New analysis indicates that forecasting infectious diseases is inherently more challenging for certain seasons, pathogens, and population sizes. This could explain why predictive model performance fluctuates.
Researchers used a forecastability metric based on spectral entropy to measure the structure and disorder within respiratory virus surveillance signals. They found that across California's influenza hospital admissions, forecastability varied significantly year-to-year. The 2017-2018 season, with the highest hospital admissions, was the most forecastable, while the 2009 H1N1 pandemic showed lower forecastability than expected due to atypical epidemic behavior.
Forecastability also increased with population size for COVID-19 and influenza hospital admissions in U.S. states and nationally. Higher forecastability was linked to improved forecast performance, particularly for ensemble models. However, the 2022-2023 influenza season was an exception, with dynamics influenced by post-COVID-19 pandemic shifts and limited historical data for newer targets.
These findings suggest that infectious disease forecastability can help contextualize changes in forecast performance across seasons and locations. The study also highlights that some disease targets may be intrinsically less predictable, rather than poorly modeled. Clinicians and health systems should interpret outbreak forecasts with this nuance in mind.