Gluon-ts Dynamic Features for deepar estimators


Very excited about the glucon-ts for building forecasting models. I did look at the tutorials, but I was wondering if we specify the dynamic features and categories in ListDataset.

For example, this is how we define the target and the start. But, DeepAR, supports dynamic features and categories. I couldn’t find any documentation on this, that describes ListDataset.

training_data = ListDataset(
[{“start”: df.index[0], “target”: df.value[:“2015-04-05 00:00:00”]}],
freq = “5min”


The source code for the ListDataset class might provide a clue. You can find that here:

It seems like there’s a boolean argument for whether your inputs are one-dimensional or not. I would expect that changing this to false might allow you have dynamic features and categories. Note I haven’t really played around with gluon-ts yet so this is just a guess.