Implication of null values in exogenous variables with LightGBM model

[matsuobasho]

I understand that missing data in exogenous variables isn't good, but in my use-case sometimes it's better to leave the missing values rather than introduce more uncertainty by filling them in with something.

I noticed that if I predict on a model with exogenous data, some of which is missing, I will get an output.

y_hat = fcst.predict(h=100,X_df = test)

However, I also see the UserWarning about missing values in the exogenous variables. What is the implication of this user warning? From what I understand if I'm using LightGMB, then it will learn the best direction to go in a node when that variable has a missing value, and that seems acceptable.

[jmoralez]

Hey. LightGBM chooses the best direction for missing values if they are also seen during training, which isn't the case for lag features with the default settings (dropna=True), that's why that warning is there. If you're confident on what you're doing you can just wrap the predict call in a warnings.catch_warnings context manager.