An Investigation Of Approaches For Temporal Hierarchical Reconciliation For Multi-step Forecasting

Abstract

In many forecasting settings, hierarchical structure exists across a collection of time series, such as product category information in a retail setting. In these cases, it is often beneficial to consider hierarchical forecasting approaches and/or hierarchical reconciliation of the predictions. Here, we consider a hierarchical structure based on temporal aggregations and investigate the performance of reconciliation methods for time series with different underlying dynamics and model structures. We find that reconciliation based on the structure of the hierarchy, combined with simpler forecasting models, can outperform more complex forecasting models for temporal hierarchies.