What is temporal fusion transformer (TFT) in time series forecasting?

TFT is a deep learning architecture combining attention, LSTM, and gating mechanisms for interpretable multi-horizon forecasting.

Temporal Fusion Transformer (TFT) integrates recurrent layers for local patterns, attention for long-range dependencies, and gating mechanisms for feature selection. It handles static, known, and observed inputs separately, making it highly flexible. TFT also provides interpretability through variable importance and attention visualization.