search

Nixtla / neuralforecast

Scalable and user friendly neural :brain: forecasting algorithms.
https://nixtlaverse.nixtla.io/neuralforecast
Apache License 2.0
3.13k stars 360 forks source link

Is global model training supported? #1070

Closed mvidela31 closed 4 months ago

mvidela31 commented 4 months ago

Description

Hi everyone, thanks for this amazing work!

I was wondering if your library supports training models on multiple time-series (global models) as in the Darts library.

In the Data Inputs documentation it says that unique_id allows to idenfify each time-series in the dataset. However, I couldn't find information about how these time-series are treated under the hood in the model training process. Particullary, I'd like to know if multiple individual models are trained for each time-series (identified by unique_id) or if a single global model is trained for all the time-series in the dataset.

Looking forward to your reply.

Link

No response

jmoralez commented 4 months ago

Hey. The models are trained on all series, the ids are used to produce the correct inputs, future dates, etc.