Constructing hierarchical time series through clustering: Is there an optimal way for forecasting?

Abstract

Forecast reconciliation has attracted significant research interest in recent years, with most studies taking the hierarchy of time series as given. We extend existing work that uses time series clustering to construct hierarchies to improve forecast accuracy in three ways. First, we investigate multiple approaches to clustering, including different clustering algorithms, how time series are represented, and how the distance between time series is defined. We find that cluster-based hierarchies improve forecast accuracy relative to two-level hierarchies. Second, we devise an approach based on random permutation of hierarchies, keeping the hierarchy structure fixed while time series are rando..