(alpha=0.5)], # creates a bug
24: [RollingMean(_window_size_=30)], # works fine
},
jan rathfelder
03/06/2024, 8:21 AMbut this means the first forecast is the correct one and how it should be and then any further predictions get biased? or is there a general issues with predict atm? also, do you need help with bug fixing? if you point me to the problem, i could have a look and see if i can help.
jan rathfelder
03/06/2024, 9:40 AMbut then when i run some tuning (especially ts-cv tuning), and each forecast is impacted by the previous one, then i am fucked basically...?
jan rathfelder
03/06/2024, 10:03 AMor is it not a problem, because the model is instantiated new every time a new set of params is tested? that could be fine then, right?
j
José Morales
03/06/2024, 1:25 PMThe issue is fixed in the latest version (0.12.0) so you should be fine by just upgrading
j
jan rathfelder
03/06/2024, 4:14 PMah, ok. thanks 🙂
2 Views