What is the difference between dynamic_dfs and X_df when making forecasts?

[davidrpugh]

Thanks for the awesome times series forecasting libraries! I am teaching myself your libraries by working my way through the Kaggle Store Sales Time Series Forecasting competition.

I am trying to understand the difference between dynamic_dfs and X_df arguments to the predict method. What is the difference?

In the store sales competition there is a variable for oil price whose values we are given for the forecasting window. This seems like it should be included in X_df as the price of oil is not impacted by store sales. However, there are other variables on promotion and transactions whose values could conceivably depend on sales. Should these variables be included as dynamic_dfs?

[jmoralez]

Hey @davidrpugh, thanks for using mlforecast. They're both meant to do the same, which is providing future values of exogenous features. The dynamic_dfs argument is deprecated and it used to take a list of dataframes and perform merges internally, you should now use X_df which takes a single dataframe with id, time and the values for the exogenous features. If you're using that please upgrade to 0.10.0, since there was a bug in the X_df processing.

For your example I think they should all be included in X_df, since their values change over time. Please also make sure to set the static_features argument of MLForecast.fit accordingly, so that those features aren't considered as static. You may find this guide useful.

Please let us know if you have further doubts.