Conformal prediction as an alternative for uncertainty quantification in probabilistic hierarchical reconciliation

Presenter

Azul Garza, Nixtla, Mexico

Abstract

This presentation introduces an agnostic approach for probabilistic hierarchical reconciliation in time series forecasting using conformal prediction as an alternative to traditional uncertainty quantification methods. Our technique is a flexible two-step approach that produces forecast distributions using only reconciled point forecasts. In the first step, point forecasts for all levels in the hierarchy are generated using any forecasting method. In the second step, any reconciliation technique can be applied to feed the conformal prediction algorithm. We demonstrate the effectiveness of our approach by comparing its performance and accuracy against state-of-the-art results. Additionally, we provide an open-source library that enables practitioners to incorporate conformal prediction into their reconciliation strategies. By providing an alternative method for uncertainty quantification, our approach offers practitioners a valuable tool for improving the accuracy and reliability of hierarchical time series forecasting. Our work contributes to advancing the field of hierarchical time series forecasting and provides a promising avenue for future research.

Discussant

Mitchell O’Hara-Wild, Monash University, Australia