“Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting” by Bryan Lim, Sercan Arik, Nicolas Loeff, and Tomas Pfister at Google was posted to arXiv in 2019. It tackles a practical reality of business forecasting: the inputs come in different flavors, and users want to understand the model, not just trust it.
The Temporal Fusion Transformer, or TFT, is built to combine three kinds of input cleanly. Static information that never changes, such as which store a series belongs to; inputs known in advance into the future, such as holidays or planned promotions; and past observed values whose future is unknown. It uses recurrent layers to capture short-range local patterns and an interpretable self-attention mechanism to capture long-range dependencies, plus gating components that let the model ignore inputs that turn out not to matter. A key selling point is interpretability: the attention weights and variable-selection components reveal which inputs and which past time steps drove a given forecast.
TFT became a popular choice when accuracy and explainability both matter, and it is implemented in several open-source forecasting libraries.
Why business readers should care: TFT gives accurate multi-step forecasts while showing which factors, such as a promotion or a season, are driving the prediction, which builds trust with the planners who must act on it.