Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting

Summary

This paper, published in the International Journal of Forecasting (2021), introduces the Temporal Fusion Transformer (TFT), a novel attention-based deep learning architecture designed for multi-horizon time series forecasting that also provides interpretable outputs. The model combines gating mechanisms, variable selection networks, and multi-head attention to identify temporally relevant patterns and rank the contribution of input variables. Evaluated across several benchmark datasets, TFT reportedly outperforms existing forecasting methods whilst offering greater transparency in how predictions are generated.

UK applicability

TFT is a general-purpose machine learning architecture not developed for UK-specific contexts; however, it is potentially applicable to UK agri-food decision-making contexts such as yield forecasting, nutrient management planning, or supply chain modelling where interpretable time series predictions are required.

Key measures

Forecast accuracy metrics (P50/P90 quantile loss, RMSE); interpretability measures (variable importance scores, attention weights); comparison against benchmark forecasting models across multiple datasets

Outcomes reported

The study evaluated a deep learning architecture (Temporal Fusion Transformer) on its ability to produce accurate, interpretable forecasts across multiple future time horizons using diverse real-world datasets. Performance was measured against benchmark models on prediction accuracy and interpretability of learned temporal patterns.

Topic tags