Include dynamic features to future predictions data

[babaniyi]

I intend to make predictions for the future (where we don't have the target). However, I can't include the dynamic features such as use_dynamic_feat_real or use_dynamic_feat_cat. How do I go about this?
@lostella

[ekerstein]

Per this page: https://ts.gluon.ai/stable/getting_started/background.html#target-and-features

Most models require dynamic features to be available in the future time-range when making predictions.

If you can't provide the dynamic features in the future then you will probably have to use a different model, or bring the features into the model as separate time series and train a global model that includes them.

[Krushi10]

What about DeepAREstimator??

[ekerstein]

It's in that same link as above:

estimator_with_features = DeepAREstimator(
    freq=ds.freq,
    prediction_length=prediction_length,
    use_feat_dynamic_real=True,
    use_feat_static_cat=True,
    cardinality=[
        2,
    ],
    trainer=Trainer(epochs=1),
)

[babaniyi]

It's in that same link as above:

estimator_with_features = DeepAREstimator(
    freq=ds.freq,
    prediction_length=prediction_length,
    use_feat_dynamic_real=True,
    use_feat_static_cat=True,
    cardinality=[
        2,
    ],
    trainer=Trainer(epochs=1),
)

@ekerstein @lostella
The issue comes with preparing the data for future predictions.

For instance say you you have the data in the following format:
Train: Jan 1, 2021 - August 31, 2022 (Have the data available)
Future prediction: September 1, 2022 - September 30, 2022 (Data not available since it is future)

In this case, suppose the training data has dynamic real variables. How do we predict for the future (Sep 1 - 30) and include the dynamic real variables of the future?

[ekerstein]

I've written up an example. See here: https://gist.github.com/ekerstein/149099bb0f9cda53d3f97fdda83db879

A couple notes:

• For training, you don't need to provide the future feat_dynamic_real items. So the target and feat_dynamic_real should be the same length. This is also true when evaluating a train/test split.
• Only when using the predict() method do you have to provide a dataset with a target length of n and a feat_dynamic_real length of n+prediction_length.
• On my example, you can change FUTURE_DYNAMIC_REAL from 1 to 0 to see the predictions change.
• On my example, you can turn use_feat_dynamic_real on/off to see the effects of using the dynamic features.

[babaniyi]

@ekerstein Thanks for the example you provided. I get it clearly now.