Hierarchical forecasting: applications in complex scenarios

Abstract

Forecast reconciliation is a state-of-art framework for hierarchical time series forecasting. Nevertheless, numerous challenges emerge when implementing it in real-world hierarchies characterized by highly diverse time series patterns. We consider two prevalent practical scenarios. The first scenario pertains to a retail product hierarchy, encompassing both fast-moving and intermittent time series. We propose an variant to the conventional reconciliation approach that keeps forecasts of a pre-specified subset of variables unchanged or “immutable”. The second scenario involves situations where all time series are discrete-valued, prompting the development of a discrete forecast reconciliation method.