This message was deleted.

Hey. For retraining there's not really a shortcut. You could save your training set the first time and try just computing the new features when you get future values, but you have to be careful with the number of samples, etc., so it's easier to just recompute the whole thing. You can also keep the same models and just update the series values to generate predictions from that point onwards (docs). For deploying we currently only support serializing the forecast object with pickle/cloudpickle. We may add some methods to save and load but they would just help reducing the size, at the end of the day they would do pretty much that.

Would you please help me understand better than the difference between refit and update_values? It's basically what you are saying right. I cannot see which scenarios is better to choose on or another

Refit involves retraining the models, so if your data changed drastically it would be necessary, otherwise you may be able to keep the same model and just update the series values in order to produce "up to date" features and being able to forecast more accurately. For example if you have a daily frequency and you trained your model with data from last week and you want to forecast next week you could use

h=14

, however if you already know the values of this week you'd get more accurate predictions by providing them to the update method and then use

h=7

, so every time you'd be forecasting one week ahead with the most recent values of your target.