Abstract
Hierarchical forecasting problems arise when time series have a natural groupstructure, and predictions at multiple levels of aggregation and disaggregationacross the groups are needed. In such problems, it is often desired to satisfythe aggregation constraints in a given hierarchy, referred to as hierarchicalcoherence in the literature. Maintaining hierarchical coherence while producingaccurate forecasts can be a challenging problem, especially in the case ofprobabilistic forecasting. We present a novel method capable of accurate andcoherent probabilistic forecasts for hierarchical time series. We call it DeepPoisson Mixture Network (DPMN). It relies on the combination of neural networksand a statistical model for the joint distribution of the hierarchicalmultivariate time series structure. By construction, the model guaranteeshierarchical coherence and provides simple rules for aggregation anddisaggregation of the predictive distributions. We perform an extensiveempirical evaluation comparing the DPMN to other state-of-the-art methods whichproduce hierarchically coherent probabilistic forecasts on multiple publicdatasets. Compared to existing coherent probabilistic models, we obtained arelative improvement in the overall Continuous Ranked Probability Score (CRPS)of 17.1% on Australian domestic tourism data, 24.2 on the Favorita grocerysales dataset, and 6.9% on a San Francisco Bay Area highway traffic dataset.