Nikos Kourentzes, PhD, University of Skövde, Sweden

Abstract: The use of multiple temporal aggregation levels of a time series has gained some momentum for improving forecast accuracy and identifying the structure of a time series. In this talk, we review the main ideas behind its use. So far, there are two main approaches to do this, the first using a forecast combination view, namely the Multiple Temporal Aggregation Algorithm; and the second using hierarchical forecasting, namely Temporal Hierarchies. While both have demonstrated benefits, the latter is more flexible, overcoming many of the limitations of the former, importantly allowing the use of forecasts from any source. Yet, in terms of forecast accuracy, both remain competitive. We show that reconciling the two views leads to a modified version of Temporal Hierarchies that substantially improves over the standard implementation. Furthermore, we show that the standard implementation violates helpful constraints established in the forecast combination literature while imposing unhelpful constraints inherited from hierarchical forecasting. Based on the modified Temporal Hierarchies, we identify conditions that using temporal hierarchies is admissible to standard forecasts. Finally, we connect the hierarchical structure to the organizations’ context and show how this impacts the evaluation of forecasts.