Forecasting in hierarchical systems

: Improving accuracy, enhancing scalability and managing risk

Abstract

We focus on the problem of reconciliation of independently produced sets of forecasts for large
collections of time series, where some subset of these are formed as linear combinations of several oth-
ers. In general terms it is the case that these reconciled forecasts are more accurate than the original
un-reconciled forecasts, as well as satisfying the natural aggregation constraints. Despite this attrac-
tive feature, there are several practical drawbacks to the approach, which we address here. Firstly the
models are heavily parameterised, and therefore not robust to errors and omissions in the underlying
data used to fit them. Secondly they do not scale to the large systems of equations where their benefits
would be most keenly felt. Our solutions address both of these problems.

Date of Award	25 Jun 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Fotios Petropoulos (Supervisor) & Michael Tipping (Supervisor)