Individual and combination approaches to forecasting hierarchical time series with correlated data: an empirical study

Hierarchical time series arise in manufacturing and service industries when the products or services have the hierarchical structure, and top-down and bottom-up methods are commonly used to forecast the hierarchical time series. One of the critical factors that affect the performance of the two methods is the correlation between the data series. This study attempts to resolve the problem and shows that the top-down method performs better when data have high positive correlation compared to high negative correlation and combination of forecasting methods may be the best solution when there is no evidence of the correlationship. We conduct the computational experiments using 240 monthly data series from the 'Industrial' category of the M3-Competition and test twelve combination methods for the hierarchical data series. The results show that the regression-based, VAR-COV and the Rank-based methods perform better compared to the other methods.