tft

tft is an R implementation of the Temporal Fusion Transformer, a neural network architecture for multi-horizon time series forecasting. It allows you to forecast multiple time series simultaneously within a single model.

The key advantage of TFT over conventional forecasting methods is how it handles different types of input data. The architecture encodes static covariates and time-varying inputs (both known and unknown) through separate pathways, rather than treating all features uniformly. It is built on torch for R and has shown strong benchmark results on standard forecasting tasks.