Forecasting interrupted time series

Working papers
Authors

Rob J Hyndman, Bahman Rostami-Tabar

Published

5 March 2024

Publication details

Working paper

Links

pdf

 

Forecasting interrupted time series data is a major challenge for forecasting teams, especially in light of events such as the COVID-19 pandemic. This paper investigates several strategies for dealing with interruptions in time series forecasting, including highly adaptable models, intervention models, marking interrupted periods as missing, forecasting what may have been, downweighting the interruption period, and ensemble models. Each approach offers specific advantages and disadvantages, such as adaptability, memory retention, data integrity, flexibility, and accuracy. We evaluate the effectiveness of these strategies using two actual datasets that were interrupted by COVID-19, and we provide recommendations for how to handle these interruptions. This work contributes to the literature on time series forecasting, offering insights for academics and practitioners dealing with interrupted data in numerous domains.