How to measure the impact of feat_static_cat in DeepAR

[rajnish-garg]

I have a time series for each city that I am forecasting using DeepAR Estimator. It is giving decent results.

I have also list of static category attributes related to each series for e.g. density (high, low.. ), seasonal etc. that i am looking to feed using feat_static_cat so that I can leverage the relationship between different sequences to improve the accuracy metric.

I have couple of qns on this:

• Is it best practice to test/update trainer parameters (layers, learning rate ?) if we are adding new static features, if yes, what are the most common one ?
• If I add 5 new static feature, how can i know which have higher impact on accuracy ?

[StatMixedML]

If I add 5 new static feature, how can i know which have higher impact on accuracy

You might be interested looking into Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting. It is an attention-based architecture which combines high-performance multi-horizon forecasting with interpretable insights.

[lorrp1]

@rajnish-garg @StatMixedML have you tried adding features to models that require a multidimensional target (i have 1 target and multiple features)? i keep getting errors and i dont know how to use to multivariategrouper with just 1 target