formatting TimeSeriesForecaster input to avoid gaps

[mu22le]

Hi, the time series I am trying to work with has many gaps. Timestamps are available, so it's easy to identify them.

I would like to use TimeSeriesForecaster (TSF from now on) while avoiding training samples with these gaps in them.

Is there a way to format the data to let TSF avoid the gaps? For instance, if the input data is split in several files, will TSF avoid building the lookback window crossing from one file to another? Or can I pre-format the data by building the lookback window myself so that I can implement the logic that avoids the gaps? Alternatively, can I make TSF aware of the timestamps?

Does TSF support TensorFlow's MaskedTimeSeries? If I use them to mark the gaps, what will the behaviour of the lookback window be?

Gaps in time series are very common, so I imagine someone has already come up with a solution to this problem.