Demand forecasting example and usage of TimeSeriesDataSet for Validation and Testing

• PyTorch-Forecasting version: 0.9.0
• PyTorch version: 1.9.0
• Python version: 3.7
• Operating System: Windows and Linux

Hello,

thanks for the amazing package, I really enjoy your work.

I had a small question regarding the TimeSeriesDataSet and it's relation with training, validation and testing data. For my elaboration, I primarily focus on your example of the Demand forecasting.

If I see it correctly, you are defining for each of the three datasets an individual TimeSeriesDataSet, thus also create a Normalizer for each dataset seperately. However, if I am not mistaken, this would include future information in the normalization process, which is highly problematic in the time series context. On the other hand, I also saw that you designed a specific function for this problem, i.e. the TimeSeriesDataSet.from_dataset().

Therefore my question, if you could elaborate, why you chose the seperate creation of the TimeSeriesDataSet instead of transfering the training Normalizers ?

Thanks in advance

Start with the Demand forecasting example and the TimeSeriesDataSet.from_dataset() entry point, then compare how the training, validation, and testing datasets create and reuse normalizers. Done means documenting whether separate dataset construction leaks future information and when the training normalizers should be transferred.