Abstract
In many real-world forecasting problems, multiple time series are interrelated through known linear constraints, and when forecasts are produced, it is natural to want them to be coherent - that is, for the forecasts to satisfy the same constraints as the original data. Forecast reconciliation is a post-forecasting process that transforms a set of incoherent forecasts (regardless of how they were obtained) into a coherent set that respects these relationships. Achieving coherence is not merely about enforcing numerical alignment; it reflects a deeper understanding of the interdependencies within the data and enables more consistent and effective decision-making across levels of aggregation. In this seminar, we present an overview of point and probabilistic forecast reconciliation techniques developed over the past 15 years. We will deal with reconciliation in three different frameworks: cross-sectional, where forecasts are structured across hierarchies or groupings (but not limited to); temporal, where constraints apply across multiple time frequency (e.g., daily, weekly, monthly); and cross-temporal, where both dimensions interact. These frameworks provide a unified perspective on coherence in multivariate time series forecasting, applicable to a wide range of domains, including economics, energy, demography, and supply chain.
Link Microsoft Teams
Organisation
Luca Trapin