This message was deleted.
# mlforecast
s
This message was deleted.
j
Hey. You can use them to train the model and for cross validation. We recently removed them from predict though, is that what you want to do?
j
Actually, I just wanted to track which features were used for training. I've noticed that in the MLForecast instance there's a
ts.features_order_
attribute with this info. Is there anything similar? Thanks for the quick answer! Edit: Oh, and regarding predictions... the global model will only use lagged and static features for prediction then?
j
I see. The distributed version has one
ts
per partition, so you could retrieve one of those, but if you've already trained a model you could get the feature names from that instead (from the
models_
attribute) which are just local versions of the distributed model, i.e. XGBRegressor or LGBMRegressor
If you used exogenous features you can't call predict, we have this issue to implement it. It used to be possible to provide a pandas dataframe which would then be broadcasted but it didn't seem like a good approach so we removed it
j
Awesome, thank you for the thorough answer!
j
Is that something you'd like to do (calling predict on the distributed model)? We've also thought of having a
to_local
method which would return a regular MLForecast instance, which would allow you to call predict normally.
j
It would be more convenient to do so depending on the number of SKUs/series that one is trying to predict and the forecasting horizon. However, for my current use case (about 1000~ series and h=30), the solution which returns a regular MLForecast instance would definitely fit the bill!