I am experimenting with the Temporal-Fusion Transformer in the forecasting library and have a question about including data without targets. I have a tremendous amount of data with most or all of the the covariate information, but no target value; an analogy would be something like I have a tremendous amount of individual historical stock values, but no stock Index value. I then have a modest dataset of time series that do have target values, e.g. something analogous to a stock index value.
Can these ‘un-labeled’ time-series be included in the training set? My understanding of the TFT is very poor, but my goal is to accomplish something akin to MAE pre-training. Sorry if this is a nonsense question in this context, still trying to wrap my head around how the model works.