Why future values of covariates are not used during training?

[doitslow]

For covariates informed tasks, why not include future values of covariates for training?

[abdulfatir]

The model is trained with future_covariates, if the keys are provided. The actual values inside future_covariates cannot be used because the corresponding target values are not present for loss computation.

[doitslow]

It is clear that future covariates should not be included for loss computation. The question is whether future values of known_covariates seen by the model during the training process? If yes, how shall these values be passed in？

[doitslow]

The defination around future_covariates is a bit misleading, really. Apparently, as your code indicates, future_covariates actually contains target+past_only_covariates+future_covariates just like context; it should really be named as future.