Clarification of FieldNames

[MFK2021]

Hi everyone, in order to convert my csv file to a format which can be used by the GluonTS library, I will use FieldNames.

FEAT_STATIC_CAT , FEAT_STATIC_REAL, FEAT_DYNAMIC_CAT and FEAT_DYNAMIC_REAL are pretty straight forward and easy to understand. But what is the meaning of the following FieldNames and how do they differ to the previous ones? :

• PAST_FEAT_DYNAMIC_REAL ?
• FEAT_DYNAMIC_REAL_LEGACY ?
• FEAT_DYNAMIC ?
• PAST_FEAT_DYNAMIC ?

Thank you for your help!

[jaheba]

Hi,

FEAT_DYNAMIC_REAL and FEAT_DYNAMIC_CAT require that features are passed both in training and prediction. For example, if you want to predict demand for an item and use price as a feature, you can provide the price as a dynamic feature. Then for prediction, you need to provide the price to get a prediction.

However, if a feature is only available in training and not prediction, some models support PAST_FEAT_DYANMIC_REAL. For example, you could use the number of views for an item's webpage as a feature in training. For prediction, you do not pass that information, but still get a forecast.

Note while many models support FEAT_DYNAMIC_REAL, only few support PAST_FEAT_DYNAMIC_REAL. So it's best to check the specific model.

---

I think FEAT_DYNAMIC_REAL_LEGACY should just be removed, it's not used as far as I can see.

It appears that FEAT_DYNAMIC and PAST_FEAT_DYNAMIC are used in some models internally, so they should also ideally removed from FieldName. In any case, you can safely ignore these for generating data-input.