Forecasting hierarchies with coherency-learning


Julien Leprince, Technical University Eindhoven, Netherlands


Many sectors nowadays require accurate and coherent predictions across their organization to effectively operate. Otherwise, decision-makers would be planning based on disparate views of the future, resulting in inconsistent decisions across their sectors. To exploit the coherency requirement of the produced forecasts, recent research has put forward a novel coherency-informed machine learning regressor founded on custom loss functions leveraging optimal reconciliation methods. While promising potentials were outlined, results exhibited discordant performances due to the considerable number of parameters the model possessed. Later, custom neural network designs were investigated inspired by the topological structures of hierarchies, effectively cutting down the complexity of these models while allowing information transfer across specific hierarchical elements. Results unveiled that, in a data-limited setting, models with fewer connections perform best and showcase the value brought by coherency knowledge in both accuracy and coherency forecasting performances, provided individual forecasts were generated within reasonable accuracy limits. Overall, coherency-learning proposes a resourceful, data-efficient, and information-rich learning process opening new pathways toward a novel generation of forecasting regressors.

Associated papers


Margaux Brégère, Sorbonne Université, France