Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

hi everyone, thank you for the work without which my phd would be a lot harder.
I was wondering if it possible to automl and use one model per step approach as described here: <https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/one_model_per_horizon.html>

Grazie mille!

Hey. This should be possible by setting `max_horizon` in the `fit_config` , e.g. `fit_config=lambda trial: {'max_horizon': 10}`

image.png

Hello again,

Maybe an additional question, I am training and hypertuning a LASSO.

The value alpha should be different for each step how to implement such behaviour in Nitxla?

At the moment the alpha value is the same for each model

Grazie ancora,
B.

You can use `MLForecast.preprocess` to generate the training set as is done <https://github.com/Nixtla/mlforecast/discussions/304#discussioncomment-8225978|here>, train your models individually (one for each column in y) and assign them to the `MLForecast.models_` attribute as a list, e.g. `mlf.models_ = {'lasso': [model1, model2, model3]}` , you can then call `MLForecast.predict` as usual