TemporalFusionTransformer quesion

[birard]

I’ve been learning how to apply the Temporal Fusion Transformer for various forecasting tasks, but I’m currently facing a dilemma that I’d like to clarify. In my dataset, one real-valued, observed-input feature has very severe missingness—only 25% coverage—and this feature is known to be correlated with the prediction target. Should I:

1. Follow the original paper’s approach and add a binary “missing” flag column,

2. Use a dedicated missing-value embedding inside the model,

3. Or simply exclude that feature entirely because the available data is too sparse?

I would appreciate any guidance on this.

[fkiraly]

All three are possible approaches to the problem - my recommendation, have a clean backtesting and evaluation setup, and see what your problem KPI are for the three?