A few months ago I asked how time series shorter than input_size were handled (whether they were padded or ignored) and was told that zero-padding was used. @Cristian (Nixtla) Could you explain what changed with this commit https://github.com/Nixtla/neuralforecast/pull/715 ? Thank you very much
08/16/2023, 12:21 PM
Hi @Manuel! Yes. Previously, we padded time series with zeros to make them of same size internally in the temporal tensor. However, this only applied to complete shorter time series compared to longer series. We realized that when all time series were shorter than input size it was raising errors. Now, we are adding this extra padding for the case when all time series are shorter than input size. Sorry for the confusion.
Even with this fix, it is not recommended to train models in this extreme case where all time series are too short. The model will only have 1 window to train per series. Statistical models will probably work better.
08/16/2023, 12:27 PM
@Cristian (Nixtla) Thank you! Does the new additional padding have available_mask=0 ?