A1179
Title: Multivariate reconciliation for hierarchical time series
Authors: Paulo Canas Rodrigues - Federal University of Bahia (Brazil) [presenting]
Abstract: Time series can often be structured hierarchically by geographic regions, known as hierarchical time series. To ensure coherent forecasts where forecasts respect the same aggregation structure as the data, predictions are typically generated for all levels of the hierarchy. Several methods have been developed to guarantee this coherence by adjusting a set of initial, or base, forecasts, a process known as forecast reconciliation. While most existing approaches focus on univariate time series, multivariate time series can also exhibit hierarchical structures. However, current reconciliation methods are limited to handling one variable at a time, requiring reconciliation to be applied separately to each variable and ignoring potential correlations between them. To overcome this limitation, a multivariate reconciliation methodology is proposed that ensures coherent forecasts while accounting for inter-variable relationships. The method is evaluated through simulation studies under different correlation scenarios across series and variables, and using various base forecasting models. It is also applied to real-world employment data on admissions and dismissals in Brazil. Results show that the proposed multivariate reconciliation approach performs well in both simulated and practical settings.
