Hi everyone! I’m having trouble understanding this paragraph—does it mean that the model’s weights are determined based on all unique_ids, making the model global with the same weights applied for predicting each unique_id? Or is the model local, with different weights for each unique_id? I’ve performed time-series clustering and am considering whether it would be beneficial to train models independently for each cluster or if this approach would have no impact on the forecasting results. Is there any difference between the AutoModels and the non AutoModels? Any insights would be greatly appreciated!

Yes, the models are global, so the same weights for all unique ids. The difference between auto and non-auto models is that auto models run hyperparameter optimization automatically and find the best configuration. For the non-auto models, you set a single value for each parameter.

Ok, thanks!! Yes, I tried with AutoModels. The strange thing here was that when I copied the hyperparameters of the AutoModels with the lowest loss into the same non auto model, they yielded better results compared to the AutoModels output. Is that because weights are calculated again of course for the non-auto models, so results can differ accordingly between AutoModels and NonAutoModels with the hyperparameter settings derived from the AutoModels?

Did you also set the random_seed in the model? Otherwise, sometimes, due to the random intialization, results might differ slightly.

Ah no, I have not set a random seed. Thanks for the tip. But results differed like 5-10%, is that reasonable?

Hard to say, but it's likely to the seed being different