Problems with padding in multivariate forecasts

[carlos-ariza3]

Hello,

We are performing multivariate forecasting for multiple time series. To achieve this, we have applied padding (to the “series” parameter) that considers the temporal position of each time series to equalize their lengths.

Based on this, we are currently facing two issues:

1. for the exact same time series, we get different forecasts depending on how the padding is applied. For example, for a given series, if two values are padded at the beginning, the forecast is different than if three values are padded at the beginning and one at the end.
We would like to know if this behavior is expected or if there is a way to solve this issue.
Note: We have taken the padding_mask parameter into account (setting the value to False for padded positions) as well as an id_mask (using distinct values for independent time series and the same value for related variables we want to predict together).
2. If we change the padding value in the “series” parameter (for example, between [0, 10, 100, 1000]), the forecasts we get are different. ¿I was wondering what padding value you would recommend?