Abstract
Relationships among time series can be exploited as inductive biases inlearning effective forecasting models. In hierarchical time series,relationships among subsets of sequences induce hard constraints (hierarchicalinductive biases) on the predicted values. In this paper, we propose agraph-based methodology to unify relational and hierarchical inductive biasesin the context of deep learning for time series forecasting. In particular, wemodel both types of relationships as dependencies in a pyramidal graphstructure, with each pyramidal layer corresponding to a level of the hierarchy.By exploiting modern - trainable - graph pooling operators we show that thehierarchical structure, if not available as a prior, can be learned directlyfrom data, thus obtaining cluster assignments aligned with the forecastingobjective. A differentiable reconciliation stage is incorporated into theprocessing architecture, allowing hierarchical constraints to act both as anarchitectural bias as well as a regularization element for predictions.Simulation results on representative datasets show that the proposed methodcompares favorably against the state of the art.