Multiple Time Series Forecasting with Temporal Fusion Transformers

The goal of this thesis is to present the Temporal Fusion Transformer model and to evaluate its forecasting capabilities across multiple time series. Its contribution to the field of multi-horizon, multiple time series forecasting is explored, with great focus on the interpretability feature offered by the model. It is observed how improvements to the model performances can be achieved when paired with a form of clustering on the target entities, either by exploiting the natural categorization of the time series considered or by associating similar entities by means of a clustering algorithm on the target variable.