Temporal hierarchies have been proposed as a new approach to model time series. At its core is the idea of looking at the same time series from multiple aggregation levels/views, to help identify both detailed (high frequency) and long-term (low frequency) information in the data. Combining these multiple views into a holistic forecast has been shown to result in improved accuracy and mitigate modelling uncertainty. Furthermore, it allows for coherent forecasts across time, that is, short and long-term forecasts point towards the same future, irrespective of the model specifics and loss functions.
In this talk we introduce the notion of using temporal hierarchies for forecasting, show its connection with hierarchical forecasting, our current understanding of why the approach works (or does not!), and current trends in the research. We make connections with other fields of time series modelling to interpret temporal hierarchies, but also show how they contribute to a diverse set of models and applications. We conclude by showing how temporal hierarchies, used together with conventional hierarchical forecasting can contribute to producing an analytics driven approach to overcoming information silos in companies and supply chains, aiming towards a single representation of the future that is informed by various forecasts and stakeholders.