How to use past_feat_dynamic_real ?

[sayedathar11]

We are using GluonTS Deepar model to predict target at a specific timeframe , we have a feature which is numerical . We compute some numerical stats on this feature and we can use this to predict Future Values of Target .

But we now realized that values of this feature won't be available in Future when we have to predict the target . We can't use feat_dynamic_real and as per our understanding we have to use past_feat_dynamic_real and instead use past stats values of this numerical feature to predict future value .

Whether this understanding is correct , also can you please add documentation regarding how to use past_feat_dynamic_real .

[pratikgehlott]

Hi @lostella , @mbohlkeschneider I am also facing the same doubt how are we suppose to use past_feat_dynamic_real ?

[mbohlkeschneider]

Hi everyone,

thank you for your questions. past_feat_dynamic_real is still not officially supported for DeepAR. There is a PR but it is not yet merged.

[spunknic]

I have the same doubt. I mean, in the real world we will never have future observations of any feature

[jimbow973]

why never? let's say you forecast sales and you know already the discount periods/discount amount etc. as future predictor. I think we can find many applications.