Temporal Fusion Transformers for interpretable multi-horizon time series forecasting

Multi-horizon forecasting often contains a complex mix of inputs – including static (i.e. time-invariant) covariates, known future inputs, and other exogenous time series that are only observed in the past without any prior information on how they interact with target. Several deep learning methods have been proposed, but typically ‘black-box’ models do not shed light use full range present practical scenarios. In this paper, we introduce Temporal Fusion Transformer (TFT) novel attention-based architecture combines high-performance multi-horizon interpretable insights into temporal dynamics. To learn relationships at different scales, TFT uses recurrent layers for local processing self-attention long-term dependencies. utilizes specialized components to select relevant features gating suppress unnecessary components, enabling high performance wide On variety real-world datasets, demonstrate significant improvements over existing benchmarks, highlight three interpretability cases TFT.